github/InvokeAI
1
1
Fork 0
mirror of https://github.com/invoke-ai/InvokeAI.git synced 2026-04-23 03:00:31 -04:00
Code Issues Packages Projects Releases Wiki Activity

Squashed commit of the following:

Browse Source 
commit 1c649e4663
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Mon Sep 12 13:29:16 2022 -0400

    fix torchvision dependency version #511

commit 4d197f699e
Merge: a3e07fb 190ba78
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Mon Sep 12 07:29:19 2022 -0400

    Merge branch 'development' of github.com:lstein/stable-diffusion into development

commit a3e07fb84a
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Mon Sep 12 07:28:58 2022 -0400

    fix grid crash

commit 9fa1f31bf2
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Mon Sep 12 07:07:05 2022 -0400

    fix opencv and realesrgan dependencies in mac install

commit 190ba78960
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Mon Sep 12 01:50:58 2022 -0400

    Update requirements-mac.txt

    Fixed dangling dash on last line.

commit 25d9ccc509
Author: Any-Winter-4079 <50542132+Any-Winter-4079@users.noreply.github.com>
Date:   Mon Sep 12 03:17:29 2022 +0200

    Update model.py

commit 9cdf3aca7d
Author: Any-Winter-4079 <50542132+Any-Winter-4079@users.noreply.github.com>
Date:   Mon Sep 12 02:52:36 2022 +0200

    Update attention.py

    Performance improvements to generate larger images in M1 #431

    Update attention.py

    Added dtype=r1.dtype to softmax

commit 49a96b90d8
Author: Mihai <299015+mh-dm@users.noreply.github.com>
Date:   Sat Sep 10 16:58:07 2022 +0300

    ~7% speedup (1.57 to 1.69it/s) from switch to += in ldm.modules.attention. (#482)

    Tested on 8GB eGPU nvidia setup so YMMV.
    512x512 output, max VRAM stays same.

commit aba94b85e8
Author: Niek van der Maas <mail@niekvandermaas.nl>
Date:   Fri Sep 9 15:01:37 2022 +0200

    Fix macOS `pyenv` instructions, add code block highlight (#441)

    Fix: `anaconda3-latest` does not work, specify the correct virtualenv, add missing init.

commit aac5102cf3
Author: Henry van Megen <h.vanmegen@gmail.com>
Date:   Thu Sep 8 05:16:35 2022 +0200

    Disabled debug output (#436)

    Co-authored-by: Henry van Megen <hvanmegen@gmail.com>

commit 0ab5a36464
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 17:19:46 2022 -0400

    fix missing lines in outputs

commit 5e433728b5
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 16:20:14 2022 -0400

    upped max_steps in v1-finetune.yaml and fixed TI docs to address #493

commit 7708f4fb98
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 16:03:37 2022 -0400

    slight efficiency gain by using += in attention.py

commit b86a1deb00
Author: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com>
Date:   Mon Sep 12 07:47:12 2022 +1200

    Remove print statement styling (#504)

    Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>

commit 4951e66103
Author: chromaticist <mhostick@gmail.com>
Date:   Sun Sep 11 12:44:26 2022 -0700

    Adding support for .bin files from huggingface concepts (#498)

    * Adding support for .bin files from huggingface concepts

    * Updating documentation to include huggingface .bin info

commit 79b445b0ca
Merge: a323070 f7662c1
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 15:39:38 2022 -0400

    Merge branch 'development' of github.com:lstein/stable-diffusion into development

commit a323070a4d
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 15:28:57 2022 -0400

    update requirements for new location of gfpgan

commit f7662c1808
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 15:00:24 2022 -0400

    update requirements for changed location of gfpgan

commit 93c242c9fb
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 14:47:58 2022 -0400

    make gfpgan_model_exists flag available to web interface

commit c7c6cd7735
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 14:43:07 2022 -0400

    Update UPSCALE.md

    New instructions needed to accommodate fact that the ESRGAN and GFPGAN packages are now installed by environment.yaml.

commit 77ca83e103
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 14:31:56 2022 -0400

    Update CLI.md

    Final documentation tweak.

commit 0ea145d188
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 14:29:26 2022 -0400

    Update CLI.md

    More doc fixes.

commit 162285ae86
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 14:28:45 2022 -0400

    Update CLI.md

    Minor documentation fix

commit 37c921dfe2
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 14:26:41 2022 -0400

    documentation enhancements

commit 4f72cb44ad
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 13:05:38 2022 -0400

    moved the notebook files into their own directory

commit 878ef2e9e0
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 12:58:06 2022 -0400

    documentation tweaks

commit 4923118610
Merge: 16f6a67 defafc0
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 12:51:25 2022 -0400

    Merge branch 'development' of github.com:lstein/stable-diffusion into development

commit defafc0e8e
Author: Dominic Letz <dominic@diode.io>
Date:   Sun Sep 11 18:51:01 2022 +0200

    Enable upscaling on m1 (#474)

commit 16f6a6731d
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 12:47:26 2022 -0400

    install GFPGAN inside SD repository in order to fix 'dark cast' issue #169

commit 0881d429f2
Author: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com>
Date:   Mon Sep 12 03:52:43 2022 +1200

    Docs Update (#466)

    Authored-by: @blessedcoolant
    Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>

commit 9a29d442b4
Author: Gérald LONLAS <gerald@lonlas.com>
Date:   Sun Sep 11 23:23:18 2022 +0800

    Revert "Add 3x Upscale option on the Web UI (#442)" (#488)

    This reverts commit f8a540881c.

commit d301836fbd
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 10:52:19 2022 -0400

    can select prior output for init_img using -1, -2, etc

commit 70aa674e9e
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 10:34:06 2022 -0400

    merge PR #495 - keep using float16 in ldm.modules.attention

commit 8748370f44
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 10:22:32 2022 -0400

    negative -S indexing recovers correct previous seed; closes issue #476

commit 839e30e4b8
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 11 10:02:44 2022 -0400

    improve CUDA VRAM monitoring

    extra check that device==cuda before getting VRAM stats

commit bfb2781279
Author: tildebyte <337875+tildebyte@users.noreply.github.com>
Date:   Sat Sep 10 10:15:56 2022 -0400

    fix(readme): add note about updating env via conda (#475)

commit 5c43988862
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 10 10:02:43 2022 -0400

    reduce VRAM memory usage by half during model loading

    * This moves the call to half() before model.to(device) to avoid GPU
    copy of full model. Improves speed and reduces memory usage dramatically

    * This fix contributed by @mh-dm (Mihai)

commit 99122708ca
Merge: 817c4a2 ecc6b75
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 10 09:54:34 2022 -0400

    Merge branch 'development' of github.com:lstein/stable-diffusion into development

commit 817c4a26de
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 10 09:53:27 2022 -0400

    remove -F option from normalized prompt; closes #483

commit ecc6b75a3e
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 10 09:53:27 2022 -0400

    remove -F option from normalized prompt

commit 723d074442
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Fri Sep 9 18:49:51 2022 -0400

    Allow ctrl c when using --from_file (#472)

    * added ansi escapes to highlight key parts of CLI session

    * adjust exception handling so that ^C will abort when reading prompts from a file

commit 75f633cda8
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Fri Sep 9 12:03:45 2022 -0400

    re-add new logo

commit 10db192cc4
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Fri Sep 9 09:26:10 2022 -0400

    changes to dogettx optimizations to run on m1
    * Author @any-winter-4079
    * Author @dogettx
    Thanks to many individuals who contributed time and hardware to
    benchmarking and debugging these changes.

commit c85ae00b33
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Thu Sep 8 23:57:45 2022 -0400

    fix bug which caused seed to get "stuck" on previous image even when UI specified -1

commit 1b5aae3ef3
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Thu Sep 8 22:36:47 2022 -0400

    add icon to dream web server

commit 6abf739315
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Thu Sep 8 22:25:09 2022 -0400

    add favicon to web server

commit db825b8138
Merge: 33874ba afee7f9
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Thu Sep 8 22:17:37 2022 -0400

    Merge branch 'deNULL-development' into development

commit 33874bae8d
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Thu Sep 8 22:16:29 2022 -0400

    Squashed commit of the following:

    commit afee7f9cea
    Merge: 6531446 171f8db
    Author: Lincoln Stein <lincoln.stein@gmail.com>
    Date:   Thu Sep 8 22:14:32 2022 -0400

        Merge branch 'development' of github.com:deNULL/stable-diffusion into deNULL-development

    commit 171f8db742
    Author: Denis Olshin <me@denull.ru>
    Date:   Thu Sep 8 03:15:20 2022 +0300

        saving full prompt to metadata when using web ui

    commit d7e67b62f0
    Author: Denis Olshin <me@denull.ru>
    Date:   Thu Sep 8 01:51:47 2022 +0300

        better logic for clicking to make variations

commit afee7f9cea
Merge: 6531446 171f8db
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Thu Sep 8 22:14:32 2022 -0400

    Merge branch 'development' of github.com:deNULL/stable-diffusion into deNULL-development

commit 653144694f
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Thu Sep 8 20:41:37 2022 -0400

    work around unexplained crash when timesteps=1000 (#440)

    * work around unexplained crash when timesteps=1000

    * this fix seems to work

commit c33a84cdfd
Author: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com>
Date:   Fri Sep 9 12:39:51 2022 +1200

    Add New Logo (#454)

    * Add instructions on how to install alongside pyenv (#393)

    Like probably many others, I have a lot of different virtualenvs, one for each project. Most of them are handled by `pyenv`.
    After installing according to these instructions I had issues with ´pyenv`and `miniconda` fighting over the $PATH of my system.
    But then I stumbled upon this nice solution on SO: https://stackoverflow.com/a/73139031 , upon which I have based my suggested changes.

    It runs perfectly on my M1 setup, with the anaconda setup as a virtual environment handled by pyenv.

    Feel free to incorporate these instructions as you see fit.

    Thanks a million for all your hard work.

    * Disabled debug output (#436)

    Co-authored-by: Henry van Megen <hvanmegen@gmail.com>

    * Add New Logo

    Co-authored-by: Håvard Gulldahl <havard@lurtgjort.no>
    Co-authored-by: Henry van Megen <h.vanmegen@gmail.com>
    Co-authored-by: Henry van Megen <hvanmegen@gmail.com>
    Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>

commit f8a540881c
Author: Gérald LONLAS <gerald@lonlas.com>
Date:   Fri Sep 9 01:45:54 2022 +0800

    Add 3x Upscale option on the Web UI (#442)

commit 244239e5f6
Author: James Reynolds <magnusviri@users.noreply.github.com>
Date:   Thu Sep 8 05:36:33 2022 -0600

    macOS CI workflow, dream.py exits with an error, but the workflow com… (#396)

    * macOS CI workflow, dream.py exits with an error, but the workflow completes.

    * Files for testing

    Co-authored-by: James Reynolds <magnsuviri@me.com>
    Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>

commit 711d49ed30
Author: James Reynolds <magnusviri@users.noreply.github.com>
Date:   Thu Sep 8 05:35:08 2022 -0600

    Cache model workflow (#394)

    * Add workflow that caches the model, step 1 for CI

    * Change name of workflow job

    Co-authored-by: James Reynolds <magnsuviri@me.com>
    Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>

commit 7996a30e3a
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Thu Sep 8 07:34:03 2022 -0400

    add auto-creation of mask for inpainting (#438)

    * now use a single init image for both image and mask

    * turn on debugging for now to write out mask and image

    * add back -M option as a fallback

commit a69ca31f34
Author: elliotsayes <elliotsayes@gmail.com>
Date:   Thu Sep 8 15:30:06 2022 +1200

    .gitignore WebUI temp files (#430)

    * Add instructions on how to install alongside pyenv (#393)

    Like probably many others, I have a lot of different virtualenvs, one for each project. Most of them are handled by `pyenv`.
    After installing according to these instructions I had issues with ´pyenv`and `miniconda` fighting over the $PATH of my system.
    But then I stumbled upon this nice solution on SO: https://stackoverflow.com/a/73139031 , upon which I have based my suggested changes.

    It runs perfectly on my M1 setup, with the anaconda setup as a virtual environment handled by pyenv.

    Feel free to incorporate these instructions as you see fit.

    Thanks a million for all your hard work.

    * .gitignore WebUI temp files

    Co-authored-by: Håvard Gulldahl <havard@lurtgjort.no>

commit 5c6b612a72
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Wed Sep 7 22:50:55 2022 -0400

    fix bug that caused same seed to be redisplayed repeatedly

commit 56f155c590
Author: Johan Roxendal <johan@roxendal.com>
Date:   Thu Sep 8 04:50:06 2022 +0200

    added support for parsing run log and displaying images in the frontend init state (#410)

    Co-authored-by: Johan Roxendal <johan.roxendal@litteraturbanken.se>
    Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>

commit 41687746be
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Wed Sep 7 20:24:35 2022 -0400

    added missing initialization of latent_noise to None

commit 171f8db742
Author: Denis Olshin <me@denull.ru>
Date:   Thu Sep 8 03:15:20 2022 +0300

    saving full prompt to metadata when using web ui

commit d7e67b62f0
Author: Denis Olshin <me@denull.ru>
Date:   Thu Sep 8 01:51:47 2022 +0300

    better logic for clicking to make variations

commit d1d044aa87
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Wed Sep 7 17:56:59 2022 -0400

    actual image seed now written into web log rather than -1 (#428)

commit edada042b3
Author: Arturo Mendivil <60411196+artmen1516@users.noreply.github.com>
Date:   Wed Sep 7 10:42:26 2022 -0700

    Improve notebook and add requirements file (#422)

commit 29ab3c2028
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Wed Sep 7 13:28:11 2022 -0400

    disable neonpixel optimizations on M1 hardware (#414)

    * disable neonpixel optimizations on M1 hardware

    * fix typo that was causing random noise images on m1

commit 7670ecc63f
Author: cody <cnmizell@gmail.com>
Date:   Wed Sep 7 12:24:41 2022 -0500

    add more keyboard support on the web server (#391)

    add ability to submit prompts with the "enter" key
    add ability to cancel generations with the "escape" key

commit dd2aedacaf
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Wed Sep 7 13:23:53 2022 -0400

    report VRAM usage stats during initial model loading (#419)

commit f6284777e6
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Tue Sep 6 17:12:39 2022 -0400

    Squashed commit of the following:

    commit 7d1344282d942a33dcecda4d5144fc154ec82915
    Merge: caf4ea3 ebeb556
    Author: Lincoln Stein <lincoln.stein@gmail.com>
    Date:   Mon Sep 5 10:07:27 2022 -0400

        Merge branch 'development' of github.com:WebDev9000/stable-diffusion into WebDev9000-development

    commit ebeb556af9
    Author: Web Dev 9000 <rirath@gmail.com>
    Date:   Sun Sep 4 18:05:15 2022 -0700

        Fixed unintentionally removed lines

    commit ff2c4b9a1b
    Author: Web Dev 9000 <rirath@gmail.com>
    Date:   Sun Sep 4 17:50:13 2022 -0700

        Add ability to recreate variations via image click

    commit c012929cda
    Author: Web Dev 9000 <rirath@gmail.com>
    Date:   Sun Sep 4 14:35:33 2022 -0700

        Add files via upload

    commit 02a6018992
    Author: Web Dev 9000 <rirath@gmail.com>
    Date:   Sun Sep 4 14:35:07 2022 -0700

        Add files via upload

commit eef788981c
Author: Olivier Louvignes <olivier@mg-crea.com>
Date:   Tue Sep 6 12:41:08 2022 +0200

    feat(txt2img): allow from_file to work with len(lines) < batch_size (#349)

commit 720e5cd651
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Mon Sep 5 20:40:10 2022 -0400

    Refactoring simplet2i (#387)

    * start refactoring -not yet functional

    * first phase of refactor done - not sure weighted prompts working

    * Second phase of refactoring. Everything mostly working.
    * The refactoring has moved all the hard-core inference work into
    ldm.dream.generator.*, where there are submodules for txt2img and
    img2img. inpaint will go in there as well.
    * Some additional refactoring will be done soon, but relatively
    minor work.

    * fix -save_orig flag to actually work

    * add @neonsecret attention.py memory optimization

    * remove unneeded imports

    * move token logging into conditioning.py

    * add placeholder version of inpaint; porting in progress

    * fix crash in img2img

    * inpainting working; not tested on variations

    * fix crashes in img2img

    * ported attention.py memory optimization #117 from basujindal branch

    * added @torch_no_grad() decorators to img2img, txt2img, inpaint closures

    * Final commit prior to PR against development
    * fixup crash when generating intermediate images in web UI
    * rename ldm.simplet2i to ldm.generate
    * add backward-compatibility simplet2i shell with deprecation warning

    * add back in mps exception, addresses @vargol comment in #354

    * replaced Conditioning class with exported functions

    * fix wrong type of with_variations attribute during intialization

    * changed "image_iterator()" to "get_make_image()"

    * raise NotImplementedError for calling get_make_image() in parent class

    * Update ldm/generate.py

    better error message

    Co-authored-by: Kevin Gibbons <bakkot@gmail.com>

    * minor stylistic fixes and assertion checks from code review

    * moved get_noise() method into img2img class

    * break get_noise() into two methods, one for txt2img and the other for img2img

    * inpainting works on non-square images now

    * make get_noise() an abstract method in base class

    * much improved inpainting

    Co-authored-by: Kevin Gibbons <bakkot@gmail.com>

commit 1ad2a8e567
Author: thealanle <35761977+thealanle@users.noreply.github.com>
Date:   Mon Sep 5 17:35:04 2022 -0700

    Fix --outdir function for web (#373)

    * Fix --outdir function for web

    * Removed unnecessary hardcoded path

commit 52d8bb2836
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Mon Sep 5 10:31:59 2022 -0400

    Squashed commit of the following:

    commit 0cd48e932f1326e000c46f4140f98697eb9bdc79
    Author: Lincoln Stein <lincoln.stein@gmail.com>
    Date:   Mon Sep 5 10:27:43 2022 -0400

        resolve conflicts with development

    commit d7bc8c12e0
    Author: Scott McMillin <scott@scottmcmillin.com>
    Date:   Sun Sep 4 18:52:09 2022 -0500

        Add title attribute back to img tag

    commit 5397c89184
    Author: Scott McMillin <scott@scottmcmillin.com>
    Date:   Sun Sep 4 13:49:46 2022 -0500

        Remove temp code

    commit 1da080b509
    Author: Scott McMillin <scott@scottmcmillin.com>
    Date:   Sun Sep 4 13:33:56 2022 -0500

        Cleaned up HTML; small style changes; image click opens image; add seed to figcaption beneath image

commit caf4ea3d89
Author: Adam Rice <adam@askadam.io>
Date:   Mon Sep 5 10:05:39 2022 -0400

    Add a 'Remove Image' button to clear the file upload field (#382)

    * added "remove image" button

    * styled a new "remove image" button

    * Update index.js

commit 95c088b303
Author: Kevin Gibbons <bakkot@gmail.com>
Date:   Sun Sep 4 19:04:14 2022 -0700

    Revert "Add CORS headers to dream server to ease integration with third-party web interfaces" (#371)

    This reverts commit 91e826e5f4.

commit a20113d5a3
Author: Kevin Gibbons <bakkot@gmail.com>
Date:   Sun Sep 4 18:59:12 2022 -0700

    put no_grad decorator on make_image closures (#375)

commit 0f93dadd6a
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 4 21:39:15 2022 -0400

    fix several dangling references to --gfpgan option, which no longer exists

commit f4004f660e
Author: tildebyte <337875+tildebyte@users.noreply.github.com>
Date:   Sun Sep 4 19:43:04 2022 -0400

    TOIL(requirements): Split requirements to per-platform (#355)

    * toil(reqs): split requirements to per-platform

    Signed-off-by: Ben Alkov <ben.alkov@gmail.com>

    * toil(reqs): fix for Win and Lin...

    ...allow pip to resolve latest torch, numpy

    Signed-off-by: Ben Alkov <ben.alkov@gmail.com>

    * toil(install): update reqs in Win install notebook

    Signed-off-by: Ben Alkov <ben.alkov@gmail.com>

    Signed-off-by: Ben Alkov <ben.alkov@gmail.com>

commit 4406fd138d
Merge: 5116c81 fd7a72e
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 4 08:23:53 2022 -0400

    Merge branch 'SebastianAigner-main' into development
    Add support for full CORS headers for dream server.

commit fd7a72e147
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 4 08:23:11 2022 -0400

    remove debugging message

commit 3a2be621f3
Merge: 91e826e 5116c81
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sun Sep 4 08:15:51 2022 -0400

    Merge branch 'development' into main

commit 5116c8178c
Author: Justin Wong <1584142+wongjustin99@users.noreply.github.com>
Date:   Sun Sep 4 07:17:58 2022 -0400

    fix save_original flag saving to the same filename (#360)

    * Update README.md with new Anaconda install steps (#347)

    pip3 version did not work for me and this is the recommended way to install Anaconda now it seems

    * fix save_original flag saving to the same filename

    Before this, the `--save_orig` flag was not working. The upscaled/GFPGAN would overwrite the original output image.

    Co-authored-by: greentext2 <112735219+greentext2@users.noreply.github.com>

commit 91e826e5f4
Author: Sebastian Aigner <SebastianAigner@users.noreply.github.com>
Date:   Sun Sep 4 10:22:54 2022 +0200

    Add CORS headers to dream server to ease integration with third-party web interfaces

commit 6266d9e8d6
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 3 15:45:20 2022 -0400

    remove stray debugging message

commit 138956e516
Author: greentext2 <112735219+greentext2@users.noreply.github.com>
Date:   Sat Sep 3 13:38:57 2022 -0500

    Update README.md with new Anaconda install steps (#347)

    pip3 version did not work for me and this is the recommended way to install Anaconda now it seems

commit 60be735e80
Author: Cora Johnson-Roberson <cora.johnson.roberson@gmail.com>
Date:   Sat Sep 3 14:28:34 2022 -0400

    Switch to regular pytorch channel and restore Python 3.10 for Macs. (#301)

    * Switch to regular pytorch channel and restore Python 3.10 for Macs.

    Although pytorch-nightly should in theory be faster, it is currently
    causing increased memory usage and slower iterations:

    https://github.com/lstein/stable-diffusion/pull/283#issuecomment-1234784885

    This changes the environment-mac.yaml file back to the regular pytorch
    channel and moves the `transformers` dep into pip for now (since it
    cannot be satisfied until tokenizers>=0.11 is built for Python 3.10).

    * Specify versions for Pip packages as well.

commit d0d95d3a2a
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 3 14:10:31 2022 -0400

    make initimg appear in web log

commit b90a215000
Merge: 1eee811 6270e31
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 3 13:47:15 2022 -0400

    Merge branch 'prixt-seamless' into development

commit 6270e313b8
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 3 13:46:29 2022 -0400

    add credit to prixt for seamless circular tiling

commit a01b7bdc40
Merge: 1eee811 9d88abe
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 3 13:43:04 2022 -0400

    add web interface for seamless option

commit 1eee8111b9
Merge: 64eca42 fb857f0
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 3 12:33:39 2022 -0400

    Merge branch 'development' of github.com:lstein/stable-diffusion into development

commit 64eca42610
Merge: 9130ad7 21a1f68
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 3 12:33:05 2022 -0400

    Merge branch 'main' into development
    * brings in small documentation fixes that were
    added directly to main during release tweaking.

commit fb857f05ba
Author: Lincoln Stein <lincoln.stein@gmail.com>
Date:   Sat Sep 3 12:07:07 2022 -0400

    fix typo in docs

commit 9d88abe2ea
Author: prixt <paraxite@naver.com>
Date:   Sat Sep 3 22:42:16 2022 +0900

    fixed typo

commit a61e49bc97
Author: prixt <paraxite@naver.com>
Date:   Sat Sep 3 22:39:35 2022 +0900

    * Removed unnecessary code
    * Added description about --seamless

commit 02bee4fdb1
Author: prixt <paraxite@naver.com>
Date:   Sat Sep 3 16:08:03 2022 +0900

    added --seamless tag logging to normalize_prompt

commit d922b53c26
Author: prixt <paraxite@naver.com>
Date:   Sat Sep 3 15:13:31 2022 +0900

    added seamless tiling mode and commands
This commit is contained in:
Lincoln Stein
2022-09-12 14:31:48 -04:00
parent 62863ac586
commit 9d6d728b51
64 changed files with 3836 additions and 2738 deletions
Download Patch File Download Diff File Expand all files Collapse all files

96
ldm/dream/conditioning.py Normal file
View File

@@ -0,0 +1,96 @@
'''
This module handles the generation of the conditioning tensors, including management of
weighted subprompts.
Useful function exports:
get_uc_and_c() get the conditioned and unconditioned latent
split_weighted_subpromopts() split subprompts, normalize and weight them
log_tokenization() print out colour-coded tokens and warn if truncated
'''
import re
import torch
def get_uc_and_c(prompt, model, log_tokens=False, skip_normalize=False):
uc = model.get_learned_conditioning([''])
# get weighted sub-prompts
weighted_subprompts = split_weighted_subprompts(
prompt, skip_normalize
)
if len(weighted_subprompts) > 1:
# i dont know if this is correct.. but it works
c = torch.zeros_like(uc)
# normalize each "sub prompt" and add it
for subprompt, weight in weighted_subprompts:
log_tokenization(subprompt, model, log_tokens)
c = torch.add(
c,
model.get_learned_conditioning([subprompt]),
alpha=weight,
)
else: # just standard 1 prompt
log_tokenization(prompt, model, log_tokens)
c = model.get_learned_conditioning([prompt])
return (uc, c)
def split_weighted_subprompts(text, skip_normalize=False)->list:
"""
grabs all text up to the first occurrence of ':'
uses the grabbed text as a sub-prompt, and takes the value following ':' as weight
if ':' has no value defined, defaults to 1.0
repeats until no text remaining
"""
prompt_parser = re.compile("""
(?P<prompt> # capture group for 'prompt'
(?:\\\:|[^:])+ # match one or more non ':' characters or escaped colons '\:'
) # end 'prompt'
(?: # non-capture group
:+ # match one or more ':' characters
(?P<weight> # capture group for 'weight'
-?\d+(?:\.\d+)? # match positive or negative integer or decimal number
)? # end weight capture group, make optional
\s* # strip spaces after weight
| # OR
$ # else, if no ':' then match end of line
) # end non-capture group
""", re.VERBOSE)
parsed_prompts = [(match.group("prompt").replace("\\:", ":"), float(
match.group("weight") or 1)) for match in re.finditer(prompt_parser, text)]
if skip_normalize:
return parsed_prompts
weight_sum = sum(map(lambda x: x[1], parsed_prompts))
if weight_sum == 0:
print(
"Warning: Subprompt weights add up to zero. Discarding and using even weights instead.")
equal_weight = 1 / len(parsed_prompts)
return [(x[0], equal_weight) for x in parsed_prompts]
return [(x[0], x[1] / weight_sum) for x in parsed_prompts]
# shows how the prompt is tokenized
# usually tokens have '</w>' to indicate end-of-word,
# but for readability it has been replaced with ' '
def log_tokenization(text, model, log=False):
if not log:
return
tokens = model.cond_stage_model.tokenizer._tokenize(text)
tokenized = ""
discarded = ""
usedTokens = 0
totalTokens = len(tokens)
for i in range(0, totalTokens):
token = tokens[i].replace('</w>', ' ')
# alternate color
s = (usedTokens % 6) + 1
if i < model.cond_stage_model.max_length:
tokenized = tokenized + f"\x1b[0;3{s};40m{token}"
usedTokens += 1
else: # over max token length
discarded = discarded + f"\x1b[0;3{s};40m{token}"
print(f"\n>> Tokens ({usedTokens}):\n{tokenized}\x1b[0m")
if discarded != "":
print(
f">> Tokens Discarded ({totalTokens-usedTokens}):\n{discarded}\x1b[0m"
)

11
ldm/dream/devices.py
View File

@@ -1,4 +1,6 @@
import torch
from torch import autocast
from contextlib import contextmanager, nullcontext
def choose_torch_device() -> str:
'''Convenience routine for guessing which GPU device to run model on'''
@@ -8,10 +10,11 @@ def choose_torch_device() -> str:
return 'mps'
return 'cpu'
def choose_autocast_device(device) -> str:
def choose_autocast_device(device):
'''Returns an autocast compatible device from a torch device'''
device_type = device.type # this returns 'mps' on M1
# autocast only supports cuda or cpu
if device_type not in ('cuda','cpu'):
return 'cpu'
return device_type
if device_type in ('cuda','cpu'):
return device_type,autocast
else:
return 'cpu',nullcontext

4
ldm/dream/generator/__init__.py Normal file
View File

@@ -0,0 +1,4 @@
'''
Initialization file for the ldm.dream.generator package
'''
from .base import Generator

158
ldm/dream/generator/base.py Normal file
View File

@@ -0,0 +1,158 @@
'''
Base class for ldm.dream.generator.*
including img2img, txt2img, and inpaint
'''
import torch
import numpy as np
import random
from tqdm import tqdm, trange
from PIL import Image
from einops import rearrange, repeat
from pytorch_lightning import seed_everything
from ldm.dream.devices import choose_autocast_device
downsampling = 8
class Generator():
def __init__(self,model):
self.model = model
self.seed = None
self.latent_channels = model.channels
self.downsampling_factor = downsampling # BUG: should come from model or config
self.variation_amount = 0
self.with_variations = []
# this is going to be overridden in img2img.py, txt2img.py and inpaint.py
def get_make_image(self,prompt,**kwargs):
"""
Returns a function returning an image derived from the prompt and the initial image
Return value depends on the seed at the time you call it
"""
raise NotImplementedError("image_iterator() must be implemented in a descendent class")
def set_variation(self, seed, variation_amount, with_variations):
self.seed = seed
self.variation_amount = variation_amount
self.with_variations = with_variations
def generate(self,prompt,init_image,width,height,iterations=1,seed=None,
image_callback=None, step_callback=None,
**kwargs):
device_type,scope = choose_autocast_device(self.model.device)
make_image = self.get_make_image(
prompt,
init_image = init_image,
width = width,
height = height,
step_callback = step_callback,
**kwargs
)
results = []
seed = seed if seed else self.new_seed()
seed, initial_noise = self.generate_initial_noise(seed, width, height)
with scope(device_type), self.model.ema_scope():
for n in trange(iterations, desc='Generating'):
x_T = None
if self.variation_amount > 0:
seed_everything(seed)
target_noise = self.get_noise(width,height)
x_T = self.slerp(self.variation_amount, initial_noise, target_noise)
elif initial_noise is not None:
# i.e. we specified particular variations
x_T = initial_noise
else:
seed_everything(seed)
if self.model.device.type == 'mps':
x_T = self.get_noise(width,height)
# make_image will do the equivalent of get_noise itself
image = make_image(x_T)
results.append([image, seed])
if image_callback is not None:
image_callback(image, seed)
seed = self.new_seed()
return results
def sample_to_image(self,samples):
"""
Returns a function returning an image derived from the prompt and the initial image
Return value depends on the seed at the time you call it
"""
x_samples = self.model.decode_first_stage(samples)
x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0)
if len(x_samples) != 1:
raise Exception(
f'>> expected to get a single image, but got {len(x_samples)}')
x_sample = 255.0 * rearrange(
x_samples[0].cpu().numpy(), 'c h w -> h w c'
)
return Image.fromarray(x_sample.astype(np.uint8))
def generate_initial_noise(self, seed, width, height):
initial_noise = None
if self.variation_amount > 0 or len(self.with_variations) > 0:
# use fixed initial noise plus random noise per iteration
seed_everything(seed)
initial_noise = self.get_noise(width,height)
for v_seed, v_weight in self.with_variations:
seed = v_seed
seed_everything(seed)
next_noise = self.get_noise(width,height)
initial_noise = self.slerp(v_weight, initial_noise, next_noise)
if self.variation_amount > 0:
random.seed() # reset RNG to an actually random state, so we can get a random seed for variations
seed = random.randrange(0,np.iinfo(np.uint32).max)
return (seed, initial_noise)
else:
return (seed, None)
# returns a tensor filled with random numbers from a normal distribution
def get_noise(self,width,height):
"""
Returns a tensor filled with random numbers, either form a normal distribution
(txt2img) or from the latent image (img2img, inpaint)
"""
raise NotImplementedError("get_noise() must be implemented in a descendent class")
def new_seed(self):
self.seed = random.randrange(0, np.iinfo(np.uint32).max)
return self.seed
def slerp(self, t, v0, v1, DOT_THRESHOLD=0.9995):
'''
Spherical linear interpolation
Args:
t (float/np.ndarray): Float value between 0.0 and 1.0
v0 (np.ndarray): Starting vector
v1 (np.ndarray): Final vector
DOT_THRESHOLD (float): Threshold for considering the two vectors as
colineal. Not recommended to alter this.
Returns:
v2 (np.ndarray): Interpolation vector between v0 and v1
'''
inputs_are_torch = False
if not isinstance(v0, np.ndarray):
inputs_are_torch = True
v0 = v0.detach().cpu().numpy()
if not isinstance(v1, np.ndarray):
inputs_are_torch = True
v1 = v1.detach().cpu().numpy()
dot = np.sum(v0 * v1 / (np.linalg.norm(v0) * np.linalg.norm(v1)))
if np.abs(dot) > DOT_THRESHOLD:
v2 = (1 - t) * v0 + t * v1
else:
theta_0 = np.arccos(dot)
sin_theta_0 = np.sin(theta_0)
theta_t = theta_0 * t
sin_theta_t = np.sin(theta_t)
s0 = np.sin(theta_0 - theta_t) / sin_theta_0
s1 = sin_theta_t / sin_theta_0
v2 = s0 * v0 + s1 * v1
if inputs_are_torch:
v2 = torch.from_numpy(v2).to(self.model.device)
return v2

72
ldm/dream/generator/img2img.py Normal file
View File

@@ -0,0 +1,72 @@
'''
ldm.dream.generator.txt2img descends from ldm.dream.generator
'''
import torch
import numpy as np
from ldm.dream.devices import choose_autocast_device
from ldm.dream.generator.base import Generator
from ldm.models.diffusion.ddim import DDIMSampler
class Img2Img(Generator):
def __init__(self,model):
super().__init__(model)
self.init_latent = None # by get_noise()
@torch.no_grad()
def get_make_image(self,prompt,sampler,steps,cfg_scale,ddim_eta,
conditioning,init_image,strength,step_callback=None,**kwargs):
"""
Returns a function returning an image derived from the prompt and the initial image
Return value depends on the seed at the time you call it.
"""
# PLMS sampler not supported yet, so ignore previous sampler
if not isinstance(sampler,DDIMSampler):
print(
f">> sampler '{sampler.__class__.__name__}' is not yet supported. Using DDIM sampler"
)
sampler = DDIMSampler(self.model, device=self.model.device)
sampler.make_schedule(
ddim_num_steps=steps, ddim_eta=ddim_eta, verbose=False
)
device_type,scope = choose_autocast_device(self.model.device)
with scope(device_type):
self.init_latent = self.model.get_first_stage_encoding(
self.model.encode_first_stage(init_image)
) # move to latent space
t_enc = int(strength * steps)
uc, c = conditioning
@torch.no_grad()
def make_image(x_T):
# encode (scaled latent)
z_enc = sampler.stochastic_encode(
self.init_latent,
torch.tensor([t_enc]).to(self.model.device),
noise=x_T
)
# decode it
samples = sampler.decode(
z_enc,
c,
t_enc,
img_callback = step_callback,
unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc,
)
return self.sample_to_image(samples)
return make_image
def get_noise(self,width,height):
device = self.model.device
init_latent = self.init_latent
assert init_latent is not None,'call to get_noise() when init_latent not set'
if device.type == 'mps':
return torch.randn_like(init_latent, device='cpu').to(device)
else:
return torch.randn_like(init_latent, device=device)

77
ldm/dream/generator/inpaint.py Normal file
View File

@@ -0,0 +1,77 @@
'''
ldm.dream.generator.inpaint descends from ldm.dream.generator
'''
import torch
import numpy as np
from einops import rearrange, repeat
from ldm.dream.devices import choose_autocast_device
from ldm.dream.generator.img2img import Img2Img
from ldm.models.diffusion.ddim import DDIMSampler
class Inpaint(Img2Img):
def __init__(self,model):
self.init_latent = None
super().__init__(model)
@torch.no_grad()
def get_make_image(self,prompt,sampler,steps,cfg_scale,ddim_eta,
conditioning,init_image,mask_image,strength,
step_callback=None,**kwargs):
"""
Returns a function returning an image derived from the prompt and
the initial image + mask. Return value depends on the seed at
the time you call it. kwargs are 'init_latent' and 'strength'
"""
mask_image = mask_image[0][0].unsqueeze(0).repeat(4,1,1).unsqueeze(0)
mask_image = repeat(mask_image, '1 ... -> b ...', b=1)
# PLMS sampler not supported yet, so ignore previous sampler
if not isinstance(sampler,DDIMSampler):
print(
f">> sampler '{sampler.__class__.__name__}' is not yet supported. Using DDIM sampler"
)
sampler = DDIMSampler(self.model, device=self.model.device)
sampler.make_schedule(
ddim_num_steps=steps, ddim_eta=ddim_eta, verbose=False
)
device_type,scope = choose_autocast_device(self.model.device)
with scope(device_type):
self.init_latent = self.model.get_first_stage_encoding(
self.model.encode_first_stage(init_image)
) # move to latent space
t_enc = int(strength * steps)
uc, c = conditioning
print(f">> target t_enc is {t_enc} steps")
@torch.no_grad()
def make_image(x_T):
# encode (scaled latent)
z_enc = sampler.stochastic_encode(
self.init_latent,
torch.tensor([t_enc]).to(self.model.device),
noise=x_T
)
# decode it
samples = sampler.decode(
z_enc,
c,
t_enc,
img_callback = step_callback,
unconditional_guidance_scale = cfg_scale,
unconditional_conditioning = uc,
mask = mask_image,
init_latent = self.init_latent
)
return self.sample_to_image(samples)
return make_image

61
ldm/dream/generator/txt2img.py Normal file
View File

@@ -0,0 +1,61 @@
'''
ldm.dream.generator.txt2img inherits from ldm.dream.generator
'''
import torch
import numpy as np
from ldm.dream.generator.base import Generator
class Txt2Img(Generator):
def __init__(self,model):
super().__init__(model)
@torch.no_grad()
def get_make_image(self,prompt,sampler,steps,cfg_scale,ddim_eta,
conditioning,width,height,step_callback=None,**kwargs):
"""
Returns a function returning an image derived from the prompt and the initial image
Return value depends on the seed at the time you call it
kwargs are 'width' and 'height'
"""
uc, c = conditioning
@torch.no_grad()
def make_image(x_T):
shape = [
self.latent_channels,
height // self.downsampling_factor,
width // self.downsampling_factor,
]
samples, _ = sampler.sample(
batch_size = 1,
S = steps,
x_T = x_T,
conditioning = c,
shape = shape,
verbose = False,
unconditional_guidance_scale = cfg_scale,
unconditional_conditioning = uc,
eta = ddim_eta,
img_callback = step_callback
)
return self.sample_to_image(samples)
return make_image
# returns a tensor filled with random numbers from a normal distribution
def get_noise(self,width,height):
device = self.model.device
if device.type == 'mps':
return torch.randn([1,
self.latent_channels,
height // self.downsampling_factor,
width // self.downsampling_factor],
device='cpu').to(device)
else:
return torch.randn([1,
self.latent_channels,
height // self.downsampling_factor,
width // self.downsampling_factor],
device=device)

6
ldm/dream/pngwriter.py
View File

@@ -59,6 +59,10 @@ class PromptFormatter:
switches.append(f'-H{opt.height or t2i.height}')
switches.append(f'-C{opt.cfg_scale or t2i.cfg_scale}')
switches.append(f'-A{opt.sampler_name or t2i.sampler_name}')
# to do: put model name into the t2i object
# switches.append(f'--model{t2i.model_name}')
if opt.seamless or t2i.seamless:
switches.append(f'--seamless')
if opt.init_img:
switches.append(f'-I{opt.init_img}')
if opt.fit:
@@ -74,6 +78,4 @@ class PromptFormatter:
if opt.with_variations:
formatted_variations = ','.join(f'{seed}:{weight}' for seed, weight in opt.with_variations)
switches.append(f'-V{formatted_variations}')
if t2i.full_precision:
switches.append('-F')
return ' '.join(switches)

8
ldm/dream/readline.py
View File

@@ -22,7 +22,7 @@ class Completer:
def complete(self, text, state):
buffer = readline.get_line_buffer()
if text.startswith(('-I', '--init_img')):
if text.startswith(('-I', '--init_img','-M','--init_mask')):
return self._path_completions(text, state, ('.png','.jpg','.jpeg'))
if buffer.strip().endswith('cd') or text.startswith(('.', '/')):
@@ -48,10 +48,15 @@ class Completer:
def _path_completions(self, text, state, extensions):
# get the path so far
# TODO: replace this mess with a regular expression match
if text.startswith('-I'):
path = text.replace('-I', '', 1).lstrip()
elif text.startswith('--init_img='):
path = text.replace('--init_img=', '', 1).lstrip()
elif text.startswith('--init_mask='):
path = text.replace('--init_mask=', '', 1).lstrip()
elif text.startswith('-M'):
path = text.replace('-M', '', 1).lstrip()
else:
path = text
@@ -94,6 +99,7 @@ if readline_available:
'--grid','-g',
'--individual','-i',
'--init_img','-I',
'--init_mask','-M',
'--strength','-f',
'--variants','-v',
'--outdir','-o',

165
ldm/dream/server.py
View File

@@ -1,16 +1,65 @@
import argparse
import json
import base64
import mimetypes
import os
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
from ldm.dream.pngwriter import PngWriter
from ldm.dream.pngwriter import PngWriter, PromptFormatter
from threading import Event
def build_opt(post_data, seed, gfpgan_model_exists):
opt = argparse.Namespace()
setattr(opt, 'prompt', post_data['prompt'])
setattr(opt, 'init_img', post_data['initimg'])
setattr(opt, 'strength', float(post_data['strength']))
setattr(opt, 'iterations', int(post_data['iterations']))
setattr(opt, 'steps', int(post_data['steps']))
setattr(opt, 'width', int(post_data['width']))
setattr(opt, 'height', int(post_data['height']))
setattr(opt, 'seamless', 'seamless' in post_data)
setattr(opt, 'fit', 'fit' in post_data)
setattr(opt, 'mask', 'mask' in post_data)
setattr(opt, 'invert_mask', 'invert_mask' in post_data)
setattr(opt, 'cfg_scale', float(post_data['cfg_scale']))
setattr(opt, 'sampler_name', post_data['sampler_name'])
setattr(opt, 'gfpgan_strength', float(post_data['gfpgan_strength']) if gfpgan_model_exists else 0)
setattr(opt, 'upscale', [int(post_data['upscale_level']), float(post_data['upscale_strength'])] if post_data['upscale_level'] != '' else None)
setattr(opt, 'progress_images', 'progress_images' in post_data)
setattr(opt, 'seed', None if int(post_data['seed']) == -1 else int(post_data['seed']))
setattr(opt, 'variation_amount', float(post_data['variation_amount']) if int(post_data['seed']) != -1 else 0)
setattr(opt, 'with_variations', [])
broken = False
if int(post_data['seed']) != -1 and post_data['with_variations'] != '':
for part in post_data['with_variations'].split(','):
seed_and_weight = part.split(':')
if len(seed_and_weight) != 2:
print(f'could not parse with_variation part "{part}"')
broken = True
break
try:
seed = int(seed_and_weight[0])
weight = float(seed_and_weight[1])
except ValueError:
print(f'could not parse with_variation part "{part}"')
broken = True
break
opt.with_variations.append([seed, weight])
if broken:
raise CanceledException
if len(opt.with_variations) == 0:
opt.with_variations = None
return opt
class CanceledException(Exception):
pass
class DreamServer(BaseHTTPRequestHandler):
model = None
outdir = None
canceled = Event()
def do_GET(self):
@@ -30,6 +79,23 @@ class DreamServer(BaseHTTPRequestHandler):
'gfpgan_model_exists': gfpgan_model_exists
}
self.wfile.write(bytes("let config = " + json.dumps(config) + ";\n", "utf-8"))
elif self.path == "/run_log.json":
self.send_response(200)
self.send_header("Content-type", "application/json")
self.end_headers()
output = []
log_file = os.path.join(self.outdir, "dream_web_log.txt")
if os.path.exists(log_file):
with open(log_file, "r") as log:
for line in log:
url, config = line.split(": {", maxsplit=1)
config = json.loads("{" + config)
config["url"] = url.lstrip(".")
if os.path.exists(url):
output.append(config)
self.wfile.write(bytes(json.dumps({"run_log": output}), "utf-8"))
elif self.path == "/cancel":
self.canceled.set()
self.send_response(200)
@@ -63,34 +129,19 @@ class DreamServer(BaseHTTPRequestHandler):
content_length = int(self.headers['Content-Length'])
post_data = json.loads(self.rfile.read(content_length))
prompt = post_data['prompt']
initimg = post_data['initimg']
strength = float(post_data['strength'])
iterations = int(post_data['iterations'])
steps = int(post_data['steps'])
width = int(post_data['width'])
height = int(post_data['height'])
fit = 'fit' in post_data
cfgscale = float(post_data['cfgscale'])
sampler_name = post_data['sampler']
gfpgan_strength = float(post_data['gfpgan_strength']) if gfpgan_model_exists else 0
upscale_level = post_data['upscale_level']
upscale_strength = post_data['upscale_strength']
upscale = [int(upscale_level),float(upscale_strength)] if upscale_level != '' else None
progress_images = 'progress_images' in post_data
seed = self.model.seed if int(post_data['seed']) == -1 else int(post_data['seed'])
opt = build_opt(post_data, self.model.seed, gfpgan_model_exists)
self.canceled.clear()
print(f">> Request to generate with prompt: {prompt}")
print(f">> Request to generate with prompt: {opt.prompt}")
# In order to handle upscaled images, the PngWriter needs to maintain state
# across images generated by each call to prompt2img(), so we define it in
# the outer scope of image_done()
config = post_data.copy() # Shallow copy
config['initimg'] = ''
config['initimg'] = config.pop('initimg_name', '')
images_generated = 0 # helps keep track of when upscaling is started
images_upscaled = 0 # helps keep track of when upscaling is completed
pngwriter = PngWriter("./outputs/img-samples/")
pngwriter = PngWriter(self.outdir)
prefix = pngwriter.unique_prefix()
# if upscaling is requested, then this will be called twice, once when
@@ -99,11 +150,24 @@ class DreamServer(BaseHTTPRequestHandler):
# entry should not be inserted into the image list.
def image_done(image, seed, upscaled=False):
name = f'{prefix}.{seed}.png'
path = pngwriter.save_image_and_prompt_to_png(image, f'{prompt} -S{seed}', name)
iter_opt = argparse.Namespace(**vars(opt)) # copy
if opt.variation_amount > 0:
this_variation = [[seed, opt.variation_amount]]
if opt.with_variations is None:
iter_opt.with_variations = this_variation
else:
iter_opt.with_variations = opt.with_variations + this_variation
iter_opt.variation_amount = 0
elif opt.with_variations is None:
iter_opt.seed = seed
normalized_prompt = PromptFormatter(self.model, iter_opt).normalize_prompt()
path = pngwriter.save_image_and_prompt_to_png(image, f'{normalized_prompt} -S{iter_opt.seed}', name)
if int(config['seed']) == -1:
config['seed'] = seed
# Append post_data to log, but only once!
if not upscaled:
with open("./outputs/img-samples/dream_web_log.txt", "a") as log:
with open(os.path.join(self.outdir, "dream_web_log.txt"), "a") as log:
log.write(f"{path}: {json.dumps(config)}\n")
self.wfile.write(bytes(json.dumps(
@@ -111,27 +175,27 @@ class DreamServer(BaseHTTPRequestHandler):
) + '\n',"utf-8"))
# control state of the "postprocessing..." message
upscaling_requested = upscale or gfpgan_strength>0
upscaling_requested = opt.upscale or opt.gfpgan_strength > 0
nonlocal images_generated # NB: Is this bad python style? It is typical usage in a perl closure.
nonlocal images_upscaled # NB: Is this bad python style? It is typical usage in a perl closure.
if upscaled:
images_upscaled += 1
else:
images_generated +=1
images_generated += 1
if upscaling_requested:
action = None
if images_generated >= iterations:
if images_upscaled < iterations:
if images_generated >= opt.iterations:
if images_upscaled < opt.iterations:
action = 'upscaling-started'
else:
action = 'upscaling-done'
if action:
x = images_upscaled+1
x = images_upscaled + 1
self.wfile.write(bytes(json.dumps(
{'event':action,'processed_file_cnt':f'{x}/{iterations}'}
{'event': action, 'processed_file_cnt': f'{x}/{opt.iterations}'}
) + '\n',"utf-8"))
step_writer = PngWriter('./outputs/intermediates/')
step_writer = PngWriter(os.path.join(self.outdir, "intermediates"))
step_index = 1
def image_progress(sample, step):
if self.canceled.is_set():
@@ -141,10 +205,10 @@ class DreamServer(BaseHTTPRequestHandler):
# since rendering images is moderately expensive, only render every 5th image
# and don't bother with the last one, since it'll render anyway
nonlocal step_index
if progress_images and step % 5 == 0 and step < steps - 1:
image = self.model._sample_to_image(sample)
name = f'{prefix}.{seed}.{step_index}.png'
metadata = f'{prompt} -S{seed} [intermediate]'
if opt.progress_images and step % 5 == 0 and step < opt.steps - 1:
image = self.model.sample_to_image(sample)
name = f'{prefix}.{opt.seed}.{step_index}.png'
metadata = f'{opt.prompt} -S{opt.seed} [intermediate]'
path = step_writer.save_image_and_prompt_to_png(image, metadata, name)
step_index += 1
self.wfile.write(bytes(json.dumps(
@@ -152,43 +216,20 @@ class DreamServer(BaseHTTPRequestHandler):
) + '\n',"utf-8"))
try:
if initimg is None:
if opt.init_img is None:
# Run txt2img
self.model.prompt2image(prompt,
iterations=iterations,
cfg_scale = cfgscale,
width = width,
height = height,
seed = seed,
steps = steps,
gfpgan_strength = gfpgan_strength,
upscale = upscale,
sampler_name = sampler_name,
step_callback=image_progress,
image_callback=image_done)
self.model.prompt2image(**vars(opt), step_callback=image_progress, image_callback=image_done)
else:
# Decode initimg as base64 to temp file
with open("./img2img-tmp.png", "wb") as f:
initimg = initimg.split(",")[1] # Ignore mime type
initimg = opt.init_img.split(",")[1] # Ignore mime type
f.write(base64.b64decode(initimg))
opt1 = argparse.Namespace(**vars(opt))
opt1.init_img = "./img2img-tmp.png"
try:
# Run img2img
self.model.prompt2image(prompt,
init_img = "./img2img-tmp.png",
strength = strength,
iterations = iterations,
cfg_scale = cfgscale,
seed = seed,
steps = steps,
sampler_name = sampler_name,
width = width,
height = height,
fit = fit,
gfpgan_strength=gfpgan_strength,
upscale = upscale,
step_callback=image_progress,
image_callback=image_done)
self.model.prompt2image(**vars(opt1), step_callback=image_progress, image_callback=image_done)
finally:
# Remove the temp file
os.remove("./img2img-tmp.png")

695
ldm/generate.py Normal file
View File

@@ -0,0 +1,695 @@
# Copyright (c) 2022 Lincoln D. Stein (https://github.com/lstein)
# Derived from source code carrying the following copyrights
# Copyright (c) 2022 Machine Vision and Learning Group, LMU Munich
# Copyright (c) 2022 Robin Rombach and Patrick Esser and contributors
import torch
import numpy as np
import random
import os
import time
import re
import sys
import traceback
import transformers
from omegaconf import OmegaConf
from PIL import Image, ImageOps
from torch import nn
from pytorch_lightning import seed_everything
from ldm.util import instantiate_from_config
from ldm.models.diffusion.ddim import DDIMSampler
from ldm.models.diffusion.plms import PLMSSampler
from ldm.models.diffusion.ksampler import KSampler
from ldm.dream.pngwriter import PngWriter
from ldm.dream.image_util import InitImageResizer
from ldm.dream.devices import choose_torch_device
from ldm.dream.conditioning import get_uc_and_c
"""Simplified text to image API for stable diffusion/latent diffusion
Example Usage:
from ldm.generate import Generate
# Create an object with default values
gr = Generate()
# do the slow model initialization
gr.load_model()
# Do the fast inference & image generation. Any options passed here
# override the default values assigned during class initialization
# Will call load_model() if the model was not previously loaded and so
# may be slow at first.
# The method returns a list of images. Each row of the list is a sub-list of [filename,seed]
results = gr.prompt2png(prompt = "an astronaut riding a horse",
outdir = "./outputs/samples",
iterations = 3)
for row in results:
print(f'filename={row[0]}')
print(f'seed ={row[1]}')
# Same thing, but using an initial image.
results = gr.prompt2png(prompt = "an astronaut riding a horse",
outdir = "./outputs/,
iterations = 3,
init_img = "./sketches/horse+rider.png")
for row in results:
print(f'filename={row[0]}')
print(f'seed ={row[1]}')
# Same thing, but we return a series of Image objects, which lets you manipulate them,
# combine them, and save them under arbitrary names
results = gr.prompt2image(prompt = "an astronaut riding a horse"
outdir = "./outputs/")
for row in results:
im = row[0]
seed = row[1]
im.save(f'./outputs/samples/an_astronaut_riding_a_horse-{seed}.png')
im.thumbnail(100,100).save('./outputs/samples/astronaut_thumb.jpg')
Note that the old txt2img() and img2img() calls are deprecated but will
still work.
The full list of arguments to Generate() are:
gr = Generate(
weights = path to model weights ('models/ldm/stable-diffusion-v1/model.ckpt')
config = path to model configuraiton ('configs/stable-diffusion/v1-inference.yaml')
iterations = <integer> // how many times to run the sampling (1)
steps = <integer> // 50
seed = <integer> // current system time
sampler_name= ['ddim', 'k_dpm_2_a', 'k_dpm_2', 'k_euler_a', 'k_euler', 'k_heun', 'k_lms', 'plms'] // k_lms
grid = <boolean> // false
width = <integer> // image width, multiple of 64 (512)
height = <integer> // image height, multiple of 64 (512)
cfg_scale = <float> // condition-free guidance scale (7.5)
)
"""
class Generate:
"""Generate class
Stores default values for multiple configuration items
"""
def __init__(
self,
iterations = 1,
steps = 50,
cfg_scale = 7.5,
weights = 'models/ldm/stable-diffusion-v1/model.ckpt',
config = 'configs/stable-diffusion/v1-inference.yaml',
grid = False,
width = 512,
height = 512,
sampler_name = 'k_lms',
ddim_eta = 0.0, # deterministic
precision = 'autocast',
full_precision = False,
strength = 0.75, # default in scripts/img2img.py
seamless = False,
embedding_path = None,
device_type = 'cuda',
ignore_ctrl_c = False,
):
self.iterations = iterations
self.width = width
self.height = height
self.steps = steps
self.cfg_scale = cfg_scale
self.weights = weights
self.config = config
self.sampler_name = sampler_name
self.grid = grid
self.ddim_eta = ddim_eta
self.precision = precision
self.full_precision = True if choose_torch_device() == 'mps' else full_precision
self.strength = strength
self.seamless = seamless
self.embedding_path = embedding_path
self.device_type = device_type
self.ignore_ctrl_c = ignore_ctrl_c # note, this logic probably doesn't belong here...
self.model = None # empty for now
self.sampler = None
self.device = None
self.generators = {}
self.base_generator = None
self.seed = None
if device_type == 'cuda' and not torch.cuda.is_available():
device_type = choose_torch_device()
print(">> cuda not available, using device", device_type)
self.device = torch.device(device_type)
# for VRAM usage statistics
device_type = choose_torch_device()
self.session_peakmem = torch.cuda.max_memory_allocated() if device_type == 'cuda' else None
transformers.logging.set_verbosity_error()
def prompt2png(self, prompt, outdir, **kwargs):
"""
Takes a prompt and an output directory, writes out the requested number
of PNG files, and returns an array of [[filename,seed],[filename,seed]...]
Optional named arguments are the same as those passed to Generate and prompt2image()
"""
results = self.prompt2image(prompt, **kwargs)
pngwriter = PngWriter(outdir)
prefix = pngwriter.unique_prefix()
outputs = []
for image, seed in results:
name = f'{prefix}.{seed}.png'
path = pngwriter.save_image_and_prompt_to_png(
image, f'{prompt} -S{seed}', name)
outputs.append([path, seed])
return outputs
def txt2img(self, prompt, **kwargs):
outdir = kwargs.pop('outdir', 'outputs/img-samples')
return self.prompt2png(prompt, outdir, **kwargs)
def img2img(self, prompt, **kwargs):
outdir = kwargs.pop('outdir', 'outputs/img-samples')
assert (
'init_img' in kwargs
), 'call to img2img() must include the init_img argument'
return self.prompt2png(prompt, outdir, **kwargs)
def prompt2image(
self,
# these are common
prompt,
iterations = None,
steps = None,
seed = None,
cfg_scale = None,
ddim_eta = None,
skip_normalize = False,
image_callback = None,
step_callback = None,
width = None,
height = None,
sampler_name = None,
seamless = False,
log_tokenization= False,
with_variations = None,
variation_amount = 0.0,
# these are specific to img2img and inpaint
init_img = None,
init_mask = None,
fit = False,
strength = None,
# these are specific to GFPGAN/ESRGAN
gfpgan_strength= 0,
save_original = False,
upscale = None,
**args,
): # eat up additional cruft
"""
ldm.generate.prompt2image() is the common entry point for txt2img() and img2img()
It takes the following arguments:
prompt // prompt string (no default)
iterations // iterations (1); image count=iterations
steps // refinement steps per iteration
seed // seed for random number generator
width // width of image, in multiples of 64 (512)
height // height of image, in multiples of 64 (512)
cfg_scale // how strongly the prompt influences the image (7.5) (must be >1)
seamless // whether the generated image should tile
init_img // path to an initial image
strength // strength for noising/unnoising init_img. 0.0 preserves image exactly, 1.0 replaces it completely
gfpgan_strength // strength for GFPGAN. 0.0 preserves image exactly, 1.0 replaces it completely
ddim_eta // image randomness (eta=0.0 means the same seed always produces the same image)
step_callback // a function or method that will be called each step
image_callback // a function or method that will be called each time an image is generated
with_variations // a weighted list [(seed_1, weight_1), (seed_2, weight_2), ...] of variations which should be applied before doing any generation
variation_amount // optional 0-1 value to slerp from -S noise to random noise (allows variations on an image)
To use the step callback, define a function that receives two arguments:
- Image GPU data
- The step number
To use the image callback, define a function of method that receives two arguments, an Image object
and the seed. You can then do whatever you like with the image, including converting it to
different formats and manipulating it. For example:
def process_image(image,seed):
image.save(f{'images/seed.png'})
The callback used by the prompt2png() can be found in ldm/dream_util.py. It contains code
to create the requested output directory, select a unique informative name for each image, and
write the prompt into the PNG metadata.
"""
# TODO: convert this into a getattr() loop
steps = steps or self.steps
width = width or self.width
height = height or self.height
seamless = seamless or self.seamless
cfg_scale = cfg_scale or self.cfg_scale
ddim_eta = ddim_eta or self.ddim_eta
iterations = iterations or self.iterations
strength = strength or self.strength
self.seed = seed
self.log_tokenization = log_tokenization
with_variations = [] if with_variations is None else with_variations
model = (
self.load_model()
) # will instantiate the model or return it from cache
for m in model.modules():
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
m.padding_mode = 'circular' if seamless else m._orig_padding_mode
assert cfg_scale > 1.0, 'CFG_Scale (-C) must be >1.0'
assert (
0.0 < strength < 1.0
), 'img2img and inpaint strength can only work with 0.0 < strength < 1.0'
assert (
0.0 <= variation_amount <= 1.0
), '-v --variation_amount must be in [0.0, 1.0]'
# check this logic - doesn't look right
if len(with_variations) > 0 or variation_amount > 1.0:
assert seed is not None,\
'seed must be specified when using with_variations'
if variation_amount == 0.0:
assert iterations == 1,\
'when using --with_variations, multiple iterations are only possible when using --variation_amount'
assert all(0 <= weight <= 1 for _, weight in with_variations),\
f'variation weights must be in [0.0, 1.0]: got {[weight for _, weight in with_variations]}'
width, height, _ = self._resolution_check(width, height, log=True)
if sampler_name and (sampler_name != self.sampler_name):
self.sampler_name = sampler_name
self._set_sampler()
tic = time.time()
if torch.cuda.is_available():
torch.cuda.reset_peak_memory_stats()
results = list()
init_image = None
mask_image = None
try:
uc, c = get_uc_and_c(
prompt, model=self.model,
skip_normalize=skip_normalize,
log_tokens=self.log_tokenization
)
(init_image,mask_image) = self._make_images(init_img,init_mask, width, height, fit)
if (init_image is not None) and (mask_image is not None):
generator = self._make_inpaint()
elif init_image is not None:
generator = self._make_img2img()
else:
generator = self._make_txt2img()
generator.set_variation(self.seed, variation_amount, with_variations)
results = generator.generate(
prompt,
iterations = iterations,
seed = self.seed,
sampler = self.sampler,
steps = steps,
cfg_scale = cfg_scale,
conditioning = (uc,c),
ddim_eta = ddim_eta,
image_callback = image_callback, # called after the final image is generated
step_callback = step_callback, # called after each intermediate image is generated
width = width,
height = height,
init_image = init_image, # notice that init_image is different from init_img
mask_image = mask_image,
strength = strength,
)
if upscale is not None or gfpgan_strength > 0:
self.upscale_and_reconstruct(results,
upscale = upscale,
strength = gfpgan_strength,
save_original = save_original,
image_callback = image_callback)
except KeyboardInterrupt:
print('*interrupted*')
if not self.ignore_ctrl_c:
raise KeyboardInterrupt
print(
'>> Partial results will be returned; if --grid was requested, nothing will be returned.'
)
except RuntimeError as e:
print(traceback.format_exc(), file=sys.stderr)
print('>> Could not generate image.')
toc = time.time()
print('>> Usage stats:')
print(
f'>> {len(results)} image(s) generated in', '%4.2fs' % (toc - tic)
)
if torch.cuda.is_available() and self.device.type == 'cuda':
print(
f'>> Max VRAM used for this generation:',
'%4.2fG.' % (torch.cuda.max_memory_allocated() / 1e9),
'Current VRAM utilization:'
'%4.2fG' % (torch.cuda.memory_allocated() / 1e9),
)
self.session_peakmem = max(
self.session_peakmem, torch.cuda.max_memory_allocated()
)
print(
f'>> Max VRAM used since script start: ',
'%4.2fG' % (self.session_peakmem / 1e9),
)
return results
def _make_images(self, img_path, mask_path, width, height, fit=False):
init_image = None
init_mask = None
if not img_path:
return None,None
image = self._load_img(img_path, width, height, fit=fit) # this returns an Image
init_image = self._create_init_image(image) # this returns a torch tensor
if self._has_transparency(image) and not mask_path: # if image has a transparent area and no mask was provided, then try to generate mask
print('>> Initial image has transparent areas. Will inpaint in these regions.')
if self._check_for_erasure(image):
print(
'>> WARNING: Colors underneath the transparent region seem to have been erased.\n',
'>> Inpainting will be suboptimal. Please preserve the colors when making\n',
'>> a transparency mask, or provide mask explicitly using --init_mask (-M).'
)
init_mask = self._create_init_mask(image) # this returns a torch tensor
if mask_path:
mask_image = self._load_img(mask_path, width, height, fit=fit) # this returns an Image
init_mask = self._create_init_mask(mask_image)
return init_image,init_mask
def _make_img2img(self):
if not self.generators.get('img2img'):
from ldm.dream.generator.img2img import Img2Img
self.generators['img2img'] = Img2Img(self.model)
return self.generators['img2img']
def _make_txt2img(self):
if not self.generators.get('txt2img'):
from ldm.dream.generator.txt2img import Txt2Img
self.generators['txt2img'] = Txt2Img(self.model)
return self.generators['txt2img']
def _make_inpaint(self):
if not self.generators.get('inpaint'):
from ldm.dream.generator.inpaint import Inpaint
self.generators['inpaint'] = Inpaint(self.model)
return self.generators['inpaint']
def load_model(self):
"""Load and initialize the model from configuration variables passed at object creation time"""
if self.model is None:
seed_everything(random.randrange(0, np.iinfo(np.uint32).max))
try:
config = OmegaConf.load(self.config)
model = self._load_model_from_config(config, self.weights)
if self.embedding_path is not None:
model.embedding_manager.load(
self.embedding_path, self.full_precision
)
self.model = model.to(self.device)
# model.to doesn't change the cond_stage_model.device used to move the tokenizer output, so set it here
self.model.cond_stage_model.device = self.device
except AttributeError as e:
print(f'>> Error loading model. {str(e)}', file=sys.stderr)
print(traceback.format_exc(), file=sys.stderr)
raise SystemExit from e
self._set_sampler()
for m in self.model.modules():
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
m._orig_padding_mode = m.padding_mode
return self.model
def upscale_and_reconstruct(self,
image_list,
upscale = None,
strength = 0.0,
save_original = False,
image_callback = None):
try:
if upscale is not None:
from ldm.gfpgan.gfpgan_tools import real_esrgan_upscale
if strength > 0:
from ldm.gfpgan.gfpgan_tools import run_gfpgan
except (ModuleNotFoundError, ImportError):
print(traceback.format_exc(), file=sys.stderr)
print('>> You may need to install the ESRGAN and/or GFPGAN modules')
return
for r in image_list:
image, seed = r
try:
if upscale is not None:
if len(upscale) < 2:
upscale.append(0.75)
image = real_esrgan_upscale(
image,
upscale[1],
int(upscale[0]),
seed,
)
if strength > 0:
image = run_gfpgan(
image, strength, seed, 1
)
except Exception as e:
print(
f'>> Error running RealESRGAN or GFPGAN. Your image was not upscaled.\n{e}'
)
if image_callback is not None:
image_callback(image, seed, upscaled=True)
else:
r[0] = image
# to help WebGUI - front end to generator util function
def sample_to_image(self,samples):
return self._sample_to_image(samples)
def _sample_to_image(self,samples):
if not self.base_generator:
from ldm.dream.generator import Generator
self.base_generator = Generator(self.model)
return self.base_generator.sample_to_image(samples)
def _set_sampler(self):
msg = f'>> Setting Sampler to {self.sampler_name}'
if self.sampler_name == 'plms':
self.sampler = PLMSSampler(self.model, device=self.device)
elif self.sampler_name == 'ddim':
self.sampler = DDIMSampler(self.model, device=self.device)
elif self.sampler_name == 'k_dpm_2_a':
self.sampler = KSampler(
self.model, 'dpm_2_ancestral', device=self.device
)
elif self.sampler_name == 'k_dpm_2':
self.sampler = KSampler(self.model, 'dpm_2', device=self.device)
elif self.sampler_name == 'k_euler_a':
self.sampler = KSampler(
self.model, 'euler_ancestral', device=self.device
)
elif self.sampler_name == 'k_euler':
self.sampler = KSampler(self.model, 'euler', device=self.device)
elif self.sampler_name == 'k_heun':
self.sampler = KSampler(self.model, 'heun', device=self.device)
elif self.sampler_name == 'k_lms':
self.sampler = KSampler(self.model, 'lms', device=self.device)
else:
msg = f'>> Unsupported Sampler: {self.sampler_name}, Defaulting to plms'
self.sampler = PLMSSampler(self.model, device=self.device)
print(msg)
def _load_model_from_config(self, config, ckpt):
print(f'>> Loading model from {ckpt}')
# for usage statistics
device_type = choose_torch_device()
if device_type == 'cuda':
torch.cuda.reset_peak_memory_stats()
tic = time.time()
# this does the work
pl_sd = torch.load(ckpt, map_location='cpu')
sd = pl_sd['state_dict']
model = instantiate_from_config(config.model)
m, u = model.load_state_dict(sd, strict=False)
if self.full_precision:
print(
'>> Using slower but more accurate full-precision math (--full_precision)'
)
else:
print(
'>> Using half precision math. Call with --full_precision to use more accurate but VRAM-intensive full precision.'
)
model.half()
model.to(self.device)
model.eval()
# usage statistics
toc = time.time()
print(
f'>> Model loaded in', '%4.2fs' % (toc - tic)
)
if device_type == 'cuda':
print(
'>> Max VRAM used to load the model:',
'%4.2fG' % (torch.cuda.max_memory_allocated() / 1e9),
'\n>> Current VRAM usage:'
'%4.2fG' % (torch.cuda.memory_allocated() / 1e9),
)
return model
def _load_img(self, path, width, height, fit=False):
assert os.path.exists(path), f'>> {path}: File not found'
# with Image.open(path) as img:
# image = img.convert('RGBA')
image = Image.open(path)
print(
f'>> loaded input image of size {image.width}x{image.height} from {path}'
)
if fit:
image = self._fit_image(image,(width,height))
else:
image = self._squeeze_image(image)
return image
def _create_init_image(self,image):
image = image.convert('RGB')
# print(
# f'>> DEBUG: writing the image to img.png'
# )
# image.save('img.png')
image = np.array(image).astype(np.float32) / 255.0
image = image[None].transpose(0, 3, 1, 2)
image = torch.from_numpy(image)
image = 2.0 * image - 1.0
return image.to(self.device)
def _create_init_mask(self, image):
# convert into a black/white mask
image = self._image_to_mask(image)
image = image.convert('RGB')
# BUG: We need to use the model's downsample factor rather than hardcoding "8"
from ldm.dream.generator.base import downsampling
image = image.resize((image.width//downsampling, image.height//downsampling), resample=Image.Resampling.LANCZOS)
# print(
# f'>> DEBUG: writing the mask to mask.png'
# )
# image.save('mask.png')
image = np.array(image)
image = image.astype(np.float32) / 255.0
image = image[None].transpose(0, 3, 1, 2)
image = torch.from_numpy(image)
return image.to(self.device)
# The mask is expected to have the region to be inpainted
# with alpha transparency. It converts it into a black/white
# image with the transparent part black.
def _image_to_mask(self, mask_image, invert=False) -> Image:
# Obtain the mask from the transparency channel
mask = Image.new(mode="L", size=mask_image.size, color=255)
mask.putdata(mask_image.getdata(band=3))
if invert:
mask = ImageOps.invert(mask)
return mask
def _has_transparency(self,image):
if image.info.get("transparency", None) is not None:
return True
if image.mode == "P":
transparent = image.info.get("transparency", -1)
for _, index in image.getcolors():
if index == transparent:
return True
elif image.mode == "RGBA":
extrema = image.getextrema()
if extrema[3][0] < 255:
return True
return False
def _check_for_erasure(self,image):
width, height = image.size
pixdata = image.load()
colored = 0
for y in range(height):
for x in range(width):
if pixdata[x, y][3] == 0:
r, g, b, _ = pixdata[x, y]
if (r, g, b) != (0, 0, 0) and \
(r, g, b) != (255, 255, 255):
colored += 1
return colored == 0
def _squeeze_image(self,image):
x,y,resize_needed = self._resolution_check(image.width,image.height)
if resize_needed:
return InitImageResizer(image).resize(x,y)
return image
def _fit_image(self,image,max_dimensions):
w,h = max_dimensions
print(
f'>> image will be resized to fit inside a box {w}x{h} in size.'
)
if image.width > image.height:
h = None # by setting h to none, we tell InitImageResizer to fit into the width and calculate height
elif image.height > image.width:
w = None # ditto for w
else:
pass
image = InitImageResizer(image).resize(w,h) # note that InitImageResizer does the multiple of 64 truncation internally
print(
f'>> after adjusting image dimensions to be multiples of 64, init image is {image.width}x{image.height}'
)
return image
def _resolution_check(self, width, height, log=False):
resize_needed = False
w, h = map(
lambda x: x - x % 64, (width, height)
) # resize to integer multiple of 64
if h != height or w != width:
if log:
print(
f'>> Provided width and height must be multiples of 64. Auto-resizing to {w}x{h}'
)
height = h
width = w
resize_needed = True
if (width * height) > (self.width * self.height):
print(">> This input is larger than your defaults. If you run out of memory, please use a smaller image.")
return width, height, resize_needed

100
ldm/gfpgan/gfpgan_tools.py
View File

@@ -8,18 +8,17 @@ from PIL import Image
from scripts.dream import create_argv_parser
arg_parser = create_argv_parser()
opt = arg_parser.parse_args()
model_path = os.path.join(opt.gfpgan_dir, opt.gfpgan_model_path)
opt = arg_parser.parse_args()
model_path = os.path.join(opt.gfpgan_dir, opt.gfpgan_model_path)
gfpgan_model_exists = os.path.isfile(model_path)
def _run_gfpgan(image, strength, prompt, seed, upsampler_scale=4):
print(f'>> GFPGAN - Restoring Faces: {prompt} : seed:{seed}')
def run_gfpgan(image, strength, seed, upsampler_scale=4):
print(f'>> GFPGAN - Restoring Faces for image seed:{seed}')
gfpgan = None
with warnings.catch_warnings():
warnings.filterwarnings('ignore', category=DeprecationWarning)
warnings.filterwarnings('ignore', category=UserWarning)
try:
if not gfpgan_model_exists:
raise Exception('GFPGAN model not found at path ' + model_path)
@@ -46,7 +45,10 @@ def _run_gfpgan(image, strength, prompt, seed, upsampler_scale=4):
if gfpgan is None:
print(
f'>> GFPGAN not initialized, it must be loaded via the --gfpgan argument'
f'>> WARNING: GFPGAN not initialized.'
)
print(
f'>> Download https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth to {model_path}, \nor change GFPGAN directory with --gfpgan_dir.'
)
return image
@@ -75,61 +77,59 @@ def _run_gfpgan(image, strength, prompt, seed, upsampler_scale=4):
def _load_gfpgan_bg_upsampler(bg_upsampler, upsampler_scale, bg_tile=400):
if bg_upsampler == 'realesrgan':
if not torch.cuda.is_available(): # CPU
warnings.warn(
'The unoptimized RealESRGAN is slow on CPU. We do not use it. '
'If you really want to use it, please modify the corresponding codes.'
)
bg_upsampler = None
if not torch.cuda.is_available(): # CPU or MPS on M1
use_half_precision = False
else:
model_path = {
2: 'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth',
4: 'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth',
}
use_half_precision = True
if upsampler_scale not in model_path:
return None
model_path = {
2: 'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth',
4: 'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth',
}
from basicsr.archs.rrdbnet_arch import RRDBNet
from realesrgan import RealESRGANer
if upsampler_scale not in model_path:
return None
if upsampler_scale == 4:
model = RRDBNet(
num_in_ch=3,
num_out_ch=3,
num_feat=64,
num_block=23,
num_grow_ch=32,
scale=4,
)
if upsampler_scale == 2:
model = RRDBNet(
num_in_ch=3,
num_out_ch=3,
num_feat=64,
num_block=23,
num_grow_ch=32,
scale=2,
)
from basicsr.archs.rrdbnet_arch import RRDBNet
from realesrgan import RealESRGANer
bg_upsampler = RealESRGANer(
scale=upsampler_scale,
model_path=model_path[upsampler_scale],
model=model,
tile=bg_tile,
tile_pad=10,
pre_pad=0,
half=True,
) # need to set False in CPU mode
if upsampler_scale == 4:
model = RRDBNet(
num_in_ch=3,
num_out_ch=3,
num_feat=64,
num_block=23,
num_grow_ch=32,
scale=4,
)
if upsampler_scale == 2:
model = RRDBNet(
num_in_ch=3,
num_out_ch=3,
num_feat=64,
num_block=23,
num_grow_ch=32,
scale=2,
)
bg_upsampler = RealESRGANer(
scale=upsampler_scale,
model_path=model_path[upsampler_scale],
model=model,
tile=bg_tile,
tile_pad=10,
pre_pad=0,
half=use_half_precision,
)
else:
bg_upsampler = None
return bg_upsampler
def real_esrgan_upscale(image, strength, upsampler_scale, prompt, seed):
def real_esrgan_upscale(image, strength, upsampler_scale, seed):
print(
f'>> Real-ESRGAN Upscaling: {prompt} : seed:{seed} : scale:{upsampler_scale}x'
f'>> Real-ESRGAN Upscaling seed:{seed} : scale:{upsampler_scale}x'
)
with warnings.catch_warnings():

31
ldm/models/diffusion/ddim.py
View File

@@ -171,6 +171,7 @@ class DDIMSampler(object):
)
return samples, intermediates
# This routine gets called from img2img
@torch.no_grad()
def ddim_sampling(
self,
@@ -270,6 +271,7 @@ class DDIMSampler(object):
return img, intermediates
# This routine gets called from ddim_sampling() and decode()
@torch.no_grad()
def p_sample_ddim(
self,
@@ -372,14 +374,16 @@ class DDIMSampler(object):
@torch.no_grad()
def decode(
self,
x_latent,
cond,
t_start,
img_callback=None,
unconditional_guidance_scale=1.0,
unconditional_conditioning=None,
use_original_steps=False,
self,
x_latent,
cond,
t_start,
img_callback=None,
unconditional_guidance_scale=1.0,
unconditional_conditioning=None,
use_original_steps=False,
init_latent = None,
mask = None,
):
timesteps = (
@@ -395,6 +399,8 @@ class DDIMSampler(object):
iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
x_dec = x_latent
x0 = init_latent
for i, step in enumerate(iterator):
index = total_steps - i - 1
ts = torch.full(
@@ -403,6 +409,14 @@ class DDIMSampler(object):
device=x_latent.device,
dtype=torch.long,
)
if mask is not None:
assert x0 is not None
xdec_orig = self.model.q_sample(
x0, ts
) # TODO: deterministic forward pass?
x_dec = xdec_orig * mask + (1.0 - mask) * x_dec
x_dec, _ = self.p_sample_ddim(
x_dec,
cond,
@@ -412,6 +426,7 @@ class DDIMSampler(object):
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning,
)
if img_callback:
img_callback(x_dec, i)

280
ldm/modules/attention.py
View File

@@ -7,13 +7,14 @@ from einops import rearrange, repeat
from ldm.modules.diffusionmodules.util import checkpoint
import psutil
def exists(val):
return val is not None
def uniq(arr):
return {el: True for el in arr}.keys()
return{el: True for el in arr}.keys()
def default(val, d):
@@ -45,18 +46,19 @@ class GEGLU(nn.Module):
class FeedForward(nn.Module):
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
super().__init__()
inner_dim = int(dim * mult)
dim_out = default(dim_out, dim)
project_in = (
nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
if not glu
else GEGLU(dim, inner_dim)
)
project_in = nn.Sequential(
nn.Linear(dim, inner_dim),
nn.GELU()
) if not glu else GEGLU(dim, inner_dim)
self.net = nn.Sequential(
project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
project_in,
nn.Dropout(dropout),
nn.Linear(inner_dim, dim_out)
)
def forward(self, x):
@@ -73,9 +75,7 @@ def zero_module(module):
def Normalize(in_channels):
return torch.nn.GroupNorm(
num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
)
return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
class LinearAttention(nn.Module):
@@ -83,28 +83,17 @@ class LinearAttention(nn.Module):
super().__init__()
self.heads = heads
hidden_dim = dim_head * heads
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
self.to_out = nn.Conv2d(hidden_dim, dim, 1)
def forward(self, x):
b, c, h, w = x.shape
qkv = self.to_qkv(x)
q, k, v = rearrange(
qkv,
'b (qkv heads c) h w -> qkv b heads c (h w)',
heads=self.heads,
qkv=3,
)
k = k.softmax(dim=-1)
q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
k = k.softmax(dim=-1)
context = torch.einsum('bhdn,bhen->bhde', k, v)
out = torch.einsum('bhde,bhdn->bhen', context, q)
out = rearrange(
out,
'b heads c (h w) -> b (heads c) h w',
heads=self.heads,
h=h,
w=w,
)
out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
return self.to_out(out)
@@ -114,18 +103,26 @@ class SpatialSelfAttention(nn.Module):
self.in_channels = in_channels
self.norm = Normalize(in_channels)
self.q = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.k = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.v = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.proj_out = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.q = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.k = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.v = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.proj_out = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
def forward(self, x):
h_ = x
@@ -135,12 +132,12 @@ class SpatialSelfAttention(nn.Module):
v = self.v(h_)
# compute attention
b, c, h, w = q.shape
b,c,h,w = q.shape
q = rearrange(q, 'b c h w -> b (h w) c')
k = rearrange(k, 'b c h w -> b c (h w)')
w_ = torch.einsum('bij,bjk->bik', q, k)
w_ = w_ * (int(c) ** (-0.5))
w_ = w_ * (int(c)**(-0.5))
w_ = torch.nn.functional.softmax(w_, dim=2)
# attend to values
@@ -150,18 +147,16 @@ class SpatialSelfAttention(nn.Module):
h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
h_ = self.proj_out(h_)
return x + h_
return x+h_
class CrossAttention(nn.Module):
def __init__(
self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0
):
def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
super().__init__()
inner_dim = dim_head * heads
context_dim = default(context_dim, query_dim)
self.scale = dim_head**-0.5
self.scale = dim_head ** -0.5
self.heads = heads
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
@@ -169,69 +164,136 @@ class CrossAttention(nn.Module):
self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
self.to_out = nn.Sequential(
nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
nn.Linear(inner_dim, query_dim),
nn.Dropout(dropout)
)
if not torch.cuda.is_available():
mem_av = psutil.virtual_memory().available / (1024**3)
if mem_av > 32:
self.einsum_op = self.einsum_op_v1
elif mem_av > 12:
self.einsum_op = self.einsum_op_v2
else:
self.einsum_op = self.einsum_op_v3
del mem_av
else:
self.einsum_op = self.einsum_op_v4
# mps 64-128 GB
def einsum_op_v1(self, q, k, v, r1):
if q.shape[1] <= 4096: # for 512x512: the max q.shape[1] is 4096
s1 = einsum('b i d, b j d -> b i j', q, k) * self.scale # aggressive/faster: operation in one go
s2 = s1.softmax(dim=-1, dtype=q.dtype)
del s1
r1 = einsum('b i j, b j d -> b i d', s2, v)
del s2
else:
# q.shape[0] * q.shape[1] * slice_size >= 2**31 throws err
# needs around half of that slice_size to not generate noise
slice_size = math.floor(2**30 / (q.shape[0] * q.shape[1]))
for i in range(0, q.shape[1], slice_size):
end = i + slice_size
s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
del s2
return r1
# mps 16-32 GB (can be optimized)
def einsum_op_v2(self, q, k, v, r1):
slice_size = math.floor(2**30 / (q.shape[0] * q.shape[1]))
for i in range(0, q.shape[1], slice_size): # conservative/less mem: operation in steps
end = i + slice_size
s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
del s2
return r1
# mps 8 GB
def einsum_op_v3(self, q, k, v, r1):
slice_size = 1
for i in range(0, q.shape[0], slice_size): # iterate over q.shape[0]
end = min(q.shape[0], i + slice_size)
s1 = einsum('b i d, b j d -> b i j', q[i:end], k[i:end]) # adapted einsum for mem
s1 *= self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[i:end] = einsum('b i j, b j d -> b i d', s2, v[i:end]) # adapted einsum for mem
del s2
return r1
# cuda
def einsum_op_v4(self, q, k, v, r1):
stats = torch.cuda.memory_stats(q.device)
mem_active = stats['active_bytes.all.current']
mem_reserved = stats['reserved_bytes.all.current']
mem_free_cuda, _ = torch.cuda.mem_get_info(torch.cuda.current_device())
mem_free_torch = mem_reserved - mem_active
mem_free_total = mem_free_cuda + mem_free_torch
gb = 1024 ** 3
tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * 4
mem_required = tensor_size * 2.5
steps = 1
if mem_required > mem_free_total:
steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
if steps > 64:
max_res = math.floor(math.sqrt(math.sqrt(mem_free_total / 2.5)) / 8) * 64
raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). '
f'Need: {mem_required/64/gb:0.1f}GB free, Have:{mem_free_total/gb:0.1f}GB free')
slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
for i in range(0, q.shape[1], slice_size):
end = min(q.shape[1], i + slice_size)
s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
del s2
return r1
def forward(self, x, context=None, mask=None):
h = self.heads
q = self.to_q(x)
q_in = self.to_q(x)
context = default(context, x)
k = self.to_k(context)
v = self.to_v(context)
k_in = self.to_k(context)
v_in = self.to_v(context)
device_type = 'mps' if x.device.type == 'mps' else 'cuda'
del context, x
q, k, v = map(
lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)
)
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in))
del q_in, k_in, v_in
r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
r1 = self.einsum_op(q, k, v, r1)
del q, k, v
sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
del r1
if exists(mask):
mask = rearrange(mask, 'b ... -> b (...)')
max_neg_value = -torch.finfo(sim.dtype).max
mask = repeat(mask, 'b j -> (b h) () j', h=h)
sim.masked_fill_(~mask, max_neg_value)
# attention, what we cannot get enough of
attn = sim.softmax(dim=-1)
out = einsum('b i j, b j d -> b i d', attn, v)
out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
return self.to_out(out)
return self.to_out(r2)
class BasicTransformerBlock(nn.Module):
def __init__(
self,
dim,
n_heads,
d_head,
dropout=0.0,
context_dim=None,
gated_ff=True,
checkpoint=True,
):
def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True):
super().__init__()
self.attn1 = CrossAttention(
query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
) # is a self-attention
self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout) # is a self-attention
self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
self.attn2 = CrossAttention(
query_dim=dim,
context_dim=context_dim,
heads=n_heads,
dim_head=d_head,
dropout=dropout,
) # is self-attn if context is none
self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
heads=n_heads, dim_head=d_head, dropout=dropout) # is self-attn if context is none
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.norm3 = nn.LayerNorm(dim)
self.checkpoint = checkpoint
def forward(self, x, context=None):
return checkpoint(
self._forward, (x, context), self.parameters(), self.checkpoint
)
return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
def _forward(self, x, context=None):
x = x.contiguous() if x.device.type == 'mps' else x
@@ -249,43 +311,29 @@ class SpatialTransformer(nn.Module):
Then apply standard transformer action.
Finally, reshape to image
"""
def __init__(
self,
in_channels,
n_heads,
d_head,
depth=1,
dropout=0.0,
context_dim=None,
):
def __init__(self, in_channels, n_heads, d_head,
depth=1, dropout=0., context_dim=None):
super().__init__()
self.in_channels = in_channels
inner_dim = n_heads * d_head
self.norm = Normalize(in_channels)
self.proj_in = nn.Conv2d(
in_channels, inner_dim, kernel_size=1, stride=1, padding=0
)
self.proj_in = nn.Conv2d(in_channels,
inner_dim,
kernel_size=1,
stride=1,
padding=0)
self.transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock(
inner_dim,
n_heads,
d_head,
dropout=dropout,
context_dim=context_dim,
)
for d in range(depth)
]
[BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
for d in range(depth)]
)
self.proj_out = zero_module(
nn.Conv2d(
inner_dim, in_channels, kernel_size=1, stride=1, padding=0
)
)
self.proj_out = zero_module(nn.Conv2d(inner_dim,
in_channels,
kernel_size=1,
stride=1,
padding=0))
def forward(self, x, context=None):
# note: if no context is given, cross-attention defaults to self-attention

939
ldm/modules/diffusionmodules/model.py
View File

File diff suppressed because it is too large Load Diff

4
ldm/modules/diffusionmodules/util.py
View File

@@ -81,7 +81,9 @@ def make_ddim_timesteps(
# assert ddim_timesteps.shape[0] == num_ddim_timesteps
# add one to get the final alpha values right (the ones from first scale to data during sampling)
steps_out = ddim_timesteps + 1
# steps_out = ddim_timesteps + 1
steps_out = ddim_timesteps
if verbose:
print(f'Selected timesteps for ddim sampler: {steps_out}')
return steps_out

27
ldm/modules/embedding_manager.py
View File

@@ -24,9 +24,9 @@ def get_clip_token_for_string(tokenizer, string):
return_tensors='pt',
)
tokens = batch_encoding['input_ids']
assert (
""" assert (
torch.count_nonzero(tokens - 49407) == 2
), f"String '{string}' maps to more than a single token. Please use another string"
), f"String '{string}' maps to more than a single token. Please use another string" """
return tokens[0, 1]
@@ -57,8 +57,9 @@ class EmbeddingManager(nn.Module):
):
super().__init__()
self.string_to_token_dict = {}
self.embedder = embedder
self.string_to_token_dict = {}
self.string_to_param_dict = nn.ParameterDict()
self.initial_embeddings = (
@@ -217,12 +218,28 @@ class EmbeddingManager(nn.Module):
def load(self, ckpt_path, full=True):
ckpt = torch.load(ckpt_path, map_location='cpu')
self.string_to_token_dict = ckpt["string_to_token"]
self.string_to_param_dict = ckpt["string_to_param"]
# Handle .pt textual inversion files
if 'string_to_token' in ckpt and 'string_to_param' in ckpt:
self.string_to_token_dict = ckpt["string_to_token"]
self.string_to_param_dict = ckpt["string_to_param"]
# Handle .bin textual inversion files from Huggingface Concepts
# https://huggingface.co/sd-concepts-library
else:
for token_str in list(ckpt.keys()):
token = get_clip_token_for_string(self.embedder.tokenizer, token_str)
self.string_to_token_dict[token_str] = token
ckpt[token_str] = torch.nn.Parameter(ckpt[token_str])
self.string_to_param_dict.update(ckpt)
if not full:
for key, value in self.string_to_param_dict.items():
self.string_to_param_dict[key] = torch.nn.Parameter(value.half())
print(f'Added terms: {", ".join(self.string_to_param_dict.keys())}')
def get_embedding_norms_squared(self):
all_params = torch.cat(
list(self.string_to_param_dict.values()), axis=0

851
ldm/simplet2i.py
View File

@@ -1,844 +1,13 @@
# Copyright (c) 2022 Lincoln D. Stein (https://github.com/lstein)
'''
This module is provided for backward compatibility with the
original (hasty) API.
# Derived from source code carrying the following copyrights
# Copyright (c) 2022 Machine Vision and Learning Group, LMU Munich
# Copyright (c) 2022 Robin Rombach and Patrick Esser and contributors
Please use ldm.generate instead.
'''
import torch
import numpy as np
import random
import os
import traceback
from omegaconf import OmegaConf
from PIL import Image
from tqdm import tqdm, trange
from itertools import islice
from einops import rearrange, repeat
from torchvision.utils import make_grid
from pytorch_lightning import seed_everything
from torch import autocast
from contextlib import contextmanager, nullcontext
import transformers
import time
import re
import sys
from ldm.generate import Generate
from ldm.util import instantiate_from_config
from ldm.models.diffusion.ddim import DDIMSampler
from ldm.models.diffusion.plms import PLMSSampler
from ldm.models.diffusion.ksampler import KSampler
from ldm.dream.pngwriter import PngWriter
from ldm.dream.image_util import InitImageResizer
from ldm.dream.devices import choose_autocast_device, choose_torch_device
"""Simplified text to image API for stable diffusion/latent diffusion
Example Usage:
from ldm.simplet2i import T2I
# Create an object with default values
t2i = T2I(model = <path> // models/ldm/stable-diffusion-v1/model.ckpt
config = <path> // configs/stable-diffusion/v1-inference.yaml
iterations = <integer> // how many times to run the sampling (1)
steps = <integer> // 50
seed = <integer> // current system time
sampler_name= ['ddim', 'k_dpm_2_a', 'k_dpm_2', 'k_euler_a', 'k_euler', 'k_heun', 'k_lms', 'plms'] // k_lms
grid = <boolean> // false
width = <integer> // image width, multiple of 64 (512)
height = <integer> // image height, multiple of 64 (512)
cfg_scale = <float> // unconditional guidance scale (7.5)
)
# do the slow model initialization
t2i.load_model()
# Do the fast inference & image generation. Any options passed here
# override the default values assigned during class initialization
# Will call load_model() if the model was not previously loaded and so
# may be slow at first.
# The method returns a list of images. Each row of the list is a sub-list of [filename,seed]
results = t2i.prompt2png(prompt = "an astronaut riding a horse",
outdir = "./outputs/samples",
iterations = 3)
for row in results:
print(f'filename={row[0]}')
print(f'seed ={row[1]}')
# Same thing, but using an initial image.
results = t2i.prompt2png(prompt = "an astronaut riding a horse",
outdir = "./outputs/,
iterations = 3,
init_img = "./sketches/horse+rider.png")
for row in results:
print(f'filename={row[0]}')
print(f'seed ={row[1]}')
# Same thing, but we return a series of Image objects, which lets you manipulate them,
# combine them, and save them under arbitrary names
results = t2i.prompt2image(prompt = "an astronaut riding a horse"
outdir = "./outputs/")
for row in results:
im = row[0]
seed = row[1]
im.save(f'./outputs/samples/an_astronaut_riding_a_horse-{seed}.png')
im.thumbnail(100,100).save('./outputs/samples/astronaut_thumb.jpg')
Note that the old txt2img() and img2img() calls are deprecated but will
still work.
"""
class T2I:
"""T2I class
Attributes
----------
model
config
iterations
steps
seed
sampler_name
width
height
cfg_scale
latent_channels
downsampling_factor
precision
strength
embedding_path
The vast majority of these arguments default to reasonable values.
"""
def __init__(
self,
iterations=1,
steps=50,
seed=None,
cfg_scale=7.5,
weights='models/ldm/stable-diffusion-v1/model.ckpt',
config='configs/stable-diffusion/v1-inference.yaml',
grid=False,
width=512,
height=512,
sampler_name='k_lms',
latent_channels=4,
downsampling_factor=8,
ddim_eta=0.0, # deterministic
precision='autocast',
full_precision=False,
strength=0.75, # default in scripts/img2img.py
embedding_path=None,
device_type = 'cuda',
# just to keep track of this parameter when regenerating prompt
# needs to be replaced when new configuration system implemented.
latent_diffusion_weights=False,
):
self.iterations = iterations
self.width = width
self.height = height
self.steps = steps
self.cfg_scale = cfg_scale
self.weights = weights
self.config = config
self.sampler_name = sampler_name
self.latent_channels = latent_channels
self.downsampling_factor = downsampling_factor
self.grid = grid
self.ddim_eta = ddim_eta
self.precision = precision
self.full_precision = True if choose_torch_device() == 'mps' else full_precision
self.strength = strength
self.embedding_path = embedding_path
self.device_type = device_type
self.model = None # empty for now
self.sampler = None
self.device = None
self.latent_diffusion_weights = latent_diffusion_weights
if device_type == 'cuda' and not torch.cuda.is_available():
device_type = choose_torch_device()
print(">> cuda not available, using device", device_type)
self.device = torch.device(device_type)
# for VRAM usage statistics
device_type = choose_torch_device()
self.session_peakmem = torch.cuda.max_memory_allocated() if device_type == 'cuda' else None
if seed is None:
self.seed = self._new_seed()
else:
self.seed = seed
transformers.logging.set_verbosity_error()
def prompt2png(self, prompt, outdir, **kwargs):
"""
Takes a prompt and an output directory, writes out the requested number
of PNG files, and returns an array of [[filename,seed],[filename,seed]...]
Optional named arguments are the same as those passed to T2I and prompt2image()
"""
results = self.prompt2image(prompt, **kwargs)
pngwriter = PngWriter(outdir)
prefix = pngwriter.unique_prefix()
outputs = []
for image, seed in results:
name = f'{prefix}.{seed}.png'
path = pngwriter.save_image_and_prompt_to_png(
image, f'{prompt} -S{seed}', name)
outputs.append([path, seed])
return outputs
def txt2img(self, prompt, **kwargs):
outdir = kwargs.pop('outdir', 'outputs/img-samples')
return self.prompt2png(prompt, outdir, **kwargs)
def img2img(self, prompt, **kwargs):
outdir = kwargs.pop('outdir', 'outputs/img-samples')
assert (
'init_img' in kwargs
), 'call to img2img() must include the init_img argument'
return self.prompt2png(prompt, outdir, **kwargs)
def prompt2image(
self,
# these are common
prompt,
iterations = None,
steps = None,
seed = None,
cfg_scale = None,
ddim_eta = None,
skip_normalize = False,
image_callback = None,
step_callback = None,
width = None,
height = None,
# these are specific to img2img
init_img = None,
fit = False,
strength = None,
gfpgan_strength= 0,
save_original = False,
upscale = None,
sampler_name = None,
log_tokenization= False,
with_variations = None,
variation_amount = 0.0,
**args,
): # eat up additional cruft
"""
ldm.prompt2image() is the common entry point for txt2img() and img2img()
It takes the following arguments:
prompt // prompt string (no default)
iterations // iterations (1); image count=iterations
steps // refinement steps per iteration
seed // seed for random number generator
width // width of image, in multiples of 64 (512)
height // height of image, in multiples of 64 (512)
cfg_scale // how strongly the prompt influences the image (7.5) (must be >1)
init_img // path to an initial image - its dimensions override width and height
strength // strength for noising/unnoising init_img. 0.0 preserves image exactly, 1.0 replaces it completely
gfpgan_strength // strength for GFPGAN. 0.0 preserves image exactly, 1.0 replaces it completely
ddim_eta // image randomness (eta=0.0 means the same seed always produces the same image)
step_callback // a function or method that will be called each step
image_callback // a function or method that will be called each time an image is generated
with_variations // a weighted list [(seed_1, weight_1), (seed_2, weight_2), ...] of variations which should be applied before doing any generation
variation_amount // optional 0-1 value to slerp from -S noise to random noise (allows variations on an image)
To use the step callback, define a function that receives two arguments:
- Image GPU data
- The step number
To use the image callback, define a function of method that receives two arguments, an Image object
and the seed. You can then do whatever you like with the image, including converting it to
different formats and manipulating it. For example:
def process_image(image,seed):
image.save(f{'images/seed.png'})
The callback used by the prompt2png() can be found in ldm/dream_util.py. It contains code
to create the requested output directory, select a unique informative name for each image, and
write the prompt into the PNG metadata.
"""
# TODO: convert this into a getattr() loop
steps = steps or self.steps
width = width or self.width
height = height or self.height
cfg_scale = cfg_scale or self.cfg_scale
ddim_eta = ddim_eta or self.ddim_eta
iterations = iterations or self.iterations
strength = strength or self.strength
self.log_tokenization = log_tokenization
with_variations = [] if with_variations is None else with_variations
model = (
self.load_model()
) # will instantiate the model or return it from cache
assert cfg_scale > 1.0, 'CFG_Scale (-C) must be >1.0'
assert (
0.0 <= strength <= 1.0
), 'can only work with strength in [0.0, 1.0]'
assert (
0.0 <= variation_amount <= 1.0
), '-v --variation_amount must be in [0.0, 1.0]'
if len(with_variations) > 0 or variation_amount > 0.0:
assert seed is not None,\
'seed must be specified when using with_variations'
if variation_amount == 0.0:
assert iterations == 1,\
'when using --with_variations, multiple iterations are only possible when using --variation_amount'
assert all(0 <= weight <= 1 for _, weight in with_variations),\
f'variation weights must be in [0.0, 1.0]: got {[weight for _, weight in with_variations]}'
seed = seed or self.seed
width, height, _ = self._resolution_check(width, height, log=True)
# TODO: - Check if this is still necessary to run on M1 devices.
# - Move code into ldm.dream.devices to live alongside other
# special-hardware casing code.
if self.precision == 'autocast' and torch.cuda.is_available():
scope = autocast
else:
scope = nullcontext
if sampler_name and (sampler_name != self.sampler_name):
self.sampler_name = sampler_name
self._set_sampler()
tic = time.time()
if torch.cuda.is_available():
torch.cuda.reset_peak_memory_stats()
results = list()
try:
if init_img:
assert os.path.exists(init_img), f'{init_img}: File not found'
init_image = self._load_img(init_img, width, height, fit).to(self.device)
with scope(self.device.type):
init_latent = self.model.get_first_stage_encoding(
self.model.encode_first_stage(init_image)
) # move to latent space
#print(f' DEBUG: seed at make_image time ={seed}')
make_image = self._img2img(
prompt,
steps=steps,
cfg_scale=cfg_scale,
ddim_eta=ddim_eta,
skip_normalize=skip_normalize,
init_latent=init_latent,
strength=strength,
callback=step_callback,
)
else:
init_latent = None
make_image = self._txt2img(
prompt,
steps=steps,
cfg_scale=cfg_scale,
ddim_eta=ddim_eta,
skip_normalize=skip_normalize,
width=width,
height=height,
callback=step_callback,
)
initial_noise = None
if variation_amount > 0 or len(with_variations) > 0:
# use fixed initial noise plus random noise per iteration
seed_everything(seed)
initial_noise = self._get_noise(init_latent,width,height)
for v_seed, v_weight in with_variations:
seed = v_seed
seed_everything(seed)
next_noise = self._get_noise(init_latent,width,height)
initial_noise = self.slerp(v_weight, initial_noise, next_noise)
if variation_amount > 0:
random.seed() # reset RNG to an actually random state, so we can get a random seed for variations
seed = random.randrange(0,np.iinfo(np.uint32).max)
device_type = choose_autocast_device(self.device)
with scope(device_type), self.model.ema_scope():
for n in trange(iterations, desc='Generating'):
x_T = None
if variation_amount > 0:
seed_everything(seed)
target_noise = self._get_noise(init_latent,width,height)
x_T = self.slerp(variation_amount, initial_noise, target_noise)
elif initial_noise is not None:
# i.e. we specified particular variations
x_T = initial_noise
else:
seed_everything(seed)
if self.device.type == 'mps':
x_T = self._get_noise(init_latent,width,height)
# make_image will do the equivalent of get_noise itself
#print(f' DEBUG: seed at make_image() invocation time ={seed}')
image = make_image(x_T)
results.append([image, seed])
if image_callback is not None:
image_callback(image, seed)
seed = self._new_seed()
if upscale is not None or gfpgan_strength > 0:
for result in results:
image, seed = result
try:
if upscale is not None:
from ldm.gfpgan.gfpgan_tools import (
real_esrgan_upscale,
)
if len(upscale) < 2:
upscale.append(0.75)
image = real_esrgan_upscale(
image,
upscale[1],
int(upscale[0]),
prompt,
seed,
)
if gfpgan_strength > 0:
from ldm.gfpgan.gfpgan_tools import _run_gfpgan
image = _run_gfpgan(
image, gfpgan_strength, prompt, seed, 1
)
except Exception as e:
print(
f'>> Error running RealESRGAN - Your image was not upscaled.\n{e}'
)
if image_callback is not None:
if save_original:
image_callback(image, seed)
else:
image_callback(image, seed, upscaled=True)
else: # no callback passed, so we simply replace old image with rescaled one
result[0] = image
except KeyboardInterrupt:
print('*interrupted*')
print(
'>> Partial results will be returned; if --grid was requested, nothing will be returned.'
)
except RuntimeError as e:
print(traceback.format_exc(), file=sys.stderr)
print('>> Are you sure your system has an adequate NVIDIA GPU?')
toc = time.time()
print('>> Usage stats:')
print(
f'>> {len(results)} image(s) generated in', '%4.2fs' % (toc - tic)
)
print(
f'>> Max VRAM used for this generation:',
'%4.2fG' % (torch.cuda.max_memory_allocated() / 1e9),
)
if self.session_peakmem:
self.session_peakmem = max(
self.session_peakmem, torch.cuda.max_memory_allocated()
)
print(
f'>> Max VRAM used since script start: ',
'%4.2fG' % (self.session_peakmem / 1e9),
)
return results
@torch.no_grad()
def _txt2img(
self,
prompt,
steps,
cfg_scale,
ddim_eta,
skip_normalize,
width,
height,
callback,
):
"""
Returns a function returning an image derived from the prompt and the initial image
Return value depends on the seed at the time you call it
"""
sampler = self.sampler
def make_image(x_T):
uc, c = self._get_uc_and_c(prompt, skip_normalize)
shape = [
self.latent_channels,
height // self.downsampling_factor,
width // self.downsampling_factor,
]
samples, _ = sampler.sample(
batch_size=1,
S=steps,
x_T=x_T,
conditioning=c,
shape=shape,
verbose=False,
unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc,
eta=ddim_eta,
img_callback=callback
)
return self._sample_to_image(samples)
return make_image
@torch.no_grad()
def _img2img(
self,
prompt,
steps,
cfg_scale,
ddim_eta,
skip_normalize,
init_latent,
strength,
callback, # Currently not implemented for img2img
):
"""
Returns a function returning an image derived from the prompt and the initial image
Return value depends on the seed at the time you call it
"""
# PLMS sampler not supported yet, so ignore previous sampler
if self.sampler_name != 'ddim':
print(
f">> sampler '{self.sampler_name}' is not yet supported. Using DDIM sampler"
)
sampler = DDIMSampler(self.model, device=self.device)
else:
sampler = self.sampler
sampler.make_schedule(
ddim_num_steps=steps, ddim_eta=ddim_eta, verbose=False
)
t_enc = int(strength * steps)
def make_image(x_T):
uc, c = self._get_uc_and_c(prompt, skip_normalize)
# encode (scaled latent)
z_enc = sampler.stochastic_encode(
init_latent,
torch.tensor([t_enc]).to(self.device),
noise=x_T
)
# decode it
samples = sampler.decode(
z_enc,
c,
t_enc,
img_callback=callback,
unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc,
)
return self._sample_to_image(samples)
return make_image
# TODO: does this actually need to run every loop? does anything in it vary by random seed?
def _get_uc_and_c(self, prompt, skip_normalize):
uc = self.model.get_learned_conditioning([''])
# get weighted sub-prompts
weighted_subprompts = T2I._split_weighted_subprompts(
prompt, skip_normalize)
if len(weighted_subprompts) > 1:
# i dont know if this is correct.. but it works
c = torch.zeros_like(uc)
# normalize each "sub prompt" and add it
for subprompt, weight in weighted_subprompts:
self._log_tokenization(subprompt)
c = torch.add(
c,
self.model.get_learned_conditioning([subprompt]),
alpha=weight,
)
else: # just standard 1 prompt
self._log_tokenization(prompt)
c = self.model.get_learned_conditioning([prompt])
return (uc, c)
def _sample_to_image(self, samples):
x_samples = self.model.decode_first_stage(samples)
x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0)
if len(x_samples) != 1:
raise Exception(
f'>> expected to get a single image, but got {len(x_samples)}')
x_sample = 255.0 * rearrange(
x_samples[0].cpu().numpy(), 'c h w -> h w c'
)
return Image.fromarray(x_sample.astype(np.uint8))
def _new_seed(self):
self.seed = random.randrange(0, np.iinfo(np.uint32).max)
return self.seed
def load_model(self):
"""Load and initialize the model from configuration variables passed at object creation time"""
if self.model is None:
seed_everything(self.seed)
try:
config = OmegaConf.load(self.config)
model = self._load_model_from_config(config, self.weights)
if self.embedding_path is not None:
model.embedding_manager.load(
self.embedding_path, self.full_precision
)
self.model = model.to(self.device)
# model.to doesn't change the cond_stage_model.device used to move the tokenizer output, so set it here
self.model.cond_stage_model.device = self.device
except AttributeError as e:
print(f'>> Error loading model. {str(e)}', file=sys.stderr)
print(traceback.format_exc(), file=sys.stderr)
raise SystemExit from e
self._set_sampler()
return self.model
# returns a tensor filled with random numbers from a normal distribution
def _get_noise(self,init_latent,width,height):
if init_latent is not None:
if self.device.type == 'mps':
return torch.randn_like(init_latent, device='cpu').to(self.device)
else:
return torch.randn_like(init_latent, device=self.device)
else:
if self.device.type == 'mps':
return torch.randn([1,
self.latent_channels,
height // self.downsampling_factor,
width // self.downsampling_factor],
device='cpu').to(self.device)
else:
return torch.randn([1,
self.latent_channels,
height // self.downsampling_factor,
width // self.downsampling_factor],
device=self.device)
def _set_sampler(self):
msg = f'>> Setting Sampler to {self.sampler_name}'
if self.sampler_name == 'plms':
self.sampler = PLMSSampler(self.model, device=self.device)
elif self.sampler_name == 'ddim':
self.sampler = DDIMSampler(self.model, device=self.device)
elif self.sampler_name == 'k_dpm_2_a':
self.sampler = KSampler(
self.model, 'dpm_2_ancestral', device=self.device
)
elif self.sampler_name == 'k_dpm_2':
self.sampler = KSampler(self.model, 'dpm_2', device=self.device)
elif self.sampler_name == 'k_euler_a':
self.sampler = KSampler(
self.model, 'euler_ancestral', device=self.device
)
elif self.sampler_name == 'k_euler':
self.sampler = KSampler(self.model, 'euler', device=self.device)
elif self.sampler_name == 'k_heun':
self.sampler = KSampler(self.model, 'heun', device=self.device)
elif self.sampler_name == 'k_lms':
self.sampler = KSampler(self.model, 'lms', device=self.device)
else:
msg = f'>> Unsupported Sampler: {self.sampler_name}, Defaulting to plms'
self.sampler = PLMSSampler(self.model, device=self.device)
print(msg)
def _load_model_from_config(self, config, ckpt):
print(f'>> Loading model from {ckpt}')
pl_sd = torch.load(ckpt, map_location='cpu')
# if "global_step" in pl_sd:
# print(f"Global Step: {pl_sd['global_step']}")
sd = pl_sd['state_dict']
model = instantiate_from_config(config.model)
m, u = model.load_state_dict(sd, strict=False)
model.to(self.device)
model.eval()
if self.full_precision:
print(
'Using slower but more accurate full-precision math (--full_precision)'
)
else:
print(
'>> Using half precision math. Call with --full_precision to use more accurate but VRAM-intensive full precision.'
)
model.half()
return model
def _load_img(self, path, width, height, fit=False):
with Image.open(path) as img:
image = img.convert('RGB')
print(
f'>> loaded input image of size {image.width}x{image.height} from {path}'
)
# The logic here is:
# 1. If "fit" is true, then the image will be fit into the bounding box defined
# by width and height. It will do this in a way that preserves the init image's
# aspect ratio while preventing letterboxing. This means that if there is
# leftover horizontal space after rescaling the image to fit in the bounding box,
# the generated image's width will be reduced to the rescaled init image's width.
# Similarly for the vertical space.
# 2. Otherwise, if "fit" is false, then the image will be scaled, preserving its
# aspect ratio, to the nearest multiple of 64. Large images may generate an
# unexpected OOM error.
if fit:
image = self._fit_image(image,(width,height))
else:
image = self._squeeze_image(image)
image = np.array(image).astype(np.float32) / 255.0
image = image[None].transpose(0, 3, 1, 2)
image = torch.from_numpy(image)
return 2.0 * image - 1.0
def _squeeze_image(self,image):
x,y,resize_needed = self._resolution_check(image.width,image.height)
if resize_needed:
return InitImageResizer(image).resize(x,y)
return image
def _fit_image(self,image,max_dimensions):
w,h = max_dimensions
print(
f'>> image will be resized to fit inside a box {w}x{h} in size.'
)
if image.width > image.height:
h = None # by setting h to none, we tell InitImageResizer to fit into the width and calculate height
elif image.height > image.width:
w = None # ditto for w
else:
pass
image = InitImageResizer(image).resize(w,h) # note that InitImageResizer does the multiple of 64 truncation internally
print(
f'>> after adjusting image dimensions to be multiples of 64, init image is {image.width}x{image.height}'
)
return image
# TO DO: Move this and related weighted subprompt code into its own module.
def _split_weighted_subprompts(text, skip_normalize=False):
"""
grabs all text up to the first occurrence of ':'
uses the grabbed text as a sub-prompt, and takes the value following ':' as weight
if ':' has no value defined, defaults to 1.0
repeats until no text remaining
"""
prompt_parser = re.compile("""
(?P<prompt> # capture group for 'prompt'
(?:\\\:|[^:])+ # match one or more non ':' characters or escaped colons '\:'
) # end 'prompt'
(?: # non-capture group
:+ # match one or more ':' characters
(?P<weight> # capture group for 'weight'
-?\d+(?:\.\d+)? # match positive or negative integer or decimal number
)? # end weight capture group, make optional
\s* # strip spaces after weight
| # OR
$ # else, if no ':' then match end of line
) # end non-capture group
""", re.VERBOSE)
parsed_prompts = [(match.group("prompt").replace("\\:", ":"), float(
match.group("weight") or 1)) for match in re.finditer(prompt_parser, text)]
if skip_normalize:
return parsed_prompts
weight_sum = sum(map(lambda x: x[1], parsed_prompts))
if weight_sum == 0:
print(
"Warning: Subprompt weights add up to zero. Discarding and using even weights instead.")
equal_weight = 1 / len(parsed_prompts)
return [(x[0], equal_weight) for x in parsed_prompts]
return [(x[0], x[1] / weight_sum) for x in parsed_prompts]
# shows how the prompt is tokenized
# usually tokens have '</w>' to indicate end-of-word,
# but for readability it has been replaced with ' '
def _log_tokenization(self, text):
if not self.log_tokenization:
return
tokens = self.model.cond_stage_model.tokenizer._tokenize(text)
tokenized = ""
discarded = ""
usedTokens = 0
totalTokens = len(tokens)
for i in range(0, totalTokens):
token = tokens[i].replace('</w>', ' ')
# alternate color
s = (usedTokens % 6) + 1
if i < self.model.cond_stage_model.max_length:
tokenized = tokenized + f"\x1b[0;3{s};40m{token}"
usedTokens += 1
else: # over max token length
discarded = discarded + f"\x1b[0;3{s};40m{token}"
print(f"\nTokens ({usedTokens}):\n{tokenized}\x1b[0m")
if discarded != "":
print(
f"Tokens Discarded ({totalTokens-usedTokens}):\n{discarded}\x1b[0m")
def _resolution_check(self, width, height, log=False):
resize_needed = False
w, h = map(
lambda x: x - x % 64, (width, height)
) # resize to integer multiple of 64
if h != height or w != width:
if log:
print(
f'>> Provided width and height must be multiples of 64. Auto-resizing to {w}x{h}'
)
height = h
width = w
resize_needed = True
if (width * height) > (self.width * self.height):
print(">> This input is larger than your defaults. If you run out of memory, please use a smaller image.")
return width, height, resize_needed
def slerp(self, t, v0, v1, DOT_THRESHOLD=0.9995):
'''
Spherical linear interpolation
Args:
t (float/np.ndarray): Float value between 0.0 and 1.0
v0 (np.ndarray): Starting vector
v1 (np.ndarray): Final vector
DOT_THRESHOLD (float): Threshold for considering the two vectors as
colineal. Not recommended to alter this.
Returns:
v2 (np.ndarray): Interpolation vector between v0 and v1
'''
inputs_are_torch = False
if not isinstance(v0, np.ndarray):
inputs_are_torch = True
v0 = v0.detach().cpu().numpy()
if not isinstance(v1, np.ndarray):
inputs_are_torch = True
v1 = v1.detach().cpu().numpy()
dot = np.sum(v0 * v1 / (np.linalg.norm(v0) * np.linalg.norm(v1)))
if np.abs(dot) > DOT_THRESHOLD:
v2 = (1 - t) * v0 + t * v1
else:
theta_0 = np.arccos(dot)
sin_theta_0 = np.sin(theta_0)
theta_t = theta_0 * t
sin_theta_t = np.sin(theta_t)
s0 = np.sin(theta_0 - theta_t) / sin_theta_0
s1 = sin_theta_t / sin_theta_0
v2 = s0 * v0 + s1 * v1
if inputs_are_torch:
v2 = torch.from_numpy(v2).to(self.device)
return v2
class T2I(Generate):
def __init__(self,**kwargs):
print(f'>> The ldm.simplet2i module is deprecated. Use ldm.generate instead. It is a drop-in replacement.')
super().__init__(kwargs)