* WIP: feat(flux2): add FLUX 2 Kontext model support
- Add new invocation nodes for FLUX 2:
- flux2_denoise: Denoising invocation for FLUX 2
- flux2_klein_model_loader: Model loader for Klein architecture
- flux2_klein_text_encoder: Text encoder for Qwen3-based encoding
- flux2_vae_decode: VAE decoder for FLUX 2
- Add backend support:
- New flux2 module with denoise and sampling utilities
- Extended model manager configs for FLUX 2 models
- Updated model loaders for Klein architecture
- Update frontend:
- Extended graph builder for FLUX 2 support
- Added FLUX 2 model types and configurations
- Updated readiness checks and UI components
* fix(flux2): correct VAE decode with proper BN denormalization
FLUX.2 VAE uses Batch Normalization in the patchified latent space
(128 channels). The decode must:
1. Patchify latents from (B, 32, H, W) to (B, 128, H/2, W/2)
2. Apply BN denormalization using running_mean/running_var
3. Unpatchify back to (B, 32, H, W) for VAE decode
Also fixed image normalization from [-1, 1] to [0, 255].
This fixes washed-out colors in generated FLUX.2 Klein images.
* feat(flux2): add FLUX.2 Klein model support with ComfyUI checkpoint compatibility
- Add FLUX.2 transformer loader with BFL-to-diffusers weight conversion
- Fix AdaLayerNorm scale-shift swap for final_layer.adaLN_modulation weights
- Add VAE batch normalization handling for FLUX.2 latent normalization
- Add Qwen3 text encoder loader with ComfyUI FP8 quantization support
- Add frontend components for FLUX.2 Klein model selection
- Update configs and schema for FLUX.2 model types
* Chore Ruff
* Fix Flux1 vae probing
* Fix Windows Paths schema.ts
* Add 4B und 9B klein to Starter Models.
* feat(flux2): add non-commercial license indicator for FLUX.2 Klein 9B
- Add isFlux2Klein9BMainModelConfig and isNonCommercialMainModelConfig functions
- Update MainModelPicker and InitialStateMainModelPicker to show license icon
- Update license tooltip text to include FLUX.2 Klein 9B
* feat(flux2): add Klein/Qwen3 variant support and encoder filtering
Backend:
- Add klein_4b/klein_9b variants for FLUX.2 Klein models
- Add qwen3_4b/qwen3_8b variants for Qwen3 encoder models
- Validate encoder variant matches Klein model (4B↔4B, 9B↔8B)
- Auto-detect Qwen3 variant from hidden_size during probing
Frontend:
- Show variant field for all model types in ModelView
- Filter Qwen3 encoder dropdown to only show compatible variants
- Update variant type definitions (zFlux2VariantType, zQwen3VariantType)
- Remove unused exports (isFluxDevMainModelConfig, isFlux2Klein9BMainModelConfig)
* Chore Ruff
* feat(flux2): add Klein 9B Base (undistilled) variant support
Distinguish between FLUX.2 Klein 9B (distilled) and Klein 9B Base (undistilled)
models by checking guidance_embeds in diffusers config or guidance_in keys in
safetensors. Klein 9B Base requires more steps but offers higher quality.
* feat(flux2): improve diffusers compatibility and distilled model support
Backend changes:
- Update text encoder layers from [9,18,27] to (10,20,30) matching diffusers
- Use apply_chat_template with system message instead of manual formatting
- Change position IDs from ones to zeros to match diffusers implementation
- Add get_schedule_flux2() with empirical mu computation for proper schedule shifting
- Add txt_embed_scale parameter for Qwen3 embedding magnitude control
- Add shift_schedule toggle for base (28+ steps) vs distilled (4 steps) models
- Zero out guidance_embedder weights for Klein models without guidance_embeds
UI changes:
- Clear Klein VAE and Qwen3 encoder when switching away from flux2 base
- Clear Qwen3 encoder when switching between different Klein model variants
- Add toast notification informing user to select compatible encoder
* feat(flux2): fix distilled model scheduling with proper dynamic shifting
- Configure scheduler with FLUX.2 Klein parameters from scheduler_config.json
(use_dynamic_shifting=True, shift=3.0, time_shift_type="exponential")
- Pass mu parameter to scheduler.set_timesteps() for resolution-aware shifting
- Remove manual shift_schedule parameter (scheduler handles this automatically)
- Simplify get_schedule_flux2() to return linear sigmas only
- Remove txt_embed_scale parameter (no longer needed)
This matches the diffusers Flux2KleinPipeline behavior where the
FlowMatchEulerDiscreteScheduler applies dynamic timestep shifting
based on image resolution via the mu parameter.
Fixes 4-step distilled Klein 9B model quality issues.
* fix(ui): fix FLUX.1 graph building with posCondCollect node lookup
The posCondCollect node was created with getPrefixedId() which generates
a random suffix (e.g., 'pos_cond_collect:abc123'), but g.getNode() was
called with the plain string 'pos_cond_collect', causing a node lookup
failure.
Fix by declaring posCondCollect as a module-scoped variable and
referencing it directly instead of using g.getNode().
* Remove Flux2 Klein Base from Starter Models
* Remove Logging
* Add Default Values for Flux2 Klein and add variant as additional info to from_base
* Add migrations for the z-image qwen3 encoder without a variant value
* Add img2img, inpainting and outpainting support for FLUX.2 Klein
- Add flux2_vae_encode invocation for encoding images to FLUX.2 latents
- Integrate inpaint_extension into FLUX.2 denoise loop for proper mask handling
- Apply BN normalization to init_latents and noise for consistency in inpainting
- Use manual Euler stepping for img2img/inpaint to preserve exact timestep schedule
- Add flux2_img2img, flux2_inpaint, flux2_outpaint generation modes
- Expand starter models with FP8 variants, standalone transformers, and separate VAE/encoders
- Fix outpainting to always use full denoising (0-1) since strength doesn't apply
- Improve error messages in model loader with clear guidance for standalone models
* Add GGUF quantized model support and Diffusers VAE loader for FLUX.2 Klein
- Add Main_GGUF_Flux2_Config for GGUF-quantized FLUX.2 transformer models
- Add VAE_Diffusers_Flux2_Config for FLUX.2 VAE in diffusers format
- Add Flux2GGUFCheckpointModel loader with BFL-to-diffusers conversion
- Add Flux2VAEDiffusersLoader for AutoencoderKLFlux2
- Add FLUX.2 Klein 4B/9B hardware requirements to documentation
- Update starter model descriptions to clarify dependencies install together
- Update frontend schema for new model configs
* Fix FLUX.2 model detection and add FP8 weight dequantization support
- Improve FLUX.2 variant detection for GGUF/checkpoint models (BFL format keys)
- Fix guidance_embeds logic: distilled=False, undistilled=True
- Add FP8 weight dequantization for ComfyUI-style quantized models
- Prevent FLUX.2 models from being misidentified as FLUX.1
- Preserve user-editable fields (name, description, etc.) on model reidentify
- Improve Qwen3Encoder detection by variant in starter models
- Add defensive checks for tensor operations
* Chore ruff format
* Chore Typegen
* Fix FLUX.2 Klein 9B model loading by detecting hidden_size from weights
Previously num_attention_heads was hardcoded to 24, which is correct for
Klein 4B but causes size mismatches when loading Klein 9B checkpoints.
Now dynamically calculates num_attention_heads from the hidden_size
dimension of context_embedder weights:
- Klein 4B: hidden_size=3072 → num_attention_heads=24
- Klein 9B: hidden_size=4096 → num_attention_heads=32
Fixes both Checkpoint and GGUF loaders for FLUX.2 models.
* Only clear Qwen3 encoder when FLUX.2 Klein variant changes
Previously the encoder was cleared whenever switching between any Klein
models, even if they had the same variant. Now compares the variant of
the old and new model and only clears the encoder when switching between
different variants (e.g., klein_4b to klein_9b).
This allows users to switch between different Klein 9B models without
having to re-select the Qwen3 encoder each time.
* Add metadata recall support for FLUX.2 Klein parameters
The scheduler, VAE model, and Qwen3 encoder model were not being
recalled correctly for FLUX.2 Klein images. This adds dedicated
metadata handlers for the Klein-specific parameters.
* Fix FLUX.2 Klein denoising scaling and Z-Image VAE compatibility
- Apply exponential denoising scaling (exponent 0.2) to FLUX.2 Klein,
matching FLUX.1 behavior for more intuitive inpainting strength
- Add isFlux1VAEModelConfig type guard to filter FLUX 1.0 VAEs only
- Restrict Z-Image VAE selection to FLUX 1.0 VAEs, excluding FLUX.2
Klein 32-channel VAEs which are incompatible
* chore pnpm fix
* Add FLUX.2 Klein to starter bundles and documentation
- Add FLUX.2 Klein hardware requirements to quick start guide
- Create flux2_klein_bundle with GGUF Q4 model, VAE, and Qwen3 encoder
- Add "What's New" entry announcing FLUX.2 Klein support
* Add FLUX.2 Klein built-in reference image editing support
FLUX.2 Klein has native multi-reference image editing without requiring
a separate model (unlike FLUX.1 which needs a Kontext model).
Backend changes:
- Add Flux2RefImageExtension for encoding reference images with FLUX.2 VAE
- Apply BN normalization to reference image latents for correct scaling
- Use T-coordinate offset scale=10 like diffusers (T=10, 20, 30...)
- Concatenate reference latents with generated image during denoising
- Extract only generated portion in step callback for correct preview
Frontend changes:
- Add flux2_reference_image config type without model field
- Hide model selector for FLUX.2 reference images (built-in support)
- Add type guards to handle configs without model property
- Update validators to skip model validation for FLUX.2
- Add 'flux2' to SUPPORTS_REF_IMAGES_BASE_MODELS
* Chore windows path fix
* Add reference image resizing for FLUX.2 Klein
Resize large reference images to match BFL FLUX.2 sampling.py limits:
- Single reference: max 2024² pixels (~4.1M)
- Multiple references: max 1024² pixels (~1M)
Uses same scaling approach as BFL's cap_pixels() function.
Changed the default value of model_cache_keep_alive from 0 (indefinite)
to 5 minutes as requested. This means models will now be automatically
cleared from cache after 5 minutes of inactivity by default, unless
users explicitly configure a different value.
Users can still set it to 0 in their config to get the old behavior
of keeping models indefinitely.
Co-authored-by: lstein <111189+lstein@users.noreply.github.com>
* (bugfix)(mm) work around Windows being unable to rmtree tmp directories after GGUF install
* (style) fix ruff error
* (fix) add workaround for Windows Permission Denied on GGUF file move() call
* (fix) perform torch copy() in GGUF reader to avoid deletion failures on Windows
* (style) fix ruff formatting issues
* fix(model-install): support multi-subfolder downloads for Z-Image Qwen3 encoder
The Z-Image Qwen3 text encoder requires both text_encoder and tokenizer
subfolders from the HuggingFace repo, but the previous implementation
only downloaded the text_encoder subfolder, causing model identification
to fail.
Changes:
- Add subfolders property to HFModelSource supporting '+' separated paths
- Extend filter_files() and download_urls() to handle multiple subfolders
- Update _multifile_download() to preserve subfolder structure
- Make Qwen3Encoder probe check both nested and direct config.json paths
- Update Qwen3EncoderLoader to handle both directory structures
- Change starter model source to text_encoder+tokenizer
* ruff format
* fix schema description
* fix schema description
---------
Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>
- Remove all trailing whitespace (W293 errors)
- Add debug logging when timeout fires but activity detected
- Add debug logging when timeout fires but cache is empty
- Only log "Clearing model cache" message when actually clearing
- Prevents misleading timeout messages during active generation
Co-authored-by: lstein <111189+lstein@users.noreply.github.com>
* Rework graph, add documentation
* Minor fixes to README.md
* Updated schema
* Fixed test to match behavior - all nodes executed, parents before children
* Update invokeai/app/services/shared/graph.py
Cleaned up code
Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>
* Change silent corrections to enforcing invariants
---------
Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>
Add route and model record service method to reidentify a model. This
re-probes the model files and replaces the model's config with the new
one if it does not error.
* feat(mm): add UnknownModelConfig
* refactor(ui): move model categorisation-ish logic to central location, simplify model manager models list
* refactor(ui)refactor(ui): more cleanup of model categories
* refactor(ui): remove unused excludeSubmodels
I can't remember what this was for and don't see any reference to it.
Maybe it's just remnants from a previous implementation?
* feat(nodes): add unknown as model base
* chore(ui): typegen
* feat(ui): add unknown model base support in ui
* feat(ui): allow changing model type in MM, fix up base and variant selects
* feat(mm): omit model description instead of making it "base type filename model"
* feat(app): add setting to allow unknown models
* feat(ui): allow changing model format in MM
* feat(app): add the installed model config to install complete events
* chore(ui): typegen
* feat(ui): toast warning when installed model is unidentified
* docs: update config docstrings
* chore(ui): typegen
* tests(mm): fix test for MM, leave the UnknownModelConfig class in the list of configs
* tidy(ui): prefer types from zod schemas for model attrs
* chore(ui): lint
* fix(ui): wrong translation string
* feat(mm): normalized model storage
Store models in a flat directory structure. Each model is in a dir named
its unique key (a UUID). Inside that dir is either the model file or the
model dir.
* feat(mm): add migration to flat model storage
* fix(mm): normalized multi-file/diffusers model installation no worky
now worky
* refactor: port MM probes to new api
- Add concept of match certainty to new probe
- Port CLIP Embed models to new API
- Fiddle with stuff
* feat(mm): port TIs to new API
* tidy(mm): remove unused probes
* feat(mm): port spandrel to new API
* fix(mm): parsing for spandrel
* fix(mm): loader for clip embed
* fix(mm): tis use existing weight_files method
* feat(mm): port vae to new API
* fix(mm): vae class inheritance and config_path
* tidy(mm): patcher types and import paths
* feat(mm): better errors when invalid model config found in db
* feat(mm): port t5 to new API
* feat(mm): make config_path optional
* refactor(mm): simplify model classification process
Previously, we had a multi-phase strategy to identify models from their
files on disk:
1. Run each model config classes' `matches()` method on the files. It
checks if the model could possibly be an identified as the candidate
model type. This was intended to be a quick check. Break on the first
match.
2. If we have a match, run the config class's `parse()` method. It
derive some additional model config attrs from the model files. This was
intended to encapsulate heavier operations that may require loading the
model into memory.
3. Derive the common model config attrs, like name, description,
calculate the hash, etc. Some of these are also heavier operations.
This strategy has some issues:
- It is not clear how the pieces fit together. There is some
back-and-forth between different methods and the config base class. It
is hard to trace the flow of logic until you fully wrap your head around
the system and therefore difficult to add a model architecture to the
probe.
- The assumption that we could do quick, lightweight checks before
heavier checks is incorrect. We often _must_ load the model state dict
in the `matches()` method. So there is no practical perf benefit to
splitting up the responsibility of `matches()` and `parse()`.
- Sometimes we need to do the same checks in `matches()` and `parse()`.
In these cases, splitting the logic is has a negative perf impact
because we are doing the same work twice.
- As we introduce the concept of an "unknown" model config (i.e. a model
that we cannot identify, but still record in the db; see #8582), we will
_always_ run _all_ the checks for every model. Therefore we need not try
to defer heavier checks or resource-intensive ops like hashing. We are
going to do them anyways.
- There are situations where a model may match multiple configs. One
known case are SD pipeline models with merged LoRAs. In the old probe
API, we relied on the implicit order of checks to know that if a model
matched for pipeline _and_ LoRA, we prefer the pipeline match. But, in
the new API, we do not have this implicit ordering of checks. To resolve
this in a resilient way, we need to get all matches up front, then use
tie-breaker logic to figure out which should win (or add "differential
diagnosis" logic to the matchers).
- Field overrides weren't handled well by this strategy. They were only
applied at the very end, if a model matched successfully. This means we
cannot tell the system "Hey, this model is type X with base Y. Trust me
bro.". We cannot override the match logic. As we move towards letting
users correct mis-identified models (see #8582), this is a requirement.
We can simplify the process significantly and better support "unknown"
models.
Firstly, model config classes now have a single `from_model_on_disk()`
method that attempts to construct an instance of the class from the
model files. This replaces the `matches()` and `parse()` methods.
If we fail to create the config instance, a special exception is raised
that indicates why we think the files cannot be identified as the given
model config class.
Next, the flow for model identification is a bit simpler:
- Derive all the common fields up-front (name, desc, hash, etc).
- Merge in overrides.
- Call `from_model_on_disk()` for every config class, passing in the
fields. Overrides are handled in this method.
- Record the results for each config class and choose the best one.
The identification logic is a bit more verbose, with the special
exceptions and handling of overrides, but it is very clear what is
happening.
The one downside I can think of for this strategy is we do need to check
every model type, instead of stopping at the first match. It's a bit
less efficient. In practice, however, this isn't a hot code path, and
the improved clarity is worth far more than perf optimizations that the
end user will likely never notice.
* refactor(mm): remove unused methods in config.py
* refactor(mm): add model config parsing utils
* fix(mm): abstractmethod bork
* tidy(mm): clarify that model id utils are private
* fix(mm): fall back to UnknownModelConfig correctly
* feat(mm): port CLIPVisionDiffusersConfig to new api
* feat(mm): port SigLIPDiffusersConfig to new api
* feat(mm): make match helpers more succint
* feat(mm): port flux redux to new api
* feat(mm): port ip adapter to new api
* tidy(mm): skip optimistic override handling for now
* refactor(mm): continue iterating on config
* feat(mm): port flux "control lora" and t2i adapter to new api
* tidy(ui): use Extract to get model config types
* fix(mm): t2i base determination
* feat(mm): port cnet to new api
* refactor(mm): add config validation utils, make it all consistent and clean
* feat(mm): wip port of main models to new api
* feat(mm): wip port of main models to new api
* feat(mm): wip port of main models to new api
* docs(mm): add todos
* tidy(mm): removed unused model merge class
* feat(mm): wip port main models to new api
* tidy(mm): clean up model heuristic utils
* tidy(mm): clean up ModelOnDisk caching
* tidy(mm): flux lora format util
* refactor(mm): make config classes narrow
Simpler logic to identify, less complexity to add new model, fewer
useless attrs that do not relate to the model arch, etc
* refactor(mm): diffusers loras
w
* feat(mm): consistent naming for all model config classes
* fix(mm): tag generation & scattered probe fixes
* tidy(mm): consistent class names
* refactor(mm): split configs into separate files
* docs(mm): add comments for identification utils
* chore(ui): typegen
* refactor(mm): remove legacy probe, new configs dir structure, update imports
* fix(mm): inverted condition
* docs(mm): update docsstrings in factory.py
* docs(mm): document flux variant attr
* feat(mm): add helper method for legacy configs
* feat(mm): satisfy type checker in flux denoise
* docs(mm): remove extraneous comment
* fix(mm): ensure unknown model configs get unknown attrs
* fix(mm): t5 identification
* fix(mm): sdxl ip adapter identification
* feat(mm): more flexible config matching utils
* fix(mm): clip vision identification
* feat(mm): add sanity checks before probing paths
* docs(mm): add reminder for self for field migrations
* feat(mm): clearer naming for main config class hierarchy
* feat(mm): fix clip vision starter model bases, add ref to actual models
* feat(mm): add model config schema migration logic
* fix(mm): duplicate import
* refactor(mm): split big migration into 3
Split the big migration that did all of these things into 3:
- Migration 22: Remove unique contraint on base/name/type in models
table
- Migration 23: Migrate configs to v6.8.0 schemas
- Migration 24: Normalize file storage
* fix(mm): pop base/type/format when creating unknown model config
* fix(db): migration 22 insert only real cols
* fix(db): migration 23 fall back to unknown model when config change fails
* feat(db): run migrations 23 and 24
* fix(mm): false negative on flux lora
* fix(mm): vae checkpoint probe checking for dir instead of file
* fix(mm): ModelOnDisk skips dirs when looking for weights
Previously a path w/ any of the known weights suffixes would be seen as
a weights file, even if it was a directory. We now check to ensure the
candidate path is actually a file before adding it to the list of
weights.
* feat(mm): add method to get main model defaults from a base
* feat(mm): do not log when multiple non-unknown model matches
* refactor(mm): continued iteration on model identifcation
* tests(mm): refactor model identification tests
Overhaul of model identification (probing) tests. Previously we didn't
test the correctness of probing except in a few narrow cases - now we
do.
See tests/model_identification/README.md for a detailed overview of the
new test setup. It includes instructions for adding a new test case. In
brief:
- Download the model you want to add as a test case
- Run a script against it to generate the test model files
- Fill in the expected model type/format/base/etc in the generated test
metadata JSON file
Included test cases:
- All starter models
- A handful of other models that I had installed
- Models present in the previous test cases as smoke tests, now also
tested for correctness
* fix(mm): omit type/format/base when creating unknown config instance
* feat(mm): use ValueError for model id sanity checks
* feat(mm): add flag for updating models to allow class changes
* tests(mm): fix remaining MM tests
* feat: allow users to edit models freely
* feat(ui): add warning for model settings edit
* tests(mm): flux state dict tests
* tidy: remove unused file
* fix(mm): lora state dict loading in model id
* feat(ui): use translation string for model edit warning
* docs(db): update version numbers in migration comments
* chore: bump version to v6.9.0a1
* docs: update model id readme
* tests(mm): attempt to fix windows model id tests
* fix(mm): issue with deleting single file models
* feat(mm): just delete the dir w/ rmtree when deleting model
* tests(mm): windows CI issue
* fix(ui): typegen schema sync
* fix(mm): fixes for migration 23
- Handle CLIP Embed and Main SD models missing variant field
- Handle errors when calling the discriminator function, previously only
handled ValidationError but it could be a ValueError or something else
- Better logging for config migration
* chore: bump version to v6.9.0a2
* chore: bump version to v6.9.0a3
Fixes a test failure introduced by
https://github.com/pydantic/pydantic/pull/11957
TL;DR: "after" model validators should be instance methods, not class
methods. Batch model updated to use an instance method, which fixes the
failing test.
