(note that this is actually release candidate 7, but I made the mistake
of including an old rc number in the branch and can't easily change it)
## Updating Root directory
- Introduced new mechanism for updating the root directory when
necessary. Currently only used to update the invoke.sh script using new
dialog colors.
- Fixed ROCm torch module version number
## Loading legacy 2.0/2.1 models
- Due to not converting the torch.dtype precision correctly, the
`load_pipeline_from_original_stable_diffusion_ckpt()` was returning
models of dtype float32 regardless of the precision setting. This caused
a precision mismatch crash.
- Problem now fixed (also see #3057 for the same fix to `main`)
## Support for a fourth textual inversion embedding file format
- This variant, exemplified by "easynegative.safetensors" has a single
'embparam' key containing a Tensor.
- Also refactored code to make it easier to read.
- Handle both pickle and safetensor formats.
## Persistent model selection
- To be consistent with WebUI parameter behavior, the currently selected
model is saved on exit and restored on restart for both WebUI and CLI
## Bug fixes
- Name of VAE cache directory was "hug", not "hub". This is fixed.
## VAE fixes
- Allow custom VAEs to be assigned to a legacy model by placing a
like-named vae file adjacent to the checkpoint file.
- The custom VAE will be picked up and incorporated into the diffusers
model if the user chooses to convert/optimize.
## Custom config file loading
- Some of the civitai models instruct users to place a custom .yaml file
adjacent to the checkpoint file. This generally wasn't working because
some of the .yaml files use FrozenCLIPEmbedder rather than
WeightedFrozenCLIPEmbedder, and our FrozenCLIPEmbedder class doesn't
handle the `personalization_config` section used by the the textual
inversion manager. Other .yaml files don't have the
`personalization_config` section at all. Both these issues are
fixed.#1685
## Consistent pytorch version
- There was an inconsistency between the pytorch version requirement in
`pyproject.toml` and the requirement in the installer (which does a
little jiggery-pokery to load torch with the right CUDA/ROCm version
prior to the main pip install. This was causing torch to be installed,
then uninstalled, and reinstalled with a different version number. This
is now fixed.
Instructions:
1. Download LoRA .safetensors files of your choice and place in
`INVOKEAIROOT/loras`. Unlike the draft version of this, the file
names can contain underscores and alphanumerics. Names with
arbitrary unicode characters are not supported.
2. Add `withLora(lora-file-basename,weight)` to your prompt. The
weight is optional and will default to 1.0. A few examples, assuming
that a LoRA file named `loras/sushi.safetensors` is present:
```
family sitting at dinner table eating sushi withLora(sushi,0.9)
family sitting at dinner table eating sushi withLora(sushi, 0.75)
family sitting at dinner table eating sushi withLora(sushi)
```
Multiple `withLora()` prompt fragments are allowed. The weight can be
arbitrarily large, but the useful range is roughly 0.5 to 1.0. Higher
weights make the LoRA's influence stronger.
In my limited testing, I found it useful to reduce the CFG to avoid
oversharpening. Also I got better results when running the LoRA on top
of the model on which it was based during training.
Don't try to load a SD 1.x-trained LoRA into a SD 2.x model, and vice
versa. You will get a nasty stack trace. This needs to be cleaned up.
3. You can change the location of the `loras` directory by passing the
`--lora_directory` option to `invokeai.
Documentation can be found in docs/features/LORAS.md.
