This commit enhances support for V2 variant (epsilon and v-predict)
import and conversion to diffusers, by prompting the user to select
the proper config file during startup time autoimport as well as
in the invokeai installer script..
This PR ports the `main` PR #2871 to the v2.3 branch. This adjusts
the global diffusers model cache to work with the 0.14 diffusers
layout of placing models in HF_HOME/hub rather than HF_HOME/diffusers.
When a legacy ckpt model was converted into diffusers in RAM, the
built-in NSFW checker was not being disabled, in contrast to models
converted and saved to disk. Because InvokeAI does its NSFW checking
as a separate post-processing step (in order to generate blurred
images rather than black ones), this defeated the
--nsfw and --no-nsfw switches.
This closes#2836 and #2580.
- Crash would occur at the end of this sequence:
- launch CLI
- !convert <URL pointing to a legacy ckpt file>
- Answer "Y" when asked to delete original .ckpt file
- This commit modifies model_manager.heuristic_import()
to silently delete the downloaded legacy file after
it has been converted into a diffusers model. The user
is no longer asked to approve deletion.
NB: This should be cherry-picked into main once refactor
is done.
- Final list can be found in invokeai/configs/INITIAL_MODELS.yaml
- After installing all the models, I discovered a bug in the file
selection form that caused a crash when no remaining uninstalled
models remained. So had to fix this.
- Discord member @marcus.llewellyn reported that some civitai 2.1-derived checkpoints were
not converting properly (probably dreambooth-generated):
https://discord.com/channels/1020123559063990373/1078386197589655582/1078387806122025070
- @blessedcoolant tracked this down to a missing key that was used to
derive vector length of the CLIP model used by fetching the second
dimension of the tensor at "cond_stage_model.model.text_projection".
His proposed solution was to hardcode a value of 1024.
- On inspection, I found that the same second dimension can be
recovered from key 'cond_stage_model.model.ln_final.bias', and use
that instead. I hope this is correct; tested on multiple v1, v2 and
inpainting models and they converted correctly.
- While debugging this, I found and fixed several other issues:
- model download script was not pre-downloading the OpenCLIP
text_encoder or text_tokenizer. This is fixed.
- got rid of legacy code in `ckpt_to_diffuser.py` and replaced
with calls into `model_manager`
- more consistent status reporting in the CLI.
Root directory finding algorithm is:
2) use --root argument
2) use INVOKEAI_ROOT environment variable
3) use VIRTUAL_ENV environment variable
4) use ~/invokeai
Since developer's are liable to put virtual environments in their
favorite places, not necessarily in the invokeai root directory, this
PR adds a sanity check that looks for the existence of
VIRTUAL_ENV/invokeai.init, and moves to (4) if not found.
- Fix a bug in the CLI which prevented diffusers imported by their repo_ids
from being correctly registered in the current session (though they install
correctly)
- Capitalize the "i" in Imported in the autogenerated descriptions.
1. resize installer window to give more room for configure and download forms
2. replace '\' with '/' in directory names to allow user to drag-and-drop
folders into the dialogue boxes that accept directories.
3. similar change in CLI for the !import_model and !convert_model commands
4. better error reporting when a model download fails due to network errors
5. put the launcher scripts into a loop so that menu reappears after
invokeai, merge script, etc exits. User can quit with "Q".
6. do not try to download fp16 of sd-ft-mse-vae, since it doesn't exist.
7. cleaned up status reporting when installing models
- Detect when install failed for some reason and print helpful error
message rather than stack trace.
- Detect window size and resize to minimum acceptable values to provide
better display of configure and install forms.
