release 0.5.0 – comprehensive ModelConfig refactor to support compatible HuggingFace dev/schnell models

[anthonywu]

change

1. a re-attempt of support HF models - freepik as first example #84 but this time I don't have to explicitly endorse any org (i.e. Freepik or Shuttle) It should reduce expectation that the maintainers actually support the models, and eliminate future asks to support yet-another pre-config.
2. support issues like Use other model file based on Flux-schnell? #102 (let users go ahead and test any HF model, as long as they bring the --base-model arg when needed
3. re-establish model config before going into Support Flux.1 Tools #101 to support the Flux.1 Tools

goals

1. keep the prior behavior as-is and pass existing tests and add backwards compatibility test cases
2. maintainers can continue to add the default configs list
3. early adopters can get readme/issues-level hints on how to use reportedly-compatible models

demo: shuttle-3

https://huggingface.co/shuttleai/shuttle-3-diffusion

# didn't expect this to work, given shuttle-3 derives from schnell 
mflux-generate \
  --seed 42 \
  --base-model dev \
  --model shuttleai/shuttle-3-diffusion \
  --steps 4 \
  --prompt "A cat holding a sign that says hello world"

now, running as schnell, actually looks... better??

# same as above but with schnell
mflux-generate \
  --seed 42 \
  --base-model schnell \ 
  --model shuttleai/shuttle-3-diffusion \
  --steps 4 \
  --prompt "A cat holding a sign that says hello world"

performance

• M1 Max 64GB, ~24s/step, 4 steps ~= 1m38s runtime

demo: Freepik

https://huggingface.co/Freepik/flux.1-lite-8B-alpha

mflux-generate --base-model dev --steps 22 --model Freepik/flux.1-lite-8B-alpha --prompt "A cat holding a sign that says hello world"

[anthonywu]

Feels pretty good to me, I will make sure my image gen local tests pass.

Users who asked for #102 please help test!

[anthonywu]

I leave from_alias usable so there's no breakage, it's in the test cases

[anthonywu]

defining the magic numbers

[anthonywu]

other values to come with #101 will be canny-dev, depth-dev and redux-dev - the intention is to have the aliases support first party BFL models according to their naming convention

[anthonywu]

backward compatibility is maintained in the test suite

[anthonywu]

⚠️ turns out this broad assumption about guidance support may be invalid, the assumption is from the official schnell model docs: https://huggingface.co/black-forest-labs/FLUX.1-schnell

but shuttle-3-diffusion claims to support guidance even though it's based on schnell: https://huggingface.co/shuttleai/shuttle-3-diffusion

Shuttle 3 Diffusion uses Flux.1 Schnell as its base.

and their docs allow guidance

[filipstrand]

@anthonywu Hi, sorry for my delayed response! I've been busy with the final things for the Dreambooth training feature so I have not had the time to look at you PR yet but will do that first thing when I have a PR up for that. Your results here look really good with the different models and nice that you could achieve this with a relatively small code change!

Dreabooth is pretty near completion and will be a big update (~2000 added lines now when I checked, but mostly new files in a speparate package, so hopefully not that many conflicts) . There are still more things I would like to improve, but I think it is good enough for a v1 release of this kind of feature. I was thinking of including that in the 0.5 release as well, but let's discuss that once I have a PR up for that feature, hopefully in a few days :)

[anthonywu]

@filipstrand Feel free to merge when you're satisfied with this PR and make edits as needed. I am traveling in the weeks ahead and have limited ability to make significant edits.

[filipstrand]

@filipstrand Feel free to merge when you're satisfied with this PR and make edits as needed. I am traveling in the weeks ahead and have limited ability to make significant edits.

Alright, sounds good!