support HF models - freepik as first example

[anthonywu]

Change

I nominate that we support compatible and trending 3rd party models based on:

1. curation for quality and reputation
2. compatibility

The first model to meet this criteria for me is: https://huggingface.co/Freepik/flux.1-lite-8B-alpha - it's a distilled version of flux.1 dev, read more at the link.

the claim:

Flux.1 Lite, an 8B parameter transformer model distilled from the FLUX.1-dev model. This version uses 7 GB less RAM and runs 23% faster while maintaining the same precision (bfloat16) as the original model.

Tradeoffs

the model configs are hard-coded for now, to hypothetically let anyone choose any huggingface-hosted model is possible but requires some config refactoring, but also invites anyone to attempt an incompatible model and creating Issues tracker noise

For now, I think this can make some users happy (those with less RAM) while mflux can be seen as inclusive of third party releases.

[anthonywu]

from their readme:

Flux.1 Lite is ready to unleash your creativity! For the best results, we strongly recommend using a guidance_scale of 3.5 and setting n_steps between 22 and 30.

I feel conflicted about having this value higher than dev, will play around to see if we can live with 10-14 range and give users faster generations by default.

[anthonywu]

considered naming this FLUX1_FREEPIK_LITE_8B and pointing to the beta/final releases as the model matures. However, in python zen we should be more explicit than not..., so I prefer to still let users know exactly what they have to ask for and receive.

[anthonywu]

Todo:

• Needs a bug fix to allowing guidance param for this model.

[anthonywu]

not able to exactly match the diffusers output for now, I suspect I'm missing a few key parameters.

using the example in the model card: https://huggingface.co/Freepik/flux.1-lite-8B-alpha#text-to-image

mflux-generate -m freepik-lite-8b-alpha --prompt "A close-up image of a green alien with fluorescent skin in the middle of a dark purple forest" --seed 11 --steps 28 --guidance 3.5

creates

while running the diffusers example as is (but changing device to mps) creates

a head scratcher for now, I think this may be worth solving to see where mflux might differ from diffusers on default implementations.

[anthonywu]

closing for now, not interesting enough to prioritize. Issue #101 is much more interesting and can use some of the work already done here.