[Feature] Flux.2 small decoder

[stduhpf]

Feature Summary

A smaller and faster VAE decoder architecture distilled from the original Flux.2 VAE. It seems to be the same overall AutoEncoderKL VAE architecture, but with narrower channels ([96, 192, 384, 384] vs [128, 256, 512, 512]).

Detailed Description

See https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/tree/main, and Comfy-Org/ComfyUI#13314

I'm not 100% sure how useful supporting this would be since TAEF2 already exist and is well supported, but this should in theory be closer to the original VAE since it's also a VAE? (This needs to be tested.)

Alternatives you considered

Using either the full VAE or TAEF2 already work.

Additional context

No response

[JohnLoveJoy]

TAEF2 causes a visible loss of quality.

[Green-Sky]

see #1402

[stduhpf]

@JohnLoveJoy In my experience, TAEF2 looks almost flawless. Actually it had a lot fewer artifacts than the VAE until that was fixed with #1297.

Without looking at metadata, can you really tell which one has visibly lost some quality here?

[JohnLoveJoy]

Both have slight noise in their hair.
But the second one has lighter colors and unnatural patterns on her skin. Perhaps the range of the color palette is worse.

[stduhpf]

But the second one has lighter colors and unnatural patterns on her skin.

And it just happens that the second one is actually the one decoded with the original full Flux.2 VAE. You can check by inspecting metadata.

It is easy to notice that the two images are slightly different, but the difference in visual quality is apparently hard to tell.