[enhancement]: flux kontext dev support

[Facetwe]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

No response

What should this feature add?

Hello. Flux kontext dev model was recently released. I would love to see support for it in the future.

Alternatives

No response

Additional Content

No response

[AlgorithmicKing737]

yes please

[magnusviri]

There's several pull requests that added this and it was demoed on the Discord channel. #8134

[hugodopradofernandes]

The quantized versions somehow "already work" in v5.15.0

[chrononaut79]

Yeah, I noticed this as well, but there's a number of pipeline changes with the new model. It appears to be capable of image generation using the default Flux tools, but it's likely missing some magic. For instance, I believe there to be a new tokenizer and a new VAE, as well as a replacement for the default clip prompt interpreter. So chances are what we're seeing isn't really what it's capable of. I think even the denoiser might have been altered, but so the images you're seeing probably aren't representative of what the model is capable of once they tool for it.