[FeatureRequest] support FLUX.1 text to image model

[mzbac]

It would be great if MLX supported the new FLUX.1 text-to-image model, as it seems to require a fairly large amount of VRAM to run ( would be advantage on Apple Silicon given unified memory). I don't have much experience with diffusion models, but I am happy to help in any way possible.

[mark-lord]

Seconded! I managed to get Flux sort of working on my M1 Max using ComfyUI, but the speed was abysmal (presumably very unoptimised since the vast majority of the community is CUDA-based). Not to mention the output is unfortunately broken. A dedicated MLX implementation would be faster and, hopefully, not broken :)

[Edit: Eventually got it working, was a PyTorch issue as mentioned by Vargol]

[mzbac]

Just found out DiffusionKit is already working on Flux. Hopefully, it will be supported soon
argmaxinc/DiffusionKit#11

[QueryType]

I too tried. And ran into issues.
comfyanonymous/ComfyUI#4165 (comment)

Seems we need to be patient.

[mgierschdev]

Following the topic too

[Vargol]

The ComfyUI noisy image issues are a PyTorch issue, and can be fixed by downgrading to PyTorch 2.3.1 or (according to one person, I've not tried it myself) upgrading to macOS 15 beta and a PyTorch Nightly, you get the same issue with Diffusers.

But +1 for a FLUX.1 MLX example

[mzbac]

Closed as supported by https://github.com/filipstrand/mflux and https://github.com/argmaxinc/DiffusionKit