Support for more image models?

[jjhaggar]

Thanks for CoMPaSS, it’s impressive work!

Are you planning to extend it to more image models, such as Qwen-Image / Qwen-Image-Edit, Stable Diffusion XL, or Wan2.2 (which can also generate single frames), or others?

If you already have others in mind for the future, it’d be great to hear which ones.

[blurgyy]

Hi @jjhaggar,

Thank you for the kind words and for your interest in CoMPaSS!

While I would love to extend support for the latest image generation models, I don't have immediate plans to implement this myself due to other commitments.

However, extending CoMPaSS to these models is certainly feasible, and I'm happy to provide some thoughts for anyone interested in exploring this:

• For the SCOP data engine & dataset: Any model that accepts both image and text as input can be used for training. Our best results were achieved with a small (53MiB) Rank-16 LoRA on FLUX.1-dev, so I suspect it would be straightforward to train a similar LoRA for the Qwen series, especially since it also uses MMDiT as its building blocks.
• For the TENOR module: The effectiveness of TENOR can depend on the specific model architecture and implementation. Our paper (Table 1) shows that Stable Diffusion XL is affected by the text-encoder deficiency we identified, as it uses both CLIP and OpenCLIP. Therefore, SDXL should benefit significantly from a trained TENOR module. Regarding the Qwen/Wan series, I haven't investigated their architectures in detail, so I can't comment definitively. However, TENOR would likely be beneficial if they rely on similar CLIP- or T5-based text encoders.

I will leave this issue open for visibility and to encourage potential community contributions. I'd be happy to review a pull request if someone decides to take this on.

[Immac]

I think just SDXL would create a massive ripple effect for adoption on other models, this is why ELLA failed to become a thing, they never released the SDXL version.

[shockgun92]

wait compass for qwen_image :(