Other Stable Diffusion Version

[Yux1angJi]

Hi. Thanks for your interesting work!
Did you try other stable diffusion like stable diffusion 3 with DiT architectures?

[kliyer-ai]

Hey, we tried our method on DiT (PixArt alpha and sigma) and MMDiT (FLUX.1-dev) and found that while CleanDIFT does lead to an overall performance increase, the feature maps still underperform compared to Stable Diffusion 1 and 2, which is why we did not include it.

[Yux1angJi]

Thanks for your reply.
I also tried using DiT model feature maps for related perception research, and the results were similar to yours—not as good as Stable Diffusion 1 and 2. Did you consider the possible reasons for these results? (like feature maps from Transformer blocks not being suitable for dense prediction tasks)

[KoljaBauer]

Hi, our hypothesis is that the CNN-based U-Net structure of Stable Diffusion fosters local features that fit the task of zero-shot semantic correspondence well. In contrast, DiT features seem to be less local, presumably due to the global self-attention between feature tokens.

[clf28]

Hi, may I ask is there any similar work based on features from DiT and could you share them? I want to try an experiment based on features of models like FLUX.

[kliyer-ai]

You may find this work interesting: https://arxiv.org/abs/2505.18584

[clf28]

Thanks so much for reply!

[xcltql666]

Hey, we tried our method on DiT (PixArt alpha and sigma) and MMDiT (FLUX.1-dev) and found that while CleanDIFT does lead to an overall performance increase, the feature maps still underperform compared to Stable Diffusion 1 and 2, which is why we did not include it.

Thank you for sharing your insights about CleanDIFT with DiT architectures! I’m particularly interested in your experiments with the FLUX.1-dev (MMDiT) version. Would it be possible to share the code or implementation details for that specific variant?
My Email is: 1532481149@qq.com