Questions about DiffusionNFT baseline

[gold3fluoride]

Dear authors,

Thanks for the amazing work and releasing the code!

I am trying to reproduce your work, and I noticed from the qualitative results from your paper that the DiffusionNFT column in pages 18-21 looked systematically blurry or under-denoised, especially compared with the baseline. Paper link

I am wondering if you have ever considered the issue or could provide a good explanation for this degradation. Or could these results have come from a mismatched training / inference configuration? Could you please clarify the setups that created those images?

Thanks!

[weichow23]

Thank you for raising this😊. We agree that the DiffusionNFT images in Figs. 11–14 appear noticeably soft.

1. Configuration. For SD3.5-M, our DiffusionNFT baseline follows the official released implementation and its method-specific configuration. We compare all methods under the same 100-update budget. The official DiffusionNFT release does not provide a Z-Image-Turbo configuration. For Turbo, we therefore combined the DiffusionNFT training recipe with Z-Image-Turbo’s official native settings: BF16, (1024^2), nine-step Euler sampling, and guidance scale 0. The qualitative comparisons use the same prompts and initial latents across methods, and DiffusionNFT is rendered with the full native inference schedule. Thus, the softness is not due to using fewer inference steps. (Z-Image-Turbo model card; our paper, App. C.1). We ported the released DiffusionNFT implementation and combined it with Z-Image-Turbo’s official model settings, but did not extensively tune a new DiffusionNFT recipe specifically for Turbo. We therefore cannot rule out a better model-specific configuration. If the community develops a stronger, verified implementation, we would welcome it, rerun the comparison, and update both the paper and baseline.
2. Possible explanation. Z-Image-Turbo is a few-step distilled model. DiffusionNFT’s forward-noising and endpoint-fitting objective includes an SFT-like component that may interfere with the denoising dynamics calibrated by Turbo’s distillation, weakening fine detail. Generally speaking, models obtained via step-wise distillation tend to become "blurry" after SFT, similar to the behavior observed in our paper. (Z-Image technical report)
3. DiffusionOPSD also outperforms DiffusionNFT based on the official SD3.5-M configuration. Importantly, the SD3.5-M result does not depend on the Z-Image-Turbo adaptation: under the official DiffusionNFT recipe and matched 100-update protocol, DiffusionOPSD outperforms DiffusionNFT in all ten reward-specific comparisons in Table 1. (Our paper, Table 1)