Questions about CFG, loss behavior, and data usage in Qwen-Image-Edit distillation

[ThreeSRR]

Hi, thanks for open-sourcing Qwen-Image-Lightning. We are distilling Qwen-Image-Edit with DMD2, and we have a few quick questions about training details:

1. CFG during training
   How do you set real_scale_guidance and fake_scale_guidance in your experiments?
   We use real_scale_guidance = 4 and fake_scale_guidance = 1, but in the later stage of training the generated images show color artifacts and oil-painting–like textures. Is this expected, or do you recommend different values or scheduling strategies?

2. Loss behavior
   In our experiments, the generator loss keeps decreasing, while the critic loss keeps increasing.
   Is this behavior normal in DMD2 distillation, or does it indicate instability or imbalance between the generator and critic?

3. Training data for edit distillation
   When distilling the editing model Qwen-Image-Edit, do you include text-to-image data, or only edit-specific data?
   If T2I data is included, is there a recommended mixing ratio?

Thanks a lot for your help!

[X-niper]

@ThreeSRR

Hi, these are all excellent questions. Our approach may not be optimal—here are the settings we’ve used in training:

CFG: We used a real guidance scale of 4 and a fake guidance scale of 1. During training, we also observed over-saturation issues, which is something we still need to address.

Loss: Our generator loss isn’t very monotonic—it roughly goes through four phases, rising, falling, rising, falling. Meanwhile, the critic loss generally trends upward overall.

Data: When distilling the edit model, we currently use only edit data(both prompts and reference images). However, incorporating prompt-only data could be a good idea.

[ThreeSRR]

Hi @X-niper ,

Thanks for the quick reply! This explanation is very helpful. I’ll try it following your suggestion and report back if I have further findings.

[ThreeSRR]

Hi @X-niper ,

In our experiments, we observed grid-like artifacts in the generated images, especially in regions with complex textures or high-frequency details (as shown below). Have you encountered similar issues during training or evaluation? If so, how did you mitigate them? Do you have any suggestions?

[GaryHuangJL]

Hi @X-niper ,

In our experiments, we observed grid-like artifacts in the generated images, especially in regions with complex textures or high-frequency details (as shown below). Have you encountered similar issues during training or evaluation? If so, how did you mitigate them? Do you have any suggestions?

I'm facing a similar issue. Does anyone know a solution for this?