DiffSynth-Studio After training DiffSynth-Studio/Qwen-Image-Distill-LoRA on the data, the inferred images are blurry. What could be the reason?

before train after train

Nov 18 '25 09:11 min-star

Same result. What is the reason？

Nov 23 '25 13:11 BryantGary

Same result. What is the reason？

no idea

Nov 25 '25 02:11 min-star

step？

Dec 02 '25 08:12 Deng-Xian-Sheng

the train step config is same with official repo

---Original--- From: @.> Date: Tue, Dec 2, 2025 16:54 PM To: @.>; Cc: @.@.>; Subject: Re: [modelscope/DiffSynth-Studio] After trainingDiffSynth-Studio/Qwen-Image-Distill-LoRA on the data, the inferred images areblurry. What could be the reason? (Issue #1059)

Deng-Xian-Sheng left a comment (modelscope/DiffSynth-Studio#1059)

step？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

Dec 02 '25 08:12 min-star

Why is that ？ Can anyone explain this ？ I also use the same config as official repo.. The mse distill loss is not stable, and the lora seems to make images blurry...

Dec 03 '25 02:12 BryantGary

数据量多少呢，可以贴一些详细的信息，我用了直接蒸馏，rank=32，数据量大约2k(场景单一)，感觉比你的好点，但是还是有些问题，我在调参，但更多的怀疑是直接蒸馏的方式能力不够，简单的mse 太简单粗暴了

Dec 15 '25 07:12 hello-program