Community trains fix LoRA that salvages Qwen Image 2.1's generation quality

Incognit0ErgoSum · reddit · 2026-09-23

A Reddit user trained a LoRA that fixes Qwen Image 2.1's generation issues and drastically improves edit quality, releasing it with a ComfyUI workflow on Civitai and Hugging Face. Key details:

Trained on 100 high-detail images (1024/768/512), then stripped layers 22-31 via trial and error; it's a DoRA with 32 layers + 16 conv layers that loads as a plain LoRA in ComfyUI
Best settings: 20 steps, seeds2/sgmuniform, cfg 3-4, Adaptive Projected Guidance + FreSca (both vanilla nodes)
An A/B-tested negative prompt (artifacts, gpt-image, washed-out colors, AI slop, missing fingers, etc.) noticeably improves quality even without the LoRA whenever cfg > 1

Hands remain unreliable and 3D renders/food photos occasionally show "gpt-image splotchiness"; a v2 is in training. The author argues Qwen 2.1 has an excellent base — seemingly good pretraining ruined by a bad finetune — and with the fix it beats other local models for editing, with detail quality surpassing even Krea 2.

Original post →

Community trains fix LoRA that salvages Qwen Image 2.1's generation quality

More from Multimodal