Community LoRA fixes Qwen Image 2.1 edit drift: median offset drops from 24.3px to 1.6px
huggingface · x · 2026-09-29
HF user ausboss released the Qwen-Image-2.1-Consistency-LoRA, fixing Qwen Image 2.1's drift and unwanted repainting during image edits.
Problem: edits and restyles (watercolor/comic/anime) come back slightly taller or nudged sideways, varying by seed — a waistband 40px lower, a misaligned face — and local edits repaint areas you didn't ask about (hair, signs, texture).
Solution: the LoRA pins the edit to the original frame — same prompt, same edit, nothing moves. No trigger word; load it in ComfyUI on Comfy-Org INT8 weights (rank 32, all 384 tensors).
Results (36 held-out restyles):
- No LoRA: worst-corner offset median 24.3px, worst case 61.1px
- Step 1500: median 1.6px, worst 12.9px
- Step 2000: tightest alignment (0.9px median, 3.5px worst) but paintings come out slightly paler
Usage: start with step 1500 to keep Qwen's look; step 2000 for tightest alignment. Limits: comic/anime restyles may still nudge ≤12.9px; English-trained.
More from Multimodal
- Agentic filmmaking experiment: GPT-6 Astra edits, Grok Imagine renders, Suno v6 scores — TinfoilTricorn · 2026-09-29
- Fizgig 6.6 adds edit-LoRA training for Qwen Image 2.1, learns grading from 40 pairs — shootthesound · 2026-09-29
- Deforum keeps going and going: an endless morphing chair animation — makeitrad1 · 2026-09-29
- GPT Image 2.5 vs open-source Krea 2 Turbo: same-prompt test highlights OpenAI's heavy censorship — Due_Research9042 · 2026-09-29
- Hugging Face release: Yue2-based quiet-storm model for 90s Jazz-Soul vibes — -becausereasons- · 2026-09-29
- One prompt, six shots: multi-shot video generation shows cinematic chops — minchoi · 2026-09-29