Qwen-Image-2.1-TailSteer

TailSteer is the final training stage for Qwen-Image-2.1. The authors state that it largely preserves the base model's generation and editing capabilities while reducing the accumulation of texture drift when using the previous round's output as the input for the next round. In the Add & Preserve evaluation at round 10, PSNR outside the edited region is higher by 1.22 dB, and LPIPS is lower by 11.5%. The repository contains full weights, with an architecture and loading method consistent with the official version; you only need to replace the model ID. ComfyUI usage is also provided, and both the evaluation set and training dataset are publicly available.

TailSteer is the final training stage for Qwen-Image-2.1. The authors state that it largely preserves the base model's generation and editing capabilities while reducing the accumulation of texture drift when using the previous round's output as the input for the next round. In the Add & Preserve evaluation at round 10, PSNR outside the edited region is higher by 1.22 dB, and LPIPS is lower by 11.5%. The repository contains full weights, with an architecture and loading method consistent with the official version; you only need to replace the model ID. ComfyUI usage is also provided, and both the evaluation set and training dataset are publicly available.

Why it matters In multi-turn editing, using each output as the input for the next step can cause details to become increasingly blurry. This version specifically reduces this cumulative drift.

Hugging Face likes 40 · Hugging Face trending score 39 · Hugging Face downloads 141

License: Research only · ComfyUI: Mentioned in the model card · VRAM: ≈18 GB (estimated)

Reasons

  • Ranked 3 on the Hugging Face any-to-any trending list (reflects attention only)
  • The model card includes comparison GIFs and all prompts for continuous editing over 20 rounds, contrasting the base model with TailSteer

Usage

  • In diffusers, use QwenImage21Pipeline.from_pretrained("Lyricccco/Qwen-Image-2.1-TailSteer"), feeding each round's output as the input for the next; for ComfyUI, replace the base model file as specified in the model card

Risks

  • Uses the Qwen research license, so commercial use is not permitted
  • Improvement metrics were measured by the authors on their custom evaluation set; with approximately 7B parameters, VRAM requirements are comparable to the base model

Open

抱脸 any-to-any 上升榜第 3

Explore the latest AI Radar updates