This LoRA model is derived from https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo. I extracted it into a LoRA file, so you can adjust the weight at any time for easy mixing with other LoRA models.

Brief Review: You need to lower the strength to 0.3–0.8 (0.6 recommended). For some images, the improvement is subtle, but overall it can suppress the probability of the model outputting "bad anatomy" and has image repair capabilities. Most importantly, it adds details when generating with 8 steps. The sampling configuration used in the showcase is er_sde + beta.

Drawbacks: This model may cause severe moiré patterns when used for editing. However, for editing, it follows instructions better than other turbo models, at the cost of (a bit) lower quality. You can try lowering the weight to find a balance point.

这个lora模型来源于 https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo

我将其提取为 lora 文件，这样你就可以随时调整权重，以方便和其他 lora 模型混合使用

简评：需要降低强度到 0.3-0.8 (推荐0.6)，对于某些图片改进很细微，但是整体上可以抑制模型输出 "bad anatomy" 的概率，有修复图片的能力。最重要的是 8steps 生成时会增加细节。

showcase使用的采样配置是 er_sde+beta 。

缺陷：这模型用于编辑的时候可能带来严重的摩尔纹效果。但是，编辑时使用会比其他的加速模型更遵循指令一点，代价就是质量变差。可以尝试降低权重可能可以找到平衡点。

[backup huggingface]