CoolFace
Modelpublic

xingewh/Qwen-Image-2.1-viggle-turbo-v0.2-5step-lora-r256_comfy

sourceHugging Facemitupdated 3d agoView on Hugging Face
7likes
Model Card

Qwen-Image-2.1-viggle-turbo-v0.2-5step-lora-r256_comfy

<Gallery />

Model description

language:

  • —en
  • —zh base_model: Qwen&#x2F;Qwen-Image-2.1 tags:
  • —comfyui
  • —lora
  • —qwen-image
  • —text-to-image
  • —diffusers-to-comfyui library_name: peft ---

Qwen-Image-2.1-viggle-turbo-v0.2-5step-lora-r256-comfy

What is this &#x2F; 这是什么

This is a ComfyUI-compatible format conversion of the original LoRA from Viggle&#x2F;Qwen-Image-2.1-viggle-turbo.

The original LoRA was trained under the Diffusers&#x2F;PEFT framework, with weight keys prefixed as &#x60;transformer.&#x60;. ComfyUI&#39;s standard LoRA loader expects keys prefixed as &#x60;diffusion_model.transformer.&#x60;, so loading the original directly will silently fail (it loads but has no effect).

This repository provides a converted file where all 454 weight keys have been re-prefixed from &#x60;transformer.&#x60; to &#x60;diffusion_model.transformer.&#x60;, making it directly loadable in ComfyUI via the &#x60;Load LoRA&#x60; node.

&gt; The conversion script was written with the help of DeepSeek.

这是对 Viggle&#x2F;Qwen-Image-2.1-viggle-turbo 原版 LoRA 的 ComfyUI 兼容格式转换版。

原版 LoRA 是在 Diffusers&#x2F;PEFT 框架下训练的，权重键名以 &#x60;transformer.&#x60; 开头。ComfyUI 的标准 LoRA 加载器期望的键名格式是 &#x60;diffusion_model.transformer.&#x60;，直接加载原版会静默失效（加载了但完全不生效）。

本仓库提供的文件，已将全部 454 个权重的键名前缀从 &#x60;transformer.&#x60; 批量替换为 &#x60;diffusion_model.transformer.&#x60;，使其可以在 ComfyUI 中直接通过 &#x60;Load LoRA&#x60; 节点加载使用。

&gt; 本转换脚本在 DeepSeek 的帮助下编写完成。

What was changed &#x2F; 转换了什么

  • —Only the key prefix was modified &#x2F; 仅修改键名前缀：&#x60;transformer.xxx&#x60; → &#x60;diffusion_model.transformer.xxx&#x60;
  • —No weight values were altered &#x2F; 未修改任何权重数值：the LoRA&#39;s actual effect is identical to the original &#x2F; LoRA 的实际效果与原版完全一致
  • —Conversion script &#x2F; 转换脚本：see &#x60;convertlora.py&#x60; in this repository &#x2F; 见仓库中的 &#x60;convertlora.py&#x60;

How to use (ComfyUI) &#x2F; 如何使用（ComfyUI）

  1. 1.Place &#x60;Qwen-Image-2.1-viggle-turbo-v0.2-5step-lora-r256comfy.safetensors&#x60; into your &#x60;ComfyUI&#x2F;models&#x2F;loras&#x2F;&#x60; folder. 将 &#x60;Qwen-Image-2.1-viggle-turbo-v0.2-5step-lora-r256comfy.safetensors&#x60; 放入 &#x60;ComfyUI&#x2F;models&#x2F;loras&#x2F;&#x60; 文件夹。
  1. 1.Add a &#x60;Load LoRA&#x60; node in your workflow and load the file. 在工作流中添加 &#x60;Load LoRA&#x60; 节点，加载该文件。
  1. 1.Connect the node to the Qwen-Image-2.1 main model. 将该节点连接到 Qwen-Image-2.1 主模型。
  1. 1.This is a 5-step acceleration LoRA. Set sampling steps to 5–10, and try LoRA weight &#x60;0.8–1.0&#x60; first, then fine-tune as needed. 这是一个 5 步加速 LoRA，建议将采样步数设为 5~10 步，LoRA 权重可先尝试 &#x60;0.8~1.0&#x60;，再根据效果微调。

Suggested parameters &#x2F; 参数建议

Parameter &#x2F; 参数Suggested &#x2F; 建议值Notes &#x2F; 说明
Steps &#x2F; 步数5–10Original LoRA is 5-step accelerated; increase for quality &#x2F; 原 LoRA 为 5 步加速，可适当增加步数换取质量
CFG1.0Official Qwen-Image-2.1 recommended default &#x2F; Qwen-Image-2.1 官方推荐默认值
LoRA Weight &#x2F; LoRA 权重0.8–1.0Adjust based on results &#x2F; 根据生成效果调整
Sampler &#x2F; 采样器euler &#x2F; simpleRecommended combination &#x2F; 推荐组合

lora convert.zip 为转换工具。 lora convert.zip is the conversion toolkit. ---

Disclaimer &#x2F; 免责声明

  • —This repository only performs format conversion on the original LoRA; no weights are modified, and the effect is identical to the original. 本仓库仅对原版 LoRA 做格式转换，不修改权重，效果与原版一致。
  • —The original model copyright belongs to Viggle. Please follow the original repository&#39;s license. 原版模型版权归 Viggle 所有，请遵循原仓库的许可证。
  • —The conversion script is for learning and exchange purposes only. 转换脚本仅供学习交流使用。

Credits &#x2F; 致谢

Download model

Download them in the Files & versions tab.