t8star/Qwen-Image-2.1-viggle-turbo-4step-r64-comfy

Qwen-Image-2.1 Viggle Turbo · ComfyUI LoRA

Built with Qwen. Viggle 训练，T8star 转换。当前提供 v0.2.1 六步 r256 / r128；仓库名称中的 4step-r64 是旧版名称，旧权重保留供复现。

v0.2.1 必须按随附工作流使用 ComfyUI 内置 `LoraLoaderBypassModelOnly`，搜索名称：`Load LoRA (Bypass, Model Only) (for debugging)`。强度设为 `1.0`，搭配 BF16 底座；无需安装新节点。 普通 LoraLoader / LoraLoaderModelOnly 会合并更新，存在作者指出的舍入精度损失。找不到该内置节点时，请更新 ComfyUI。

For the supplied v0.2.1 workflow, use ComfyUI's built-in `LoraLoaderBypassModelOnly`: search for `Load LoRA (Bypass, Model Only) (for debugging)`. Set strength to `1.0` and use the BF16 base. No custom nodes are required. Ordinary merged LoRA loaders can lose adapter updates through rounding. Update ComfyUI if the built-in node is unavailable.

简体中文

1. 1.使用 ComfyUI 0.37.0、前端 1.53.6 或更新版本。选择一份 v0.2.1 LoRA，放入 ComfyUI/models/loras/。
2. 2.从 Comfy-Org/Qwen-Image-2.1 下载三个底座文件，按下表放置。
3. 3.导入文生图工作流或图片编辑工作流，选择模型；编辑时在 LoadImage 选择参考图。默认选用 r256，使用 r128 时切换 LoRA 文件即可。

工作流使用 Euler、六步、无 CFG，按输出尺寸自动计算作者的 sigma 日程，并关闭 Qwen 前缀缓存。请保留调度组连线；仅把普通 KSampler 的步数改成 6 并不等价。另附文生图 API和编辑 API。

支持 BF16 DiT。INT8 DiT 的融合 MLP 路径可能跳过 bypass hook；上表中的 INT8 文本编码器可以使用。使用原版 Qwen3-VL-8B，不使用 Zen student；不要叠加不同版本的蒸馏 LoRA，也不要挂在 Viggle 完整微调 Transformer 上。

转换保留全部 227 组源投影，通过块对角打包适配 ComfyUI 的融合 MLP；文件只含 LoRA，不含底座，没有再次截断 rank。已在禁用全部自定义节点的环境验证 512 像素文生图与编辑、完整权重映射和六步调度。内置 bypass 加载器仍标为实验性功能；复杂多参考图编辑并非均能准确遵循指令。

English

These are v0.2.1 six-step r256 / r128 adapters converted for native ComfyUI. The repository retains its historical 4step-r64 name; the old r64 file and legacy instructions remain available.

Use ComfyUI 0.37.0 / frontend 1.53.6 or newer. Place one v0.2.1 adapter in ComfyUI/models/loras/, download the separate base files listed above, and open the text-to-image or image-edit workflow. Select your model files and, for editing, a reference image. Switch the LoRA filename to use r128.

Load through `LoraLoaderBypassModelOnly`, strength `1.0`, BF16 DiT. The workflows use Euler, six steps, no CFG, the author's resolution-dependent sigma schedule, and disabled prefix caching. Keep the schedule group connected: setting an ordinary KSampler to six steps is not equivalent. INT8 DiT's fused MLP can skip bypass hooks; the separate INT8 Qwen3-VL-8B text encoder is supported. Do not use the Zen student encoder or stack distilled adapters/full fine-tunes.

All 227 source projection adapters are preserved through block-diagonal packing for ComfyUI's fused MLP. No base weights, further rank truncation, or weight merging are included. Native 512 px generation/editing, complete mapping, and schedule equivalence were validated with all custom nodes disabled. The built-in bypass loader is experimental; this is functional validation, not a comprehensive quality benchmark.

Sources & license / 来源与许可

Upstream model / 原模型 · Conversion source / 转换源码 · SHA256SUMS

Weights retain the Qwen Research License, for non-commercial research/evaluation; commercial use requires separate permission. / 权重沿用 Qwen Research License，仅限非商业研究与评估；商业使用须另行取得许可。v0.2.1 归属和修改见 NOTICE-v0.2.1 与 上游声明；旧版见 NOTICE。

T8star

B站 / Bilibili · YouTube · Hugging Face · API · 免费画廊 / Gallery · 在线 AI 应用 / AI apps · ComfyUI 整合包 / Portable bundle