Zen Image Edit — ComfyUI single checkpoint / ComfyUI 单文件模型
Built with Qwen · Improved using Qwen
New / 新增: Viggle v0.2.1 native LoRAs / 原生 LoRA — no custom nodes required / 无需自定义节点。
This is a ComfyUI-format conversion of AiArtLab/zen-image-edit, using the native Qwen-Image-2.1 DiT and VAE from Comfy-Org/Qwen-Image-2.1. The file contains the Qwen3.5-0.8B student text encoder, Zen fusion adapter, and tokenizer/processor assets. It is a conversion for ComfyUI, not a new fine-tune or an official Qwen release.
这是 AiArtLab/zen-image-edit 的 ComfyUI 单文件转换版，整合来自 Comfy-Org/Qwen-Image-2.1 的原生 DiT/VAE，以及 Qwen3.5-0.8B 学生文本编码器、Zen 融合适配器与分词/图像处理资源。本仓库没有重新训练模型，也不是 Qwen 官方发布。
| File / 文件 | Size / 大小 | Placement / 位置 |
|---|---|---|
zen_image_edit_qwen21_single.safetensors |
16.995 GB | ComfyUI/models/checkpoints/ |
Install the Zen Image Edit T8 nodes through ComfyUI Manager and open the included text-to-image or image-edit workflow. Only this checkpoint is required. Native Qwen-Image-2.1 conditioning, KSampler, and VAE decode remain available. / 通过 ComfyUI Manager 安装 Zen Image Edit T8 节点，打开随附文生图或图片编辑工作流。只需下载本文件；可继续使用原生 Qwen-Image-2.1 条件编码、KSampler 和 VAE 解码。
License / 许可: Qwen-derived weights are under the enclosed Qwen Research License Agreement, which grants non-commercial research/evaluation rights; commercial use requires a separate license from the rightsholder. The student encoder retains its Apache-2.0 terms in LICENSE-Qwen3.5-0.8B. Preserve the agreements and NOTICE.md on redistribution. / Qwen 衍生权重受随附 Qwen Research License 约束，仅授权非商业研究/评估；商业使用须另向权利人取得许可。学生编码器的 Apache-2.0 许可见 LICENSE-Qwen3.5-0.8B。再次分发时须保留许可证与 NOTICE.md。
Viggle Turbo v0.2.1 · Native ComfyUI LoRA
Download LoRAs · Text-to-image workflow · Image-edit workflow
These converted Viggle adapters work with ComfyUI's built-in nodes. The Zen extension is not required. Use ComfyUI 0.37.0 with frontend 1.53.6 or newer.
| File | Size | Folder |
|---|---|---|
qwen_image_2.1_viggle_turbo_v0.2.1_r256_comfy.safetensors |
1.76 GB | models/loras/ |
qwen_image_2.1_viggle_turbo_v0.2.1_r128_comfy.safetensors |
881 MB | models/loras/ |
Choose one adapter. Download the separate base files from Comfy-Org/Qwen-Image-2.1: qwen_image_2.1_bf16.safetensors → models/diffusion_models/, qwen3vl_8b_int8_convrot.safetensors → models/text_encoders/, and qwen_image_2.1_vae_bf16.safetensors → models/vae/. Open a workflow and select the files; for editing, select your reference image in LoadImage.
Use the built-in Load LoRA (Bypass, Model Only) (for debugging) (LoraLoaderBypassModelOnly) at strength 1.0. It computes the adapter separately, preserving updates that ordinary merged LoRA loading can round away. This conversion retains all 227 source projection adapters, including both halves of ComfyUI's fused MLP; it contains no base weights. The larger file size comes from lossless block-diagonal packing.
The workflows use Euler, six steps, no CFG, automatically shift the author's sigma schedule to the actual output size, and disable Qwen prefix caching. Keep the schedule group connected. The supported base is BF16 DiT; INT8 DiT has a fused MLP path that can skip bypass hooks. Use the original Qwen3-VL-8B encoder, not the Zen student encoder. The built-in bypass loader is experimental. API examples are also included.
Validated with all custom nodes disabled: 512 px text-to-image and editing, complete native adapter mapping, and schedule equivalence. This is functional validation, not a full image-quality benchmark. Weights retain the upstream Qwen Research License; see the conversion notice.
简体中文
转换后的 Viggle LoRA 直接使用 ComfyUI 内置节点，无需安装 Zen 插件。请使用 ComfyUI 0.37.0、前端 1.53.6 或更新版本。
| 文件 | 大小 | 存放目录 |
|---|---|---|
qwen_image_2.1_viggle_turbo_v0.2.1_r256_comfy.safetensors |
1.76 GB | models/loras/ |
qwen_image_2.1_viggle_turbo_v0.2.1_r128_comfy.safetensors |
881 MB | models/loras/ |
两种 LoRA 任选其一。另从 Comfy-Org/Qwen-Image-2.1 下载底座：qwen_image_2.1_bf16.safetensors 放入 models/diffusion_models/，qwen3vl_8b_int8_convrot.safetensors 放入 models/text_encoders/，qwen_image_2.1_vae_bf16.safetensors 放入 models/vae/。导入工作流并选择文件；编辑时在 LoadImage 选择参考图。
使用内置 **Load LoRA (Bypass, Model Only) (for debugging)**（LoraLoaderBypassModelOnly），强度 1.0。它独立计算 LoRA 分支，保留普通合并加载可能因舍入而丢失的更新。转换完整保留原始 227 组投影适配器，包括 ComfyUI 融合 MLP 的两个分支；文件不含底座权重。体积增加来自无损块对角打包。
工作流采用 Euler、六步、无 CFG，按实际输出尺寸自动调整作者的 sigma 日程，并关闭 Qwen 前缀缓存。保留调度组连线。支持的底座为 BF16 DiT；INT8 DiT 的融合 MLP 路径可能跳过 bypass hook。文本编码器使用原版 Qwen3-VL-8B，不使用 Zen student。内置 bypass 加载器目前标为实验性功能。另附 API 示例。
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、全部适配器映射和调度等价性；这属于功能验证，不代表完整画质基准。权重沿用上游 Qwen Research License，详见转换声明。