Zen Image Edit T8

Zen Image Edit for ComfyUI / Zen Image Edit ComfyUI 节点

English · 简体中文

Model / 模型下载: t8star/Zen-Image-Edit-Comfy

English

Zen Image Edit brings AiArtLab/zen-image-edit to ComfyUI as a single-file checkpoint. Its Qwen3.5-0.8B multimodal text encoder and fusion adapter produce a native CLIP output; image generation and editing continue through ComfyUI's Qwen-Image-2.1 conditioning, sampler, and VAE nodes. No Diffusers pipeline is required at runtime.

Viggle Turbo v0.2.1 — native 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.

Required loader for these workflows: 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.

Zen installation

1. Update ComfyUI to a build with native Qwen-Image-2.1 support (validated on ComfyUI 0.37.0). Install Zen Image Edit T8 from ComfyUI Manager, or clone this repository into ComfyUI/custom_nodes/.
2. If installing manually, run python -m pip install -r ComfyUI/custom_nodes/comfyui-Zen-Image-Edit-T8/requirements.txt with ComfyUI's Python interpreter.
3. Download zen_image_edit_qwen21_single.safetensors from t8star/Zen-Image-Edit-Comfy and place it in ComfyUI/models/checkpoints/. Restart ComfyUI.
4. Open the text-to-image or image-edit workflow. For editing, select an image in LoadImage; the image-edit API example expects an uploaded reference.png. Text-to-image API example.

The checkpoint contains the Qwen-Image-2.1 DiT and VAE plus the Zen student encoder, fusion adapter, and tokenizer/processor assets. You do not need to download a separate text encoder. The checkpoint loader outputs MODEL, CLIP, and VAE; the CLIP-only loader works with separate native diffusion and VAE loaders. The advanced encoder supports an explicit output canvas. Text encoding uses BF16; the CUDA/PyTorch runtime must support it.

For editing, connect the first image as the edit target and any following images as references. The stock TextEncodeQwenImage21 node remains supported. Each reference must contain at least 65,536 pixels after native resizing; the adapter supports up to 2,304 processor tokens. Use plain prompt text: text inversion, prompt weights, CLIP skip, and text-encoder LoRA are unsupported. English prompts are recommended by the upstream model card. The validated GPU was an RTX 5090 Laptop with 24 GB VRAM; memory use rises with resolution and reference count.

The node code is Apache-2.0. The checkpoint combines weights from AiArtLab and Comfy-Org/Qwen-Image-2.1. Qwen-derived weights are limited to non-commercial research/evaluation under the Qwen Research License; commercial use requires separate permission. See the model license and NOTICE.md.

简体中文

本节点将 AiArtLab/zen-image-edit 接入 ComfyUI。单文件模型内含 Qwen-Image-2.1 DiT、VAE、Qwen3.5-0.8B 多模态文本编码器及融合适配器。节点输出原生 CLIP，可继续使用 ComfyUI 的 Qwen-Image-2.1 条件编码、采样与 VAE 节点；运行时不依赖 Diffusers 管道。

Viggle Turbo v0.2.1：原生 LoRA

下载 LoRA · 文生图工作流 · 图片编辑工作流

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。

工作流指定加载器： ComfyUI 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。

Zen 安装

1. 更新到原生支持 Qwen-Image-2.1 的 ComfyUI（已在 0.37.0 验证）。在 ComfyUI Manager 安装 Zen Image Edit T8，或将本仓库克隆至 ComfyUI/custom_nodes/。
2. 手动安装时，使用 ComfyUI 对应的 Python 执行 python -m pip install -r ComfyUI/custom_nodes/comfyui-Zen-Image-Edit-T8/requirements.txt。
3. 从 t8star/Zen-Image-Edit-Comfy 模型仓库下载 zen_image_edit_qwen21_single.safetensors，放入 ComfyUI/models/checkpoints/，然后重启 ComfyUI。
4. 打开文生图或图片编辑工作流。编辑时先在 LoadImage 选择图片；图片编辑 API 示例需要预先上传名为 reference.png 的图片。另附文生图 API 示例。

无需另外下载文本编码器。整合加载器输出 MODEL、CLIP、VAE；独立 CLIP 加载器可搭配原生 DiT/VAE 加载器使用；高级编码器可指定输出画布。文本编码使用 BF16，CUDA/PyTorch 运行环境须支持该精度。编辑时，第一张图是编辑目标，后续图片是参考图。兼容原生 TextEncodeQwenImage21。每张参考图经原生缩放后须至少有 65,536 像素，融合适配器上限为 2,304 个处理器 token。请使用普通提示词；暂不支持文本反演、提示词权重、CLIP skip 或文本编码器 LoRA。原模型建议使用英文提示词。验收显卡为 24 GB 显存的 RTX 5090 Laptop；分辨率和参考图数量越高，显存需求越大。

节点代码采用 Apache-2.0。模型权重分别来自 AiArtLab 与 Comfy-Org/Qwen-Image-2.1。Qwen 衍生权重依照 Qwen Research License 仅限非商业研究/评估；商业使用须另行取得许可。详见模型许可证和 NOTICE.md。

T8star

Bilibili · YouTube · Hugging Face · API · Free gallery / 免费画廊 · Online AI apps / 在线 AI 应用 · ComfyUI portable bundle / ComfyUI 整合包