Qwen-Image-Edit ComfyUI Workflows

Where the base model generates, Qwen-Image-Edit comfyui workflows change an image you already have. Qwen-Image-Edit is Alibaba Tongyi Lab's open editing model from the Qwen family, built on the 20B Qwen-Image transformer and released under Apache 2.0, so the weights are open and you can run the edits on your own hardware instead of uploading to a closed service. The workflows on this page are the ready-made graphs for that: load your image, describe the change in plain language, and the model applies it, with the node graph already wired.

The editing is instruction-driven, so you tell it what to change rather than masking by hand: swap an object, alter a background, adjust wording on a sign, or restyle a scene. The current build, Qwen-Image-Edit-2511, took over from the 2509 release and improves character consistency, reduces image drift across edits, and accepts up to three reference images at once for jobs like fusing two people into one group shot. Beyond straight edits, the workflows here cover relighting, which re-lights a subject as if shot under different conditions, and style transfer, which carries the look of one image onto another.

Because the graph is open, you control the strength of the edit, the conditioning, and any LoRA, and you can re-run to dial in how much changes versus how much of the original is preserved. The 2511 build also folds several popular community LoRAs straight into the base model, and you can still stack your own on top when you want a specific learned style across edits. That control is the difference between a one-shot filter and a workflow you can tune until the edit looks deliberate rather than pasted on.

Everything here is open and editable. Open a Qwen-Image-Edit workflow on Comfy Cloud with nothing to install, or download it and run it locally if you have a capable GPU. An FP8 build edits in about 16GB of VRAM and a quantized GGUF version goes lower, while the full BF16 weights want roughly 40GB. You can chain edits, add a LoRA, adjust the parameters, and there is no watermark on the result.