LongCat Image Edit

Edit by instruction. You load a photo, type "make it golden hour but keep the composition," and the image comes back changed - no mask painting, no inpainting precision, no ADetailer acrobatics. That's the pitch of LongCat-Image-Edit, and this node is the ComfyUI front door to it. For single-image edits it's refreshingly close to what the demos promise.

How it works

LongCat-Image-Edit is a separate checkpoint from the base model, not a mode you switch on. The key is its Qwen2.5-VL encoder - a vision-language model - so the pipeline can actually see the structure of your source image while your instruction steers the edit. That's the architectural trick that makes instruction editing work instead of just re-imagining the image from text. The LongCat Model Loader detects the model by the "edit" in its folder name and hands back the right pipeline automatically.

This node takes the first image in your batch, converts it to RGB, runs the edit pipeline, and returns a single IMAGE. There are no width, height, or batch controls - it edits at whatever resolution your source image is, which is a feature. Feed it a big photo and you get a big edit.

The inputs that matter

• image - the one required input and the thing you're editing. Wire in your loaded image.
• prompt - the instruction, in natural language: "turn the background into a rainy night" or "把背景换成巴黎街头". And yes, the same quoted-text rule as the T2I node applies: any text you want rendered into the edited image goes in quotes ("COFFEE").
• steps - 50 default, but the community consensus is you don't need it. People report genuinely good edits at 15–20 steps, some even down around 1. If a full 50-step edit feels like a waste of your GPU, try 20 first.
• guidance_scale (default 4.5) - same guidance-distilled flow-model logic as the T2I node: if output gets harsh, lower it, don't raise it. Some users run edits happily at cfg ~1.
• negative_prompt - optional at best on this architecture. Leave it empty unless you're fighting something specific.
• seed - set it to reproduce an edit you liked.

One output: IMAGE.

Reality check: what it's good at and what it isn't

Content edits work well: swap a subject, transplant something from one image into another, restyle the lighting, change the setting. That's the sweet spot. What people hit a wall on is camera control - one Reddit user summed it up: asking for a side-view camera angle often just rotates the subject instead of the camera. Don't expect cinematography; expect content.

One more limitation worth knowing: the underlying model can actually take multiple reference images, but this wrapper only feeds it one. It's a real constraint of this young pack, not something you can fix in settings. If multi-image editing matters to you, you'll be waiting for a richer wrapper.

The pair gotcha

Wire the base model into this node and it stops you with "Use LongCat Text to Image node for T2I models." Same rule as its sibling, mirrored: the loader picks the pipeline from the model folder name, and this node only accepts an edit model. Wrong-model pairing is the #1 wiring mistake in this pack, and the error message is the tell.

Troubleshooting

• It edited but didn't change what you asked → the model's instruction-following is decent but not magic; rephrase the instruction as a concrete visual change ("add a red umbrella") rather than a vague intent ("make it more interesting").
• Rendered text is garbled → the quotes again. Wrap any in-image text in "…".
• OOM → same fix as everywhere in this pack: keep CPU offload enabled at the loader.

The wrapper is young and barely has stars to its name, but for single-image instruction editing, LongCat-Image-Edit is one of the friendlier open options out there - and this node gets you to it without learning a Python API.