Qwen Image 3 Edit

The Qwen-Image-Edit line is where a huge slice of image work stopped needing a mask, a ControlNet, or an adapter: you hand it a picture and a sentence and the change just happens. Swap a jacket, remove a stranger from a photo, re-light a scene. This node is the hosted version of that workflow - Qwen-Image 3.0's edit model, reached through Comfy's partner proxy, handling up to three reference images at once. It's the "convince an image to change" node, and it's built into ComfyUI core.

What it is. A partner node that sends one to three reference images plus a text instruction to the Qwen-Image 3.0 edit API. No local model, no VRAM cost; billed per call against your Comfy account credits. The base qwen-image-3.0 and the pricier qwen-image-3.0-pro are the two options.

How it works. Plug images into the model combo's image slots - an autogrow field accepting one to three references, numbered image_1, image_2, image_3. Refer to them in the prompt as @Image1, @Image2, @Image3 (a bare @image means the first one). You can also just describe things plainly and let the model figure out which image you mean. The size control has three modes: "match input" reuses the first reference image's resolution, "auto" keeps the same aspect ratio but lets the model choose, and "custom" exposes explicit width/height fields.

The inputs that matter. The image slots plus the prompt are the whole game - that's where the edit actually happens. n gives you up to 6 variants, seed for reproducibility, prompt_extend (default ON) lets the model rewrite your instruction, and watermark optionally stamps the result. The negative_prompt inside the model combo works too, for "keep X out of this."

The outputs. One IMAGE - a batch of n edited results. Unlike some other partner nodes there's no final_prompt output here, so if prompt_extend's rewriting annoys you, just switch it off and keep your instructions literal.

How you get it. It's core - no install, no Manager, no git clone. You do need a Comfy account logged in with credits, and the images you upload travel to the Qwen API, so don't feed it anything you'd rather keep private.

Where people get burned. The hard limit is three reference images - a batched input counts once, and the node raises an error if you exceed it. The subtler one comes straight from the open Qwen-Edit's documented weakness, and it holds here: these models re-emit the whole frame, so pixels you never asked to touch can drift, and faces wander on a chain of edits. For a surgical change - "just remove the cup on the table" - a mask-based workflow is still the more reliable tool; reach for this node when you want the flexibility of a natural-language edit. And remember prompt_extend ON means the model may embellish your instruction; if a result goes off-brief, check that toggle before blaming the model. These partner nodes are a mid-2026 addition, so keep ComfyUI updated.