Qwen Image Edit Plus (Multi-Image)

This is the sibling node that does the thing everyone actually wants: take a pose from one image and force it onto another subject, without ever touching OpenPose or a ControlNet. The multi-image variant of Qwen-Image-Edit accepts one to five images, treats the first as the pose reference, and edits the rest to adopt it. That's a workflow that used to mean keypoint extraction, a ControlNet checkpoint, and a prayer. Here it's one node and a sentence.

The qwen/qwen-image-edit-plus model behind it is the reason the community stopped reaching for the adapter stack entirely - posing, camera angles and composition-from-reference all moved onto the edit model's context window instead of a conditioning map. You're getting the hosted version, which means no 20B checkpoint to fit in VRAM. As with its single-image sibling in the same pack, this is a convenience tax on a model you could technically run locally; the tax buys you a laptop-compatible, zero-download setup.

How it works

Mechanically it's the same shape as the rest of the pack: each image in the batch is converted to a base64 PNG data URL, all of them are passed to Replicate's qwen-image-edit-plus API as the image parameter with your prompt, and the single result is downloaded and turned back into a tensor. The node enforces the 1–5 image limit in code, so feeding it a six-image batch fails before it ever reaches the API.

The inputs that matter

• images - the IMAGE batch. Wire in one or more loaded images. Order matters: the first one is the pose reference, everything after it is what gets edited.
• prompt - describe the edit, usually in terms of the relationship: "the person in the second image adopts the pose from the first image" is the README's own example and a good template.
• api_token - your Replicate token, or leave blank if REPLICATE_API_TOKEN is set in the environment.

The optional controls match the single-image node: seed (-1 for random), go_fast (true by default - a faster, slightly rougher pass), aspect_ratio (defaults to match_input_image), output_format, output_quality, and disable_safety_checker. For a pose-transfer job you'll mostly leave these alone; the one worth revisiting is go_fast if a result comes back with mangled anatomy.

Outputs

Same two outputs as every node in this pack: image - the edited result as a standard IMAGE tensor - and output_url, the direct Replicate URL as a STRING. The image output is a single result regardless of how many images you fed in; the rest were context, not outputs.

Installing it

This node is part of the comfyUI_charly_replicate_api pack (four nodes total). ComfyUI Manager, search "ComfyUI Replicate Image Generation & Edit", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/charlyad142/comfyUI_charly_replicate_api

Dependencies are just the replicate client plus requests and Pillow, and there's nothing to download model-wise - the work happens on Replicate's servers. You need a token from replicate.com/account/api-tokens, and multi-image edits cost real money per call, so budget accordingly.

Where people get burned

The red-image trap applies here exactly as it does to the pack's other nodes: a failure returns a solid red 512×512 tensor with the error message tucked into the output_url string. Check that string before you re-architect your workflow.

Two model-level truths worth remembering. First, Qwen-Image-Edit re-emits the whole frame, so pose transfer can drag other pixels along with it - identity drift on faces is the well-documented failure mode, worst on faces the model hasn't seen, and the standard fix is an inpainting pass afterward. Second, the pose reference's body type tends to bleed into the result. If the transferred pose looks wrong in the shoulders or the proportions, the model is being literal about the source, not broken. Both are properties of the model on Replicate's side; this node is just the delivery vehicle.