Eric Qwen-Edit Inpaint Transfer

The most "oh, that's clever" node in the pack. Eric Qwen-Edit Inpaint Transfer moves content from a reference image into a masked region of another image - you want that jacket from photo B on the person in photo A, or that car dropped into a different street - and it's built to make the result look placed, not pasted.

A naive version of this is just cut-and-paste with a soft edge, which always looks composited because the lighting, color, and perspective never match. This node's trick is to let the model fix that: it composites the transfer content in first, then runs Qwen-Edit over the region with a prompt telling it to harmonize, then composites again with a feathered mask to protect the original outside the mask.

The strategy (from the README)

1. Scale the transfer image (and optional transfer mask) proportionally so the source region fits inside the target mask's bounding box.
2. Pre-composite the transfer into the masked area - the model sees the content already in place, not a hole.
3. The model harmonizes lighting, color, and edges via the prompt.
4. Post-composite with a feathered mask to keep the original pixels outside the mask untouched.

That pre-composite step is the clever bit. Because Qwen-Edit sees content already sitting in the region, it doesn't have to hallucinate what belongs there - it just has to make it blend.

Inputs that matter

• image - the target. mask - white = where the transfer content goes, black = keep original.
• transfer_image - the reference whose content you're moving in.
• transfer_mask - optional, and underrated: mark which part of the transfer image to use (white = keep). When omitted, the full transfer image is used. Provide it when the reference has background you don't want to drag along.
• prompt - optional guidance. Empty = a default prompt that tells the model to harmonize the inserted content. Add detail to steer the blend ("match the warm window light").
• transfer_vl_ref - when on, the full un-cropped transfer image also goes in as a second VL semantic reference, giving the model extra context. Disable if the pre-composite alone is enough.
• blend_strength - 1.0 fully replaces the masked area with generated content; lower mixes in some original for subtle blends.
• feather - post-composite edge softness (0–50, default 8).
• steps, true_cfg_scale, seed, max_mp - the standard edit knobs.

Output: one image.

Wiring it

Loader → Apply LoRA → Inpaint Transfer(image, mask, transfer_image) → Preview/Save

Installing it

ComfyUI Manager → "Eric Qwen-Edit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments.git

Restart, download Qwen/Qwen-Image-Edit-2511 (~54 GB), load via the pack's loader. Dependencies pinned by the pack's requirements: diffusers>=0.34,<0.38, transformers>=4.50,<5.

Troubleshooting

• The transfer looks pasted, not lit - give the model context in the prompt ("harmonise the pasted element with its surroundings" is the pack's own example), and make sure transfer_vl_ref is on so the model sees the source.
• Wrong part of the transfer got moved - add a transfer_mask to select exactly the region you want; without it, the whole reference comes along.
• Region too small for the content - the node scales the transfer to fit the target mask bounding box, but a wildly mismatched aspect ratio means heavy cropping. Resize the mask or pick a closer crop of the transfer.
• Seams at the boundary - raise feather (8 → 12–16) and lower blend_strength slightly so the composite doesn't fight the original edge.