Eric Qwen-Edit Multi-Image Fusion

This is the "two people become one photo" node. Eric Qwen-Edit Multi-Image Fusion takes 2 to 4 images and runs them through the Qwen-Image-Edit model together, so instead of you pasting pictures onto a canvas and hoping, the model composes them into a single coherent image - matching lighting, touching up the seam, putting the subjects somewhere that makes sense. Qwen-Edit-2509's multi-image editing is exactly what this builds on: person+person, person+product, person+scene. If your job is "product photo, but the product sits next to a person" or "my cat and my dog in one shot," this is the node.

How it actually works

Qwen-Image-Edit conditions on input images through two parallel paths, and this node is the best place to see them, because it exposes them per-image. The VL path runs each image through the built-in Qwen2.5-VL vision encoder at ~384 px and captures what is in it - semantics. The VAE/ref path encodes each image at output resolution and captures how the pixels look - appearance. Every image gets a vl_N toggle (include in semantic path, on by default) and a ref_N toggle (include in pixel path, off by default for images 3-4). That default is deliberate: pixel-level references for secondary images can bleed their look across the whole output, so secondary images start semantic-only.

One more knob decides the skeleton: main_image. Whichever image you pick here seeds the denoising process with its VAE latent, so it gets the strongest reconstruction. If you want a person from image_2 standing in front of image_1's background, main_image should probably be image_2 with a scene composition.

The inputs that matter

• image_1, image_2 (required), image_3, image_4 (optional) - your source pictures.
• composition_mode - group (subjects side by side), scene (image_1 as background, drop the rest into it), merge (fuse features into one subject), raw (your prompt is used exactly as written, and you reference "Picture 1" etc. yourself). group is the easiest first run.
• prompt - in non-raw modes, describe the scene and the references get auto-added. "All subjects are sitting together on a couch, looking at the camera" is the default and it's a fine template.
• subject_label - what the auto-prompt calls each subject ("person", "woman", "bear", "product"). Worth setting, because "the two people" beats "the two entities."
• steps - 8 with a lightning LoRA, 50 on the base model. Same rule as every node in this pack.
• true_cfg_scale - real classifier-free guidance, two transformer passes per step. 4.0 is the sane start.
• max_mp - output cap (0.5-16). Start at 4 and climb.

The single output is image (IMAGE), ready to save or hand to a sharpening/upscale pass.

When it goes wrong

If a subject's identity wanders - the model starts inventing a new face - that's the VAE/ref path for that image being off. Flip its ref_N to True. If instead the whole output looks like one image's color bled over everything, you've gone the other way and should drop refs back to VL-only. And remember the underlying model's quirk the KB hammers on: Qwen-Edit re-emits the whole frame, so pixels you didn't ask to change shift a little. Multi-image fusion is where that trade is worth it (the alternative is manual compositing), but don't expect pixel-identical faces on a long chain of edits. Install via ComfyUI Manager ("Eric Qwen Edit") or git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments, with the edit model already loaded by Eric Qwen-Edit Loader - this node is a consumer, not a loader.