Eric UniPic3 Compose (HOI)

This is the node that justifies the pack. Skywork trained UniPic3 with multi-image composition as its headline capability - a "unified sequence" model that fuses several pictures into one coherent scene with realistic folds, occlusions and lighting, backed by 700K human-object-interaction (HOI) training samples. Eric UniPic3 Compose (HOI) is that capability exposed as a node: a person in image1, a jacket in image2, shoes in image3, and you get a photo of that person wearing that jacket in those shoes. Not a collage. A photo.

How it differs from the Edit node

Edit changes one image. Compose fuses at least two into a new unified image. The model treats both as the same task under the hood - that's the "unified" in UniPic3, no mode switching - but the node enforces a minimum of two inputs because a single-image compose is really an edit. Image order matters, and the author is explicit about it: image1 is the primary subject (the person for try-on, the product for product shots), image2 is the first thing to add, and image3–image6 stack on accessories, props and background elements. Because the text encoder is a vision-language model, you reference them by number in the prompt: "The person from image 1 wearing the jacket from image 2". The more explicit you are about which image contributes what, the better it behaves.

The inputs that matter

• pipeline - from Eric UniPic3 Load Model.
• image1 - primary subject, required.
• image2 - first item to compose, required.
• image3–image6 - additional items, all optional.
• prompt - how to compose; the default is "The person wearing the clothing".
• true_cfg_scale, steps, seed - the usual quality controls (4.0, 50, random).
• negative_prompt - optional, what to avoid.

The output is a single image tensor, ready for Preview Image or Save Image.

What it's actually good at

• Virtual try-on - person plus one or more garments.
• Product photography - product plus background scene plus props.
• Full outfits - person plus top, bottom, shoes, bag in one pass.

This is where the pack genuinely pulls ahead of the Qwen-Edit path: Qwen-Image-Edit-2509 handled one to three input images and was tuned for that range; UniPic3 takes up to six with HOI-specific training. If your job is clothing or products, that's the whole game - try-on and product shots are exactly what this training was built for.

Troubleshooting

• Order changes everything. If the jacket comes out as the subject, swap image1 and image2 and re-run. The model leans on image position, not just your prompt.
• Same 1MP cap as Edit. Output is capped at about 1024×1024 by a hardcoded diffusers constant; docs/HIGH_RES_EDITING.md in the repo raises it to 8MP.
• Steps stay at 50 by default even for distilled variants - set 8 yourself when you loaded dmd/consistency.

Installing

Ships inside Eric_UniPic3 alongside the loader, Edit and Unload nodes - ComfyUI Manager or git clone https://github.com/EricRollei/Eric_UniPic3 into custom_nodes, pip install -r requirements.txt, then the model downloads. The Compose node is dead weight without the loader's pipeline wired in first, so get the ~58GB model loading before you build the composition graph.