Eric UniPic3 Image Edit

UniPic3 is not a text-to-image model, and Eric UniPic3 Image Edit is not a T2I node. It's an instruction editor: give it a photo and a sentence - "change the background to a sunset beach" - and it re-emits the whole image with the edit applied. If that workflow sounds familiar, it should. It's the same job Qwen-Image-Edit took from masked inpainting and ControlNet, and this pack literally builds on the Qwen-Image-Edit-2511 base for its VAE, text encoder and scheduler. The difference is the brain: Skywork's own ~13B UniPic3 transformer.

What makes it different: numbered references

Because the text encoder is Qwen2.5-VL, a vision-language model, it sees your images and your words at the same time. That's what makes this node's party trick possible: you can point at specific inputs by number. The primary image is "image 1"; each of the five reference inputs is "image 2" through "image 6". So "put the jacket from image 2 on the person in image 1" isn't a vibe, it's a pointer. The same mechanic does style transfer ("edit image 1 in the artistic style of image 2") and content addition ("add the hat from image 2"). And when you only need the primary image, skip the references entirely - the node handles a single image fine.

The inputs that matter

• pipeline - from Eric UniPic3 Load Model.
• image - the photo you're editing (that's "image 1" in your prompt).
• prompt - plain-language instruction; use numbered references when precision matters.
• ref_image1–ref_image5 - optional references ("image 2" … "image 6").
• true_cfg_scale - the main quality control, 4.0 by default.
• steps - 50 for the teacher variant; drop to 8 if you loaded dmd/consistency.
• negative_prompt - optional, what to avoid; leave empty unless you need it.

The output is a single image (an IMAGE tensor), wired straight into Preview Image or Save Image.

Where people get burned

Three things, in rough order of how often they bite:

• The whole-frame re-emission. The model returns a new image, not a patched one, so pixels outside your edit come back close but not identical, and the error compounds if you chain edits. Faces drift worst, especially faces the model hasn't seen before. If every pixel outside the change must stay put, this is the wrong tool - that's masked inpainting's job.
• The 1MP cap. Inputs get downscaled to about 1024×1024 and outputs are capped there too - hardcoded in the diffusers pipeline, not in the model. Feed it a 12MP phone photo and you get a 1MP edit back. The pack ships docs/HIGH_RES_EDITING.md with a patch that raises the ceiling to 8MP; read it before you blame the node for soft results.
• Steps don't auto-adjust. The node defaults to 50 even when you loaded a distilled variant. The tooltip tells you to use 8 for dmd/consistency, but it won't do it for you.

Installing

Same as the rest of the pack - this node ships inside Eric_UniPic3, so if the loader is installed, this is too. ComfyUI Manager or git clone https://github.com/EricRollei/Eric_UniPic3 into custom_nodes, pip install -r requirements.txt, then the ~100GB of model downloads (see the loader article for the full list). Nothing in the pack runs without the loader's pipeline, so install that first and get it loading before you build out the edit graph.