Eric_UniPic3
ComfyUI custom nodes for UniPic3 image editing
Paketprofil
- Status
- Überprüft
- Verleger
- @EricRollei
- Quelle
- GitHub
README
Eric UniPic3 Nodes for ComfyUI
Version 1.0.0 | Author: Eric Hiss (GitHub: EricRollei) | License: MIT
ComfyUI custom nodes for Skywork UniPic3 - a state-of-the-art unified multimodal framework for single-image editing and multi-image composition with Human-Object Interaction (HOI) capabilities.
Table of Contents
- About UniPic3
- What's Novel
- Prompt Format & Image Referencing
- Model Variants
- Node Reference
- Usage Examples
- Design Philosophy
- Technical Details
- Future Work
- Credits & Acknowledgments
- Installation
About UniPic3
Overview
UniPic3 is the third generation of Skywork's UniPic family, released in January 2026. Unlike its predecessors (UniPic1 and UniPic2 which include text-to-image capabilities), UniPic3 is specifically designed for image editing and multi-image composition tasks.
The model excels at:
- Single-image editing: Modify images based on text instructions while preserving unedited regions
- Multi-image composition: Combine 2-6 reference images into a coherent unified scene
- Human-Object Interaction (HOI): Create realistic compositions where people naturally interact with objects (wearing clothes, holding items, using equipment)
What UniPic3 is NOT
UniPic3 does not support pure text-to-image generation. It always requires at least one input image. For T2I capabilities, use UniPic1 or UniPic2.
Supported Tasks
| Task | Input | Description |
|---|---|---|
| Single Edit | 1 image + prompt | Edit an image with text instructions |
| Edit with Reference | 1 image + refs + prompt | Edit using style/content from references |
| Multi-Image Compose | 2-6 images + prompt | Combine multiple elements into one scene |
| Virtual Try-On | Person + clothing images | Put garments on a person naturally |
| Product Composition | Product + background + props | Create product photography compositions |
What's Novel
UniPic3 introduces several innovations that set it apart from previous image editing models:
1. Unified Sequence Modeling
UniPic3 reformulates both single-image editing and multi-image composition as conditional generation on a unified sequence representation. This means:
- A single model handles both tasks without mode switching
- Knowledge transfers between editing and composition tasks
- Consistent behavior across different input configurations
2. Arbitrary Input Flexibility
Unlike models like Qwen-Image-Edit-2509 which support limited input configurations:
- 1-6 input images supported in a single forward pass
- Arbitrary resolutions within a 1024×1024 pixel budget
- Arbitrary output resolutions (not fixed to square outputs)
3. HOI-Focused Training
The model was trained with a focus on Human-Object Interaction scenarios:
- 700K high-quality HOI training samples covering apparel, tools, instruments, furniture, sports equipment
- Comprehensive data curation pipeline specifically designed for multi-image composition
- Handles realistic occlusions, lighting, and spatial relationships
4. Distillation for Fast Inference
Three model variants provide quality/speed tradeoffs:
- Teacher: Full 50-step diffusion for maximum quality
- DMD: Distribution Matching Distillation for 8-step inference
- Consistency: Trajectory-consistent mapping for geometric alignment
The distilled models achieve 12.5× speedup while maintaining high quality.
Prompt Format & Image Referencing
Numbered Image References
UniPic3 understands numbered image references in prompts. Since it uses Qwen2.5-VL (a vision-language model) as the text encoder, you can explicitly reference which image should contribute what to
Knoten in diesem Paket
4 Knoten (Plural)
Knoten in diesem Paket
4 Knoten (Plural)- EricUniPic3Compose
- EricUniPic3ImageEdit
- EricUniPic3Loader
- EricUniPic3Unload
Quellen
3 Quellen
Quellen
3 QuellenVerifiziert 21. Okt.
Verifiziert 21. Okt.
Quellenauszüge
1 Auszug
Quellenauszüge
1 AuszugSource context: Repo EricRollei/Eric_UniPic3
