SkyworkAI/UniPic
Open-source SOTA multi-image editing model
What it solves
Skywork-UniPic is a series of unified multimodal models designed to handle image generation, editing, and understanding within a single framework. It eliminates the need for separate models for different visual tasks, providing a versatile toolset for high-fidelity synthesis and robust perception.
How it works
The project consists of three distinct modeling paradigms:
- UniPic-3: A multi-image editing model that supports 1–6 input images with flexible resolutions. It uses CM + DMD distillation to achieve 8-step inference, resulting in a 12.5x speedup.
- UniPic-2: Based on efficient architectures with Diffusion Post-Training, featuring variants like SD3.5M-Kontext and MetaQuery for text-to-image generation, fine-grained editing, and multimodal reasoning.
- UniPic-1: A 1.5B parameter model using Unified Autoregressive Modeling to handle both visual perception (understanding) and synthesis (generation) in a single transformer.
Who it’s for
Content creators, AI researchers, and developers who need a unified approach to image manipulation, high-speed image editing, and multimodal visual reasoning.
Highlights
- Unified Framework: Combines generation, editing, and understanding in one series.
- Multi-Image Support: UniPic-3 can process up to 6 input images for composition.
- High-Speed Inference: UniPic-3 offers 12.5x faster inference via 8-step distillation.
- Diverse Paradigms: Offers autoregressive and diffusion-based post-training approaches.