Zhejiang University & Harvard University · Image Editing

ICEdit

Instruction-based image editing framework using Diffusion Transformers with in-context prompting, requiring only 1% trainable parameters (200M) and 0.1% training data (50k) compared to existing methods. Achieves competitive results with commercial models like Gemini and GPT-4o while being faster (~9 seconds per image) and more efficient on consumer hardware (4GB VRAM).