IC-Edit

IC-Edit revolutionizes instruction-based image editing with its ultra-efficient Diffusion Transformer framework, achieving state-of-the-art results using just 1% trainable parameters (200M) and 0.1% training data (50k samples). Unlike commercial tools like GPT-4o, this open-source solution delivers precise edits in ~9 seconds per image—even on 4GB VRAM GPUs—while preserving object identity better than leading alternatives.

Key innovations include LoRA-MoE hybrid tuning for flexible adaptation and VLM-guided noise filtering for higher-quality outputs. Ideal for developers and creators, IC-Edit handles multi-turn edits (e.g., "change the dog to a cat, then add a hat") without costly retraining, making advanced image manipulation accessible to all.