ICEdit

Why we included this project

ICEdit treats instruction-based image editing as in-context generation on a diffusion transformer, which lets a single lightweight LoRA adapter handle consistent multi-turn edits while keeping the subject's identity stable across turns. The authors released a Hugging Face demo and MoE weights, and inference fits in roughly 4GB of GPU memory, so a developer workstation is enough rather than a GPU cluster. What makes it interesting for people fine-tuning edit models is the data and parameter budget: it needs about 0.5% of the training data and 1% of the parameters of earlier state-of-the-art methods, which makes reproduction and adaptation far more practical. One caveat: the FLUX.1 [dev] base carries a non-commercial license, so check your use case before shipping anything.