FireRedTeam/FireRed-Image-Edit

Based on the README, here's a technical summary: Built on a backbone-agnostic architecture, the model injects editing capabilities through a full Pretrain → SFT → RL pipeline transferable to any T2I foundation model. It features an Agent module for automatic instruction rewriting, ROI detection, and multi-image stitching, plus offline feature pre-extraction that decouples VLM inference from training to eliminate generation overhead. Deployment targets include native ComfyUI nodes, GGUF quantization format, and optimized inference reaching 4.5s/sample on 30GB VRAM through distillation, quantization, and static compilation—with full open-source LoRA training code and REDEdit-Bench evaluation dataset.

1,073 stars. Actively maintained with 11 commits in the last 30 days.