Step1X-Edit

Closing the gap to closed systems like GPT-4o image editing is the stated goal of Step1X-Edit, StepFun's open framework that couples a multimodal LLM, which digests the reference image and the written instruction, with a DiT-style diffusion decoder that renders the edited result. Alongside the model the team released GEdit-Bench, an evaluation benchmark assembled from genuine user editing requests rather than synthetic prompts, scored with GPT-4 and Qwen2.5-VL judges. Releases have moved quickly: v1.1 refined quality, Step1X-Edit-v1p2 arrived in November 2025 as a native reasoning edit model, and repository activity continues into 2026. Hardware demands are documented candidly in the README: full precision peaks between 42.5 and 49.8 GB of GPU memory depending on resolution, while FP8 quantization with CPU offload fits in 18 GB at a substantial speed cost. Inference uses PyTorch with a patched Diffusers branch specific to each model version, and weights download from Hugging Face. Code and models are Apache-2.0, leaving the roughly 2,200-star project free for commercial work.