Step Image Edit 2 is a lightweight, unified AI model designed for both text-to-image generation and instruction-based image editing. Developed by StepFun (阶跃星辰), the model contains approximately 3.5 billion parameters and is optimized for low-latency performance, typically delivering generation or editing results within one to two seconds. It serves as a commercial counterpart to the open-source Step1X-Edit research, prioritizing high inference efficiency and stable API access for real-time interactive workflows.
The model’s architecture incorporates Multi-Expert Self-Evolution Learning, a training technique that splits the base model into specialized branches for targeted learning before compressing the acquired knowledge back into a compact base model through iterative self-distillation. It also utilizes Distribution-Matching Reinforcement Learning (DARL) to align the model’s output distribution with reference data, enhancing training stability and visual fidelity. These methods allow Step Image Edit 2 to maintain competitive performance against significantly larger open-source models in the 12B to 20B parameter range.
Functional capabilities include localized object editing, style transfer, and visual reasoning to interpret complex instructions. The model is capable of rendering both Chinese and English text within images and maintains high subject consistency across edits. It was trained on a dataset of over 50 million specialized samples, including 20 million dedicated specifically to text-editing tasks. On the KRIS-Bench academic benchmark, Step Image Edit 2 has achieved top rankings among lightweight image editing systems for its ability to handle factual and conceptual reasoning tasks.