一句话总结：Step1X-Edit， 不只能"改图"，更能"听得懂、改得准、保得住"。 开源链接与体验地址： Github： https://github.com/stepfun-ai/Step1X-Edit HuggingFace： https://huggingface.co/stepfun-ai/Step1X-Edit ModelScope： https://www.modelscope . cn/models/stepfun-ai/Step1X-Edit/summary 技术 Report： https://arxiv.org/pdf/2504.17761 阶跃星辰 . Step1X-Edit 首次在开源体系中实现 MLLM 与 DiT 的深度融合，在编辑精度与图像保真度上实现大幅提升。在最新发布的图像编辑基准 GEdit- Bench 中，Step1X-Edit 在语义一致性、图像质量与综合得分三项指标上全面领先现有开源模型，比肩 GPT-4o 与 Gemini 2.0 Flash。 以下文章来源于阶跃星辰 ，作者拥抱开源的 智能阶跃，十倍每一个人的可能。 阶跃星辰正式发布 并开源 图像编辑大模型 ...

Core Insights - The company, Step1X-Edit, has announced the open-source release of its image editing model, achieving state-of-the-art (SOTA) performance in the open-source domain [1] Group 1: Model Specifications - The model has a total parameter count of 19 billion, comprising 7 billion for MLLM and 12 billion for DiT [1] - It features three key capabilities: precise semantic parsing, identity consistency maintenance, and high-precision regional control [1] Group 2: Supported Tasks - The model supports 11 types of high-frequency image editing tasks, including text replacement, style transfer, material transformation, and portrait retouching [1]