Step1X-Edit

What it is

Step1X-Edit is an open image-editing model released by StepFun on 25 April 2025. The release included inference code, model weights, a technical report and an evaluation set called GEdit-Bench^[1]. The technical report is arXiv:2504.17761^[2]. All of it is published under the Apache 2.0 license^[1]. An online demo opened on 26 April 2025, one day after the code and weights were made available^[1].

How it works

The model divides the editing job between two components. A multimodal large language model reads the reference image and the editing instruction together, and the latent embedding it produces is fed to a diffusion image decoder, which generates the edited image^[1]. Put simply, the language-model half works out what the instruction asks to change and where, and the diffusion half turns that understanding into pixels. A diffusion model starts from noise and denoises step by step toward a target; here the target is conditioned on the embedding the language model produced, so the instruction shapes the generation rather than being applied as a separate post-processing step.

Splitting the work this way lets instruction understanding be treated as a language-and-image problem, while the visual quality of the output is left to a generative decoder. A general editing model has to answer many different instructions with one set of weights, which is part of why the design leans on a language model for parsing the request instead of building a separate module for each kind of edit.

Benchmark and versions

GEdit-Bench ships alongside the model and is built on real-world user instructions rather than synthetic prompts^[2]; its G_ metrics are scored by GPT-4.1^[1]. On the benchmark's overall score, G_O, version 1.0 reached 6.44^[1].

Version 1.1, released on 9 July 2025, added text-to-image generation and improved editing quality, and scored 6.97 on G_O^[1].

A v1.2 preview with native reasoning-based editing came on 8 September 2025 and scored 7.42 on G_O^[1]. The full v1.2, called ReasonEdit-S in its paper, followed on 26 November 2025; it scored 7.24 on G_O, or 7.58 with thinking and reflection, while v1.2 with thinking scored 7.36^[1]. Taken together, the version history records both an increase in editing quality and a move toward letting the model reason about an edit before producing it.

Fine-tuning and everyday use

LoRA fine-tuning on a single 24GB GPU was added on 22 May 2025, and that release included a hand-fixing LoRA for anime characters^[1]. Together with the Apache 2.0 license and public weights^[1], this made it possible to adapt the model on ordinary hardware instead of a large cluster. Releasing a benchmark at the same time as the model also gives later versions and outside models a common basis for comparison.

Where it sits

The technical report's abstract claims performance comparable to closed-source models such as GPT-4o and Gemini 2 Flash^[2]; that comparison is the authors' own assessment in the paper, while the figures above come from the project's GEdit-Bench. On 29 April 2026, StepFun announced Step Image Edit 2, a separate lightweight hosted model on its open platform rather than a new version of Step1X-Edit^[1].