StepFun open-sources the Step1X-Edit image editor

Background StepFun (阶跃星辰) had been building image models as part of its Step series. Image editing, which changes an existing picture according to a written instruction, was the focus of this release.

What happened On 25 April 2025 StepFun released the inference code, weights and technical report for Step1X-Edit ^[1]. It also released GEdit-Bench, an image-editing benchmark, together with its evaluation code and data ^[1].

An online demo followed on 26 April. Within days, community members produced FP8 weights and a ComfyUI plugin for the model ^[1]. ComfyUI is a node-based tool that many people use to run image models locally.

Key details At launch the model scored 6.44 overall on GEdit-Bench ^[1]. The model is licensed under Apache 2.0 ^[1].

Later versions extended the model. Version 1.1, released in July 2025, added text-to-image generation. Version 1.2, released in November 2025, added reasoning-based editing ^[1].

Significance The release gave the community an open image-editing model together with a public benchmark for comparing editing quality. The launch scores are the developers' own figures. The later versions show the project moving from editing alone to generation and reasoning-based editing.

Terms used An image-editing model takes an existing picture and an instruction, such as changing the colour of an object or replacing a background, and returns a modified picture. A benchmark is a standard set of test cases with a scoring method. Because benchmarks are often designed by a model's developers, a score shows how the model performs on that test, and comparisons are most useful when the same test is used for every model.

Why the format matters FP8 is a numerical format that uses fewer bits for each weight than the common 16-bit format. Smaller weights need less memory, which is why community FP8 versions were a practical step for people who wanted to run the model locally.