Step1X-Edit: Open Source AI Image Editing Framework
Step1X-Edit is a newly released open-source image editing framework that uses multimodal large language models to process reference images and user editing instructions, extracting latent embeddings and integrating them with a diffusion image decoder to obtain the target image. This project aims to provide performance comparable to closed-source models like GPT-4o and Gemini2 Flash.
Key Features
- Natural Language Editing: Supports image editing via natural language instructions, such as “beautify,” “change the background,” or “convert to pixel art style”
- Open Source and Commercial Use: Released under the Apache 2.0 license, freely available for commercial use
- Flexible Hardware Requirements: Offers FP8 quantized version to reduce memory requirements
- Community Support: Already has multiple community versions, including FP8 quantized models
Technical Specifications
Step1X-Edit provides multiple versions to accommodate different hardware configurations:
| Model Version | Peak GPU Memory (512/786/1024 resolution) | 28 Steps Generation Time (seconds) |
|---|---|---|
| Standard | 42.5GB / 46.5GB / 49.8GB | 5s / 11s / 22s |
| FP8 Quantized | 31GB / 31.5GB / 34GB | 6.8s / 13.5s / 25s |
| Standard+CPU Offload | 25.9GB / 27.3GB / 29.1GB | 49.6s / 54.1s / 63.2s |
| FP8 Quantized+CPU Offload | 18GB / 18GB / 18GB | 35s / 40s / 51s |
While officially recommended to use a GPU with 80GB of memory for best performance and efficiency, Step1X-Edit can also run on graphics cards with less memory through FP8 quantization and CPU offloading technology.
Online Demo
You can directly experience Step1X-Edit through the following link:
Future Plans
The Step1X-Edit team has already completed:
- Inference code and model weights release
- Online demo (Gradio)
- FP8 quantized weights
Planned features include:
- Diffusers integration
- ComfyUI integration
Related Links
- GitHub Project Page
- HuggingFace Model
- FP8 Quantized Model
- ModelScope Model
- Technical Report (arXiv)
- GEdit-Bench Evaluation Dataset
The release of this open-source project will enable more users to access high-quality AI image editing capabilities without relying on closed-source models, while its commercializable nature provides more possibilities for developers.