Step Image Edit 2 - An image generation and editing model launched by StepFun

What is Step Image Edit 2

Step Image Edit 2 is a new-generation lightweight image generation and editing model launched by StepFun, with only 3.5B parameters, capable of surpassing 12B-20B class open-source large models in actual performance. The model takes 0.5-2 seconds for a single image generation, focusing on extremely fast response and high-quality output. The model covers image generation and editing, supports Chinese and English rendering, local editing, visual reasoning, subject consistency preservation, and style transfer, and can be applied to scenarios such as IP creation, poster design, comic generation, portrait beautification, travel photo editing, and portrait photo generation.

Main features of Step Image Edit 2

• Image generation: Quickly generate high-quality images based on text descriptions, with a single image generation taking only 0.5-2 seconds.
• Image editing: Supports operations such as local editing, subject replacement, and style transfer on existing images.
• Chinese and English rendering: Specially optimized for text editing scenarios, capable of accurately generating and modifying Chinese and English content in images.
• Local repainting: Supports fine-grained modification of specific areas of an image while keeping non-edited areas unchanged.
• Visual reasoning: Possesses the ability to understand relationships among image contents and perform reasonable editing reasoning.
• Subject consistency: Maintains stable subject features during multi-round editing or style transfer.
• Style transfer: Applies a specified artistic style to an image or local area.

Technical principles of Step Image Edit 2

• Multi-expert driven self-evolving learning: Adopts a two-stage training framework of "separate exploration, centralized aggregation". Multiple subdivided task expert branches are derived from the base model to capture high-quality editing trajectories in complex high-noise data; through iterative self-distillation, expert knowledge is aggregated back into the base model, breaking through the capability ceiling without increasing parameter scale, enabling a lightweight model to approach the performance of large models.
• Distribution-Aligned Reinforcement Learning (DARL): Redefines the reinforcement learning objective as aligning the model output distribution with a reference distribution, rather than relying on traditional single-point reward signals. By using the distribution gap between the model's overall output and the reference output as a dense reward, it avoids evaluation bias from a small number of samples, making complex task training more stable and generalization ability stronger.
• Breakthroughs in data quality and scale: Invested over 50 million scale of specialized training data, integrating three sources: real-scenario mining, targeted synthesis, and high-quality open-source data. In response to text editing difficulties, a self-developed typesetting system generated 20 million specialized data entries, and a three-level quality control system of "agent automatic cleaning - large model global evaluation - manual fine screening" was built.

How to use Step Image Edit 2

• Visit the StepFun open platform: Visit the StepFun open platform https://platform.stepfun.com/docs/zh/guides/models/step-image-edit-2。
• Obtain API access: Register and log in to the platform account to obtain API call permissions for Step Image Edit 2.
• Call the image generation/editing interface: According to the documentation guidelines, pass text prompts or images to be edited and editing instructions through the API.
• View the Step Plan integration solution: Visit https://platform.stepfun.com/docs/zh/step-plan/integrations/image-api to learn more detailed integration methods.