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

What is Step Image Edit 2?

Step Image Edit 2 is a new generation of lightweight image generation and editing model launched by Step Star. With only 3.5B parameters, it outperforms 12B-20B open-source models in practical applications. The model generates images in 0.5-2 seconds per pass, emphasizing rapid response and high-quality output. Covering image generation and editing, it supports Chinese and English rendering, partial editing, visual reasoning, subject consistency preservation, and style transfer. It can be applied to scenarios such as IP creation, poster design, comic generation, portrait beautification, travel photo retouching, and portrait photo generation.

Main functions of Step Image Edit 2

• Image generationIt can quickly generate high-quality images based on text descriptions, with each image generation taking only 0.5-2 seconds.
• Image editingIt supports operations such as partial editing, subject replacement, and style transfer on existing images.
• Chinese and English renderingSpecifically optimized for text editing scenarios, it can accurately generate and modify Chinese and English content in images.
• Partial redrawIt supports fine-grained modification of specific areas of an image while keeping non-editable areas unchanged.
• Visual reasoning: Possesses the ability to understand the relationships between image content and to make reasonable editing and reasoning.
• Subject ConsistencyMaintain the stability of the main features during multiple rounds of editing or style transfer.
• Style transferApply a specified artistic style to an image or a specific area.

Technical Principles of Step Image Edit 2

• Multi-expert-driven self-evolutionary learningThe system employs a two-stage training framework of "separate exploration and centralized aggregation". Multiple sub-task expert branches are derived from the base model to capture high-quality editing trajectories in complex and noisy data. Through iterative self-distillation, expert knowledge is aggregated back into the base model, breaking through the capability limit without increasing the parameter scale, and achieving a lightweight model that approximates the performance of a large model.
• Distribution Matching Reinforcement Learning (DARL)The reinforcement learning objective is redefined as aligning the model's output distribution with the reference distribution, rather than relying on traditional single-point reward signals. By comparing the distribution difference between the overall model output and the reference output as a dense reward, evaluation bias from a few small samples is avoided, resulting in smoother training and stronger generalization ability for complex tasks.
• Breakthroughs in data quality and scaleWe invested over 50 million pieces of specialized training data, integrating data from three sources: real-world scenario mining, targeted synthesis, and high-quality open-source data. Addressing the challenges of text editing, we developed a self-developed typesetting system that generated 20 million pieces of specialized data, constructing a three-tiered quality control system: automatic agent cleaning, large-scale model global evaluation, and meticulous manual screening.

How to use Step Image Edit 2

• Access the Stepping Star Open PlatformVisit the Stepfun Open Platform: https://platform.stepfun.com/docs/zh/guides/models/step-image-edit-2.
• Get API accessRegister and log in to your platform account to obtain API access for Step Image Edit 2.
• Calling the image generation/editing interfaceAccording to the documentation, pass in text prompts or images to be edited and editing instructions via the API.
• View the Step Plan integration solutionVisit https://platform.stepfun.com/docs/zh/step-plan/integrations/image-api for more detailed integration instructions.