HiDream-O1-Edit-1.5

HiDream-O1-Edit-1.5 is a visual generative foundation model specialized in high-precision image editing and instruction-following. Developed by HiDream, it is part of the O1 series and is built on a Pixel-level Unified Transformer (UiT) architecture. Unlike traditional diffusion models that rely on disjoint text encoders and Variational Autoencoders (VAEs), this model natively encodes raw pixels, text, and task conditions into a single shared token space, allowing it to treat image editing as a consistent in-context reasoning process.

A core feature of the model is its integration with a Reasoning-Driven Prompt Agent. This agent explicitly processes user instructions to resolve implicit knowledge and logical constraints—such as physical laws, spatial relationships, and lighting consistency—before the generation begins. This reasoning-heavy approach enables the model to handle complex multi-step editing tasks, such as replacing specific objects while maintaining the integrity of shadows and reflections or performing precise text rendering within edited regions.

Technical Architecture

The model's architecture eliminates the need for external modular components, which reduces information loss typically associated with latent-space representations. By operating directly in pixel patch space, HiDream-O1-Edit-1.5 can generate and modify images at resolutions up to 2048 × 2048. The 8B version of the model is designed to achieve performance parity with significantly larger models by leveraging the efficiency of its unified transformer structure and end-to-end training paradigm.

Editing Capabilities

HiDream-O1-Edit-1.5 supports a wide range of manipulation tasks, including object addition, removal, style transfer, and layout-controlled subject personalization. For optimal results, prompts should be descriptive and specify both the target of the edit and the desired contextual alignment. The model performs best when users provide clear logical instructions, allowing the reasoning agent to accurately map the intended changes to the visual scene. It is particularly effective for commercial design and high-fidelity retouching where text accuracy and structural consistency are critical.