HiDream-O1-Image - A native unified image generation model open-sourced by Zhixiang Future.

What is HiDream-O1-Image?

HiDream-O1-Image is an open-source 8B pixel-level native unified image generation model from Zhixiang Future. It adopts the world's first UiT architecture, which can directly generate 2048×2048 high-definition images in a single token space without VAE and independent text encoder. It surpasses FLUX.2, Qwen-Image and GPT Image 2 in six benchmark tests including GenEval and HPSv3, becoming the highest-ranked open-source weighted model in the Artificial Analysis text-to-image arena.

Main functions of HiDream-O1-Image

• Text-based image generationSupports end-to-end generation at a native resolution of up to 2048×2048, outputting cinematic-quality images without the need for multi-stage super-resolution.
• Instruction-driven editing:pass --ref_images By inputting a reference image and using natural language commands, precise editing can be achieved, such as removing objects and transferring styles.
• Subject-driven personalizationInput two or more reference images of the same subject, and the model can maintain the consistency of the identity of the person or object in a completely new scene.
• Long text renderingThe accuracy of rendering long bilingual (English and Chinese) text in CVTG-2K and LongText-Bench both reached 0.97+, which is significantly better than mainstream models.
• Storyboard generationIt supports the generation of multiple consecutive frames, maintaining consistency between characters and scenes, and meeting the needs of film and television storyboard creation.
• Inference-driven Prompt AgentBuilt-in "Thinking" agent automatically parses implicit knowledge, spatial layout, and text formatting logic before generation.

The technical principles of HiDream-O1-Image

• UiT Unified ArchitectureHiDream-O1-Image adopts the world's first UiT (Unified Transformer) unified architecture, which integrates the VAE, text encoder and diffusion network that are separate in the traditional diffusion model into a single end-to-end Transformer.
• The revolution of three-stage pipelinesUnlike the traditional Stable Diffusion and FLUX's three-stage pipeline of "VAE compression + independent text encoder + latent space diffusion", UiT performs diffusion denoising directly in the original pixel space, without the need for latent space compression and external encoder processing.
• Single shared token spaceUiT maps image pixel blocks, text tokens, and task condition tokens to the same shared representation space, and uses a self-attention mechanism to allow text semantics to be directly associated with the position and color value of each original pixel block.
• Pixel-level native generationAfter removing VAE, the model avoids the loss of detail, color boundary artifacts and resolution ceiling caused by compression reconstruction, and supports end-to-end generation at native 2048×2048 resolution without any super-resolution post-processing.
• Cross-modal direct alignmentIn a unified attention space, text tokens can directly attend to the coordinates of any pixel block, and pixel blocks can also attend to the semantic information of text instructions in reverse, achieving direct cross-modal alignment of the underlying representation space.
• Guidance Distillation AccelerationThe Dev version uses Guidance Distillation, a knowledge distillation technique, to train the Full version as the teacher model. It learns the CFG-enhanced distribution directly within 28 steps, eliminating the need for dual-path CFG computation during inference.
• Multitasking with zero switchingThe model distinguishes between tasks such as text-to-image generation, instruction editing, subject personalization, and storyboard generation through task condition tokens. All modes share the same set of 8B parameter weights, allowing for zero-switching without loading LoRA or ControlNet.

How to use HiDream-O1-Image

• Environmental preparationClone the GitHub repository and install the dependencies, ensuring you have a GPU that supports CUDA (the 8B model can run on a single card).
• Download weightPull from Hugging Face HiDream-ai/HiDream-O1-Image(Full version) or HiDream-O1-Image-Dev(Quick version).
• Wenshengtu ReasoningRun the official example script, enter the text prompt, set 50 steps (Full) or 28 steps (Dev), and directly output a 2048×2048 image.
• Command EditingAdd to the command line --ref_images input.jpg In conjunction with editing commands, such as "remove the earphones"This enables zero-sample image modification.
• Personalized Subject: Input two or more reference images of the same person or object, and the model will automatically extract the identity features and embed them into the new scene.
• API and low-code integration: Through the code library licensed under the MIT license, the model can be integrated into your own intelligent agent or low-code platform, supporting batch invocation and local private deployment.

HiDream-O1-Image's core advantages

• Architectural innovation, high-efficiency parametersThe 8B parameters outperform the 56B FLUX.2 Dev and the 27B Qwen-Image across six benchmarks, including GenEval (0.90), DPG-Bench (89.83), and HPSv3 (10.37), resulting in a 3–7x improvement in parameter efficiency.
• Pixel-native, lossless generationAfter removing VAE, it operates directly in pixel space, avoiding compression and reconstruction errors, and supports native high-resolution output of 2048×2048.
• Leading in long texts and multilingualismLongText-Bench scores 0.979 for Chinese and 0.978 for English, while CVTG-2K scores 0.9128 for complex visual text generation, both placing them among the world's top performers.
• Unified weighting, zero-switch between multiple tasksText generation, instruction editing, personalization, and storyboard generation share the same set of weights, allowing task switching without loading LoRA or ControlNet.
• Open source for commercial use, ecosystem friendlyThe code and weights are released under the MIT license, supporting local deployment, low-code integration, and commercial use. They are also available on Hugging Face and GitHub.

HiDream-O1-Image project address

• GitHub repository: https://github.com/HiDream-ai/HiDream-O1-Image
• HuggingFace model libraryhttps://huggingface.co/HiDream-ai/HiDream-O1-Image
• Technical Papers: https://github.com/HiDream-ai/HiDream-O1-Image/blob/main/assets/HiDream-O1-Image.pdf

Comparison of HiDream-O1-Image with similar competing products

Application Scenarios of HiDream-O1-Image

• Professional film and television productionLeveraging cinematic image quality and character consistency, it provides professional teams with end-to-end AI-powered film and television production support, from creative conception and storyboard design to final output.
• Cross-border e-commerce marketingBy using platforms like HiBurst to generate product display images and marketing videos in bulk, sellers can significantly reduce visual content production costs and improve overseas conversion rates.
• Social media content production: By leveraging tools such as Vivago, we enable second-level special effects generation and end-to-end short video creation, helping self-media and MCN agencies significantly improve content production efficiency.
• Localized design workflowThe 8B parameter scale is adapted for local deployment on consumer-grade graphics cards, enabling designers to generate high-resolution posters, illustrations, and brand materials in a private environment.
• Multilingual Visual PublishingWith its ability to accurately render long texts in both Chinese and English, it is suitable for professional publishing scenarios that require complex text and image layouts, such as book covers, magazine layouts, and educational courseware.