SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

HiDream-O1-Image - The Native Unified Image Generation Model Open-Sourced by HiDream.ai

Author: AI小集 · Original source: AI工具集

What is HiDream-O1-Image

HiDream-O1-Image is an 8B pixel-level native unifiedimage generationmodel open-sourced by HiDream.ai. It adopts the world's first UiT architecture, requiring no VAE or independent text encoder to directly generate 2048×2048 high-definition images within a single token space. It surpasses FLUX.2,Qwen-Image and GPT Image 2 on six benchmark tests including GenEval and HPSv3, becoming the highest-ranked open-weight model on the Artificial Analysis text-to-image arena.

HiDream-O1-Image

HiDream-O1-Image's Main Features

  • Text-to-Image Generation: Supports end-to-end generation at up to 2048×2048 native resolution, delivering cinematic-quality output without multi-stage super-resolution.
  • Instruction-Driven Editing: By passing a reference image through --ref_images along with natural language instructions, it enables precise edits such as object removal and style transfer.
  • Subject-Driven Personalization: Given 2 or more reference images of the same subject, the model can maintain character or object identity consistency in entirely new scenes.
  • Long-Text Rendering: Achieves 0.97+ accuracy in both Chinese and English long-text rendering on CVTG-2K and LongText-Bench, significantly outperforming mainstream models.
  • Storyboard Generation: Supports multi-frame continuous image generation while maintaining character and scene consistency, meeting the needs of film and television storyboard creation.
  • Reasoning-Driven Prompt Agent: Built-in "thinking" agent automatically parses implicit knowledge, spatial layout, and text layout logic before generation.

HiDream-O1-Image's Technical Principles

  • UiT Unified Architecture: HiDream-O1-Image adopts the world's first UiT (Unified Transformer) architecture, integrating the VAE, text encoder, and diffusion network that are separate in traditional diffusion models into a single end-to-end Transformer.
  • Subversion of the Three-Stage Pipeline: Unlike the three-stage pipeline of traditional Stable Diffusion and FLUX—"VAE compression + independent text encoder + latent space diffusion"—UiT performs diffusion denoising directly in the raw pixel space, without latent space compression or external encoder processing.
  • Single Shared Token Space: UiT maps image pixel patches, text tokens, and task condition tokens into the same shared representation space, using self-attention mechanisms to let text semantics directly relate to the position and color values of each raw pixel patch.
  • Pixel-Level Native Generation: By removing the VAE, the model avoids detail loss, color boundary artifacts, and resolution ceilings caused by compression and reconstruction, supporting end-to-end generation at 2048×2048 native resolution without any super-resolution post-processing.
  • Direct Cross-Modal Alignment: In the unified attention space, text tokens can directly attend to the coordinate values of any pixel patch, and pixel patches can in turn attend to the semantic information of text instructions, achieving direct cross-modal alignment at the underlying representation space.
  • Guidance Distillation Acceleration: The Dev version uses Guidance Distillation knowledge distillation technology, trained with the Full version as the teacher model, directly learning the CFG-enhanced distribution within 28 steps, eliminating the need for dual-path CFG computation during inference.
  • Multi-Task Zero Switching: The model distinguishes tasks such as text-to-image, instruction editing, subject personalization, and storyboard generation through task-conditioning tokens. All modes share the same set of 8B parameter weights, enabling zero-switching without loading LoRA or ControlNet.

How to Use HiDream-O1-Image

  • Environment Preparation: Clone the GitHub repository and install dependencies, ensuring you have a CUDA-capable GPU (the 8B model can run on a single card).
  • Download Weights: Pull from Hugging Face HiDream-ai/HiDream-O1-Image(Full version) or HiDream-O1-Image-Dev(fast version).
  • Text-to-Image Inference: Run the official example script, enter a text prompt, set 50 steps (Full) or 28 steps (Dev), and directly output a 2048×2048 image.
  • Instruction Editing: Add --ref_images input.jpg to the command line along with an editing instruction, such as "remove the earphones", to achieve zero-shot image modification.
  • Subject Personalization: Pass in more than 2 reference images of the same person or object, and the model automatically extracts identity features and embeds them into a new scene.
  • API and Low-Code Integration: Through the MIT-licensed codebase, integrate the model into your own agents or low-code platforms, supporting batch calls and local private deployment.

Core Advantages of HiDream-O1-Image

  • Architectural Innovation, Parameter Efficiency: With 8B parameters, it comprehensively surpasses the 56B FLUX.2 Dev and 27B Qwen-Image across six benchmarks including GenEval (0.90), DPG-Bench (89.83), and HPSv3 (10.37), improving parameter efficiency by 3–7 times.
  • Pixel-Native, Lossless Generation: After removing the VAE, it operates directly in pixel space, avoiding compression-reconstruction errors and supporting native 2048×2048 high-resolution output.
  • Leading in Long Text and Multilingualism: In LongText-Bench, the Chinese and English scores are 0.979 and 0.978 respectively, and CVTG-2K complex visual text generation reaches 0.9128, all in the global top tier.
  • Unified Weights, Multi-Task Zero-Switching: Text-to-image, instruction editing, personalization, and storyboard generation share the same set of weights, enabling task switching without loading LoRA or ControlNet.
  • Open Source and Commercial, Ecosystem-Friendly: The code and weights are released under the MIT license, supporting local deployment, low-code integration, and commercial use, and have been simultaneously published on Hugging Face and GitHub.

Project Address of HiDream-O1-Image

  • GitHub Repository：https://github.com/HiDream-ai/HiDream-O1-Image
  • HuggingFace Model Library：https://huggingface.co/HiDream-ai/HiDream-O1-Image
  • Technical Paper：https://github.com/HiDream-ai/HiDream-O1-Image/blob/main/assets/HiDream-O1-Image.pdf

Comparison of HiDream-O1-Image with Similar Competitors

Dimension HiDream-O1-Image Stable Diffusion 3.5 Ideogram 3.0
Developer HiDream.ai Stability AI Ideogram
Model Type Open-source image generation Open-source image generation Closed-source image generation
Parameter Scale 8B(single unified model) Approximately 8B（MM-DiT） Not disclosed
Core Architecture UiT pixel-level unified Transformer, no VAE, no separate text encoder Latent space MM-DiT, relies on VAE compression and separate text encoder Latent space Transformer, relies on VAE compression
Image Fidelity Pixel-level native generation, lossless reconstruction, sharp details Latent space generation, high-frequency details suffer VAE reconstruction loss Latent space generation, overall image quality is high but limited by the compression pipeline
Maximum native resolution 2048×2048 end-to-end, no super-resolution post-processing required Usually requires a super-resolution model or tiled generation Usually requires a super-resolution model
Long text rendering LongText-Bench English 0.979, Chinese 0.978 Moderate long-text capability, complex layouts prone to errors World's strongest, industry benchmark for complex poster-level text layout
CVTG-2K (complex visual text) 0.9128 approx. 0.85 0.95+
Open-source license MIT (commercial use allowed, modification allowed) Open license (commercial use allowed) Closed source, API/subscription use only
Local deployment 8B unified weights runnable on a single GPU Runnable on a single GPU, low hardware requirements Not supported, cloud calls only
Ecosystem and controllability A single set of weights covers text-to-image/editing/personalization/storyboard, no LoRA needed Largest ecosystem, rich plugins such as LoRA, ControlNet, IPAdapter No ecosystem plugins, relies on official feature iterations
Inference efficiency 28–50 steps, Dev version requires no dual-pass CFG computation 20–50 steps, requires CFG and loading multiple modules Cloud inference, speed depends on server load
Core advantages Ultimate parameter efficiency, native pixel-level high fidelity, unified architecture with zero switching across multiple tasks Most mature open-source ecosystem, most complete controllable toolchain, richest community resources World's strongest long text and layout, stable image aesthetics

Application scenarios of HiDream-O1-Image

  • Professional film and television creation: Relying on cinematic image quality and character consistency capabilities, it provides professional teams with full-process AI film and television production support from creative ideation and storyboard design to final output.
  • Cross-border e-commerce marketing: Through platforms such as HiBurst, it batch-generates product display images and marketing videos, significantly reducing sellers' visual content production costs and increasing overseas conversion rates.
  • Social media content production: With tools such as vivago, it enables second-level special effects generation and end-to-end short video creation, helping self-media and MCN agencies greatly improve content output efficiency.
  • Localized design workflow: The 8B parameter scale is suitable for local deployment on consumer-grade GPUs, allowing designers to complete high-resolution poster, illustration, and brand material generation in a private environment.
  • Multilingual visual publishing: With precise Chinese-English bilingual long-text rendering capabilities, it is suitable for professional publishing scenarios requiring complex mixed text and image layout, such as book covers, magazine layout, and educational courseware.