HiDream-O1-Image

HiDream-O1-Image is a natively unified image generative foundation model built on a Pixel-level Unified Transformer (UiT) without external VAEs or disjoint text encoders, which natively encodes raw pixels, text, and task-specific conditions in a single shared token space — supporting text-to-image, image editing, and subject-driven personalization at up to 2,048 × 2,048.

Project Updates

• 🤗 May 10, 2026: Try HiDream-O1-Image online on Hugging Face Spaces — 🤗 HiDream-O1-Image and 🤗 HiDream-O1-Image-Dev.
• 📕 May 10, 2026: Our technical report is now available — 📑 HiDream-O1-Image.pdf.
• 🚀 May 8, 2026: We've open-sourced HiDream-O1-Image (8B), including both the undistilled and distilled Dev variants, together with the Reasoning-Driven Prompt Agent.

HiDream-O1-Image (codename: Peanut) debuts at #8 in the Artificial Analysis Text to Image Arena, which is positioned to be the new leading open weights Text to Image model (2026-5-5).

_{Artificial Analysis Text to Image Arena at up to 2,048 × 2,048.}

_{General text-to-image generation at up to 2,048 × 2,048.}

_{Long-text rendering & layout control — accurate, multi-region, multilingual text.}

_{Subject-driven personalization — preserve identity / IP across new scenes.}

Key Features

• 🧬 Pixel-Level Unified Transformer — One end-to-end model on raw pixels, no VAE, no disjoint text encoder.
• 🎨 One Model, Many Tasks — Text-to-image, long-text rendering, instruction editing, subject-driven personalization, and storyboard generation in a single architecture.
• 🧠 Reasoning-Driven Prompt Agent — Built-in "thinking" agent that resolves implicit knowledge, layout, and text rendering before generation.
• 🖼️ Native High Resolution — Direct synthesis up to 2,048 × 2,048 with sharp fine-grained detail.
• ⚡ Exceptional Efficiency and Versatility at 8B Scale — With only 8B parameters, achieves performance parity with or even surpasses larger open-source DiTs and leading closed-source models.

Models

Evaluation

We benchmark HiDream-O1-Image against state-of-the-art open-source and proprietary models on five widely-used evaluation suites covering compositional generation, dense prompt alignment, human preference, complex visual text generation, and long-text rendering. In each table, the best result is highlighted in bold and the second-best is underlined. Click any benchmark below to expand or collapse.

Installation

1. Clone this repository:

git clone https://github.com/HiDream-ai/HiDream-O1-Image.git
cd HiDream-O1-Image

2. Install the required dependencies:

pip install -r requirements.txt

Note on flash-attn. We highly recommend installing flash-attn for optimized attention computation. If you do not (or cannot) install flash-attn, you must edit models/pipeline.py line 291 and change "use_flash_attn": True to "use_flash_attn": False — otherwise inference will fail to import the kernel.

Reasoning-Driven Prompt Agent

HiDream-O1-Image ships with a Reasoning-Driven Prompt Agent (prompt_agent.py) that explicitly reasons through layout, subject attributes, physical logic, and text-rendering details, then rewrites a raw user instruction into a self-contained English prompt. It supports two backends — pick one with --backend.

The agent prints a JSON object with three fields: prompt (rewritten English prompt), reasoning, and resolved_knowledge. Feed the prompt field into inference.py for best results on intricate, reasoning-heavy requests.

Option A — Local Backend (Gemma-4-31B-it)

1. Download the Gemma weights (requires accepting the Gemma license on HuggingFace):

huggingface-cli download google/gemma-4-31B-it --local-dir /path/to/gemma-4-31B-it

2. Run the refiner locally:

python prompt_agent.py \
    --backend local \
    --model_id /path/to/gemma-4-31B-it \
    --prompt "李白的静夜思写在古墙上"

Option B — External OpenAI-Compatible API

Use any OpenAI-compatible endpoint (OpenAI, Azure, vLLM, SGLang, DeepSeek, etc.) by providing --base_url, --api_key, and --model_name:

python prompt_agent.py \
    --backend api \
    --base_url https://api.openai.com/v1 \
    --api_key $OPENAI_API_KEY \
    --model_name deepseek-v4-pro \
    --prompt "李白的静夜思写在古墙上"

Usage

A CUDA-capable GPU is required for inference. The examples below use the undistilled model (--model_type full); see the last subsection for running the same tasks with the distilled model (--model_type dev).

1. Text-to-Image Generation

Generate an image from a text prompt:

python inference.py \
    --model_path /path/to/HiDream-O1-Image \
    --prompt "medium shot, eye-level, front view. A woman is seated in an ornate bedroom, illuminated by candlelight, with a calm and composed expression. The subject is a young woman with fair skin, light brown hair styled in an updo with loose tendrils framing her face, and blue eyes. She wears a cream-colored satin robe with delicate floral embroidery and lace trim along the neckline. Her ears are adorned with pearl drop earrings. She is seated on a bed with a dark, intricately carved wooden headboard. To her left, a wooden nightstand holds three lit white candles and a candelabra with multiple lit candles in the background. The bed is covered with patterned pillows and a dark, textured blanket. The walls are paneled with dark wood and feature a large, ornate tapestry with muted earth tones. The lighting creates soft highlights on her face and robe, with warm shadows cast across the room." \
    --output_image results/t2i.png \
    --height 2048 \
    --width 2048

2. Instruction-Based Image Editing

Provide a single reference image and an editing instruction:

python inference.py \
    --model_path /path/to/HiDream-O1-Image \
    --prompt "remove the earphones" \
    --ref_images assets/edit/test.jpg \
    --output_image results/edit.png \
    --keep_original_aspect

3. Multi-Reference Subject-Driven Personalization

Provide two or more reference images that define the subject(s), and a prompt that places them in a new scene:

python inference.py \
    --model_path /path/to/HiDream-O1-Image \
    --prompt "A young boy with blonde hair stands on steps wearing light blue jeans, a white t-shirt with logo, and blue and white sneakers. He wears a brown cord necklace with beads, a black wristwatch with digital display, and carries a yellow fanny pack with white zipper. In his hand is a red boxing glove with white top, a teal plastic toy car, and a plastic toy figure of Captain America. He wears a straw hat with cream band. Natural light illuminates the scene." \
    --ref_images assets/IP/1.jpg assets/IP/2.jpg assets/IP/3.jpg assets/IP/4.jpg assets/IP/5.jpg assets/IP/6.jpg assets/IP/7.jpg assets/IP/8.jpg assets/IP/9.jpg assets/IP/10.jpg \
    --output_image results/subject.png

4. Running with the Dev Model

All three tasks above can be run with the Dev model by switching --model_path to the Dev checkpoint and setting --model_type dev. For example:

python inference.py \
    --model_path /path/to/HiDream-O1-Image-Dev \
    --prompt "A dog holds a sign that says \"HiDream-O1-Image release.\"" \
    --output_image results/t2i_dev.png \
    --model_type dev

Command Line Arguments

• --model_path: Path to the complete HuggingFace model directory (undistilled or distilled).
• --prompt: Text prompt for the generation or editing task.
• --ref_images: Paths to one or more reference images (optional; space-separated).
• --output_image: Path to save the generated image (default: output.png).
• --height / --width: Output image dimensions (default: 2048 × 2048; values snap to valid resolutions internally).
• --model_type: full or dev (default: full). Selects the inference recipe:
  • full: 50 steps, guidance scale 5.0, shift 3.0, default scheduler.
  • dev: 28 steps, guidance scale 0.0, shift 1.0, flash scheduler with predefined timesteps.
• --seed: Random seed (default: 32).
• --guidance_scale: Guidance scale (default: 5.0). Only effective when --model_type is full.
• --noise_scale_start, --noise_scale_end: Control the scale of the noise injected by the scheduler at each denoising step; the per-step scale linearly interpolates from noise_scale_start (first step) to noise_scale_end (last step). See models/pipeline.py:262 and models/pipeline.py:273. Defaults: 7.5, 7.5.
• --noise_clip_std: Per-step clipping threshold (in units of the injected noise's standard deviation) applied to the noise added during scheduler stepping. See models/flash_scheduler.py:348-350. Default: 2.5.
• --keep_original_aspect: When exactly one reference image is provided, resize it with max_size=2048 and use its dimensions for the target image (preserves the reference's aspect ratio) if True.

Web Demo

app.py is a self-contained Flask web application that exposes all generation modes. It also integrates the Reasoning-Driven Prompt Agent.

Starting the server

python app.py \
    --model_path /path/to/HiDream-O1-Image \
    --host 0.0.0.0 \
    --port 7860

Then open http://localhost:7860 in your browser.

Command-line arguments

All four arguments can also be set via environment variables (see .env.example): HIDREAM_MODEL_PATH, HIDREAM_MODEL_TYPE, HIDREAM_HOST, and HIDREAM_PORT.

Prompt Agent in the UI

The sidebar contains a Prompt Agent panel that calls the same Reasoning-Driven Prompt Agent used by prompt_agent.py. Select either the OpenAI-compatible API backend (any endpoint, key, and model name) or the Local · Gemma backend (set HIDREAM_AGENT_MODEL in .env or the environment to point to your local Gemma-4-31B-it weights).

License

The code in this repository and the HiDream-O1-Image models are licensed under MIT License.