HiDream-O1-Image: Specs, Variants and Local Setup

hidream.org is an unofficial, independent site and is not affiliated with HiDream.ai. This guide summarizes HiDream.ai's published sources, linked throughout, as of November 5, 2026. Our own tools do not run this model.

What HiDream-O1-Image is

HiDream-O1-Image is an image model that HiDream.ai open-sourced on May 8, 2026. The official README describes it as a "natively unified image generative foundation model" built on a Pixel-level Unified Transformer (UiT), without external VAEs or disjoint text encoders. Raw pixels, text and task-specific conditions are encoded in one shared token space.

In practice, one 8B-parameter model covers several jobs, according to the README:

• Text-to-image at up to 2,048 × 2,048
• Instruction-based image editing
• Subject-driven personalization from reference images
• Long-text rendering and storyboard generation

The method is covered in the technical report on arXiv. The README also publishes HiDream.ai's own benchmark tables; we don't repeat or verify those numbers here.

Variants and specs

Version	Open-sourced	Steps	Guidance	Best suited to, per HiDream.ai	License
HiDream-O1-Image	May 8, 2026	50	5.0	Recommended for editing tasks	MIT
HiDream-O1-Image-Dev	May 8, 2026	28	0.0	Faster, distilled generation	MIT
HiDream-O1-Image-Dev-2604	May 14, 2026	28	Not listed separately	Text-to-image, with its own prompt refiner	MIT

The README's default output size is 2,048 × 2,048, and requested sizes snap to a set of predefined resolutions such as 2,304 × 1,728 and 2,560 × 1,440.

How to run it locally

These commands come from the official README. It requires a CUDA-capable GPU. Replace /path/to/ with the folder where you downloaded the weights.

git clone https://github.com/HiDream-ai/HiDream-O1-Image.git
cd HiDream-O1-Image
pip install -r requirements.txt

The README highly recommends installing flash-attn. If you can't, it says to edit models/pipeline.py (line 341, or line 291 on the dev branch) and set "use_flash_attn" to False, or inference will fail. It also advises against PyTorch 2.9.x.

Generate an image with the full model:

python inference.py \
    --model_path /path/to/HiDream-O1-Image \
    --prompt "Your prompt here" \
    --output_image results/t2i.png \
    --height 2048 \
    --width 2048

Edit an image:

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

To use the Dev version, point --model_path at the Dev checkpoint and add --model_type dev. Dev-2604 runs from the repository's dev branch, so run git checkout dev before installing.

Prompt Agent

The repository includes a Reasoning-Driven Prompt Agent, prompt_agent.py, which rewrites a raw instruction into a self-contained English prompt. It runs locally with google/gemma-4-31B-it or through an OpenAI-compatible API. Dev-2604 uses HiDream-ai/Prompt-Refine instead.

ComfyUI

ComfyUI's documentation provides native workflows for the Full and Dev versions. If you just want to test the model without installing anything, HiDream.ai also hosts demos on Hugging Face Spaces.

Hardware

HiDream.ai publishes no VRAM figure for O1-Image. The README says only that "a CUDA-capable GPU is required for inference", so treat any specific number you read elsewhere as unofficial.

License

The README states that the code and the HiDream-O1-Image models are under the MIT License. The Hugging Face cards for all three versions also list MIT. Running the Prompt Agent locally means downloading Gemma, so check Google's terms for that model too.

How it relates to the other HiDream models

Model	Main job	Open-sourced	Steps	Resolution	Parameters
HiDream-O1-Image	Generation, editing and personalization	May 2026	50 (Dev 28)	Up to 2,048 × 2,048	8B
HiDream-I1	Text-to-image	April 2025	50 (Dev 28, Fast 16)	1,024 × 1,024 in the official example	17B
HiDream-E1	Image editing	April 2025	28	768 × 768	Built on I1
HiDream-E1.1	Image editing	July 2025	28	Dynamic, about 1 megapixel	Built on I1

The older models rely on separate text encoders and a VAE, while O1-Image's README describes it as working without them.

Try an AI image generator on hidream.org

If you don't want to set up a GPU, you can try our browser tools instead: the AI image generator and the AI photo editor. They use other third-party image models, not HiDream-O1-Image, so results won't match this model's.

Sources

• HiDream-O1-Image on GitHub
• HiDream-O1-Image on Hugging Face
• HiDream-O1-Image-Dev on Hugging Face
• HiDream-O1-Image-Dev-2604 on Hugging Face
• Technical report on arXiv
• ComfyUI's HiDream-O1-Image guide