HiDream-O1-Image

HiDream's 8B image model with no separate autoencoder or text encoder: text-to-image, editing and multi-reference scenes. MIT license.

In short. HiDream-O1-Image is an image model that generates pictures from text, edits a photo from an instruction, and puts a person or object from several reference photos into a new scene, all with one set of weights and at up to 2,048 by 2,048 pixels. It is good at long, multi-line text inside images, and the weights are free to use commercially under the MIT license.

Summary

HiDream.ai published HiDream-O1-Image on 8 May 2026, and a technical report, arXiv 2605.11061, followed two days later. The card calls it an 8B model, the checkpoint files hold 8.8 billion parameters, and the config loads it as a Qwen3-VL architecture. Its main design choice is to drop the usual image autoencoder and the separate text encoder. One Pixel-level Unified Transformer reads raw pixels, text and task conditions in a single shared token space, so the same weights cover text-to-image, instruction editing, subject-driven personalization from several reference images, and storyboard generation, up to 2,048 by 2,048 pixels. The card stresses long, multi-region text in several languages, and the inference script adds layout and skeleton conditioning for the personalization task.

There are three checkpoints. The full model runs 50 steps at a guidance scale of 5.0, and it is the one the card recommends for editing. HiDream-O1-Image-Dev is distilled to 28 steps without guidance. HiDream-O1-Image-Dev-2604, added on 14 May, is tuned for text-to-image and comes with a Prompt-Refine model. A built-in prompt agent rewrites a short or reasoning-heavy instruction into a detailed English prompt before generation. It runs on Gemma-4-31B-it locally, which means accepting the Gemma terms, or on any OpenAI-compatible API. Each repository is about 35.2 GB because the weights are stored in FP32, and the script loads them in BF16 by default, so our memory estimate assumes BF16. The card says only that a CUDA GPU is required and recommends flash-attn. Its tables also list a larger HiDream-O1-Image-Pro, with more than 200 billion parameters, for which we found no public weights.

The code and the models are under the MIT license. HiDream's own tables give HiDream-O1-Image 0.90 on GenEval, against 0.87 for Qwen-Image and FLUX.2 [dev], and 0.913 on CVTG-2K complex text rendering, against 0.893 for FLUX.2 [dev]. The card also says Dev-2604 debuted at number 8 on the Artificial Analysis Text to Image Arena, a snapshot from its release. All of these results are the developer's. In the 30 days to 11 November 2026, Hugging Face counted about 15,000 downloads of the full model and a few hundred each for the two Dev builds.

Good for

  • One open model for text-to-image, instruction editing and multi-reference personalization at up to 2,048 by 2,048
  • Images with long, multi-region text, including Chinese and other languages
  • Teams that want MIT-licensed weights they can fine-tune, host and sell the outputs of
  • Prompts that need reasoning, through the built-in prompt agent

Watch out for

  • The repository is 35.2 GB because of FP32 storage, the card gives no VRAM figure, and a CUDA GPU is required
  • The best prompt-agent results need Gemma-4-31B-it, which has its own terms, or a hosted OpenAI-compatible model
  • For editing the card recommends the full model rather than a Dev build
  • The card's tables include a 200B-plus Pro model, and we found no public weights for it
  • The benchmark and arena figures are HiDream's own

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of HiDream-O1-Image. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
HiDream-O1-Image (full, 50 steps) 8.8B – 20 GB Gaming GPU
HiDream-O1-Image-Dev (distilled, 28 steps) 8.8B – 20 GB Gaming GPU
HiDream-O1-Image-Dev-2604 (28 steps, text-to-image) 8.8B – 20 GB Gaming GPU

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.

More about MIT

See also

Frequently asked questions

What is HiDream-O1-Image?
HiDream-O1-Image is an image model that generates pictures from text, edits a photo from an instruction, and puts a person or object from several reference photos into a new scene, all with one set of weights and at up to 2,048 by 2,048 pixels. It is good at long, multi-line text inside images, and the weights are free to use commercially under the MIT license.
Can I use HiDream-O1-Image commercially?
Yes. HiDream-O1-Image is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is HiDream-O1-Image open source?
HiDream-O1-Image is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run HiDream-O1-Image on my own computer?
Yes, on a good graphics card. The smallest variant, HiDream-O1-Image (full, 50 steps), needs about 20 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was HiDream-O1-Image released?
HiDream-O1-Image was released by HiDream.ai on May 8, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen-Image-2.1 Alibaba Qwen 7.1B – Qwen Research License (Non-commercial) 20 GB Gaming GPU
Z-Image Tongyi MAI 6.15B – Apache 2.0 (Permissive) 16 GB Laptop or phone
FLUX 3 Image Klein Black Forest Labs 5.6B – FLUX Kommunity (Conditional) 8 GB Laptop or phone
Krea 2 Krea 12.8B – Krea 2 Community (Conditional) 41 GB Workstation