HiDream-I1: 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-I1 is

HiDream-I1 is a text-to-image model that HiDream.ai open-sourced on April 7, 2025. The official README calls it an "open-source image generative foundation model with 17B parameters". The method is described in HiDream.ai's technical report, published in May 2025.

It uses four text encoders, according to ComfyUI's documentation: OpenCLIP ViT-bigG, OpenAI CLIP ViT-L, T5-XXL and Llama 3.1 8B Instruct. Its VAE comes from FLUX.1 [schnell].

Variants and specs

HiDream.ai publishes a full version and two distilled versions:

Version	Steps	Guidance	Type	Hugging Face
HiDream-I1 Full	50	5.0	Full	HiDream-I1-Full
HiDream-I1 Dev	28	0.0	Distilled	HiDream-I1-Dev
HiDream-I1 Fast	16	0.0	Distilled	HiDream-I1-Fast

The official diffusers example generates at 1,024 × 1,024. All three versions were open-sourced together, and each model card lists the MIT License.

How to run it locally

These commands come from the official README. After cloning the repository, install the requirements. The README asks for Flash Attention and recommends CUDA 12.4 for manual installation.

pip install -r requirements.txt
pip install -U flash-attn --no-build-isolation

The script tries to download Llama 3.1 8B Instruct automatically. That model is gated, so accept its license on your Hugging Face account and run huggingface-cli login first. Then choose a version:

python ./inference.py --model_type full
python ./inference.py --model_type dev
python ./inference.py --model_type fast

The repository also includes a Gradio demo, which you start with python gradio_demo.py.

diffusers

HiDream.ai announced official diffusers support on April 11, 2025. This is the README's example, with its comments condensed:

import torch
from transformers import PreTrainedTokenizerFast, LlamaForCausalLM
from diffusers import HiDreamImagePipeline

tokenizer_4 = PreTrainedTokenizerFast.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")
text_encoder_4 = LlamaForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3.1-8B-Instruct",
    output_hidden_states=True,
    output_attentions=True,
    torch_dtype=torch.bfloat16,
)

pipe = HiDreamImagePipeline.from_pretrained(
    "HiDream-ai/HiDream-I1-Full",  # or HiDream-I1-Dev, HiDream-I1-Fast
    tokenizer_4=tokenizer_4,
    text_encoder_4=text_encoder_4,
    torch_dtype=torch.bfloat16,
)
pipe = pipe.to('cuda')

image = pipe(
    'A cat holding a sign that says "HiDream.ai".',
    height=1024,
    width=1024,
    guidance_scale=5.0,  # 0.0 for Dev and Fast
    num_inference_steps=50,  # 28 for Dev, 16 for Fast
    generator=torch.Generator("cuda").manual_seed(0),
).images[0]

Install the current library first:

pip install git+https://github.com/huggingface/diffusers.git

See also the diffusers HiDream documentation.

ComfyUI

ComfyUI's HiDream-I1 guide has native workflows for all three versions. It notes that Dev and Fast don't need negative prompts, so cfg should be set to 1.0.

Hardware

HiDream.ai's repository and model cards give no VRAM requirement. The only figures we found are ComfyUI's: more than 27GB of VRAM to run the full-size files smoothly, and more than 16GB for the fp8 files. Treat these as ComfyUI's guidance, not an official HiDream.ai specification.

License

The README states that the code and the HiDream-I1 models are under the MIT License. The Hugging Face cards add that the MIT License covers the transformer models, and that other components keep their own terms:

• VAE: FLUX.1 [schnell], Apache 2.0
• Text encoder: google/t5-v1_1-xxl, Apache 2.0
• Text encoder: meta-llama/Meta-Llama-3.1-8B-Instruct, Llama 3.1 Community License

The cards also say that you own the content you create with the model. If you're planning commercial use, read the full cards and the Llama license yourself.

How it relates to the other HiDream models

Model	Main job	Open-sourced	Relationship to I1
HiDream-I1	Text-to-image	April 2025	This model
HiDream-E1	Instruction-based editing at 768 × 768	April 2025	Built on I1
HiDream-E1.1	Editing at about 1 megapixel, dynamic resolution	July 2025	Built on I1
HiDream-O1-Image	Generation, editing and personalization up to 2,048 × 2,048	May 2026	Separate 8B model with a unified transformer

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-I1, so results won't match this model's.

Sources

• HiDream-I1 on GitHub
• HiDream-I1-Full on Hugging Face
• HiDream-I1-Dev on Hugging Face
• HiDream-I1-Fast on Hugging Face
• Technical report on arXiv
• diffusers HiDream documentation
• ComfyUI's HiDream-I1 guide