HiDream-I1-Full

HiDream-I1-Full

Generates images from text prompts with a 17B-parameter open model that scores at the top of the card's prompt-following tables. For designers, marketers and developers who have a capable GPU.

Made by HiDream-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: HiDream-I1-Full is an open text-to-image model with 17.1 billion parameters, released under the MIT license. You describe a picture in words and it draws it, in styles from photorealistic to cartoon. The repository also points to faster distilled variants, called dev and fast.

Who it suits: It suits people who want to generate images on their own hardware and are comfortable running Python scripts and installing GPU libraries.

Industry | Who uses it | How they use it | Example
Marketing | Designers and content teams | Draft campaign visuals from a written brief | A product scene for a banner
Software | App developers | Add image generation to a product | Generating placeholder art for a game
Research | Image model researchers | Compare prompt following against other open models | Running GenEval-style prompts
Creative work | Illustrators | Explore styles before painting by hand | A cartoon and a photoreal version of one idea

Skip it if you have no strong GPU, since the F16 weights alone are about 34.2 GB.

Where to use it:

• Python scripts run from the project's code, on Windows, macOS or Linux
• A Gradio web demo that you start on your own machine
• The diffusers library, which the model is listed for

Use case | Who | What they get back
Text to image | Designers | A picture generated from your prompt
Speed trade-offs | Developers | The choice of full, dev or fast variants
Interactive tries | Anyone testing prompts | A web page where you type prompts and see results

How to set it up: You need these first.

• Python with pip, and a GPU with plenty of memory
• Flash Attention installed, where the card recommends CUDA 12.4 for a manual install
• The project's code from its GitHub repository, cloned to your machine
• Network access on the first run, because the script fetches Meta's Llama 3.1 8B Instruct files

1. Clone the project's GitHub repository.
2. Install its requirements with pip.
3. Optionally fetch the HiDream-I1-Full files with the Hugging Face tools.
4. Run the inference script for the variant you want.

pip install -r requirements.txt

pip install -U huggingface_hub
hf download HiDream-ai/HiDream-I1-Full

How to use it: Run the inference script with the variant you want, or start the Gradio demo for interactive use. The first run pulls the Llama model files on its own, so a blocked connection stops it before it starts.

# For full model inference
python ./inference.py --model_type full

# For distilled dev model inference
python ./inference.py --model_type dev

# For distilled fast model inference
python ./inference.py --model_type fast

python gradio_demo.py

Tool | What it does
inference.py --model_type full | Runs the full model
inference.py --model_type dev | Runs the distilled dev model
inference.py --model_type fast | Runs the distilled fast model
gradio_demo.py | Starts the interactive Gradio demo

Example: Start the Gradio demo and type a prompt such as "a watercolor fox in a snowy forest". The model generates an image from that text and shows it on the page, ready to save.

For newcomers and experts: If you are new, start with the Gradio demo, since you can type a prompt and see an image without editing code. Expect a big download and a long first run, because the Llama files come along too.

If you know the field, the technical report names it a sparse diffusion transformer. The card ranks it first in the tables shown for GenEval and DPG-Bench, with an overall GenEval score of 0.83 and 85.89 on DPG-Bench. These are the authors' own results, and the same table shows it behind some models on single categories such as the global score.

You write a text prompt
Script loads the HiDream-I1 weights
Llama 3.1 8B Instruct files are fetched
Chosen variant generates the image
Picture saved or shown in the demo

Good to know:

• Size: 17.1B parameters in Safetensors, about 34.2 GB in F16 and about 8.6 GB at 4-bit by estimate
• Other builds: other people have made 5 quantized and 13 fine-tuned versions
• Editing: the authors released separate image-editing models, HiDream-E1-Full and HiDream-E1.1
• Commercial use: the card says generated images can be used in personal, research and commercial work

Title: GenEval overall score, from the card's table
HiDream-I1: 0.83 score
Janus-Pro-7B: 0.80 score
SD3-Medium: 0.74 score
CogView4-6B: 0.73 score
DALL-E 3: 0.67 score
Flux.1-dev: 0.66 score
SDXL: 0.55 score
PixArt-alpha: 0.48 score

• License: MIT
• Cost to run: none named, the source lists no API key or paid service, though the GPU needs are large
• Data it reaches: your prompts and generated images on your own machine, plus a first-run download of Llama 3.1 8B Instruct files
• Not for: machines without a strong GPU, since the card gives no CPU-only route

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Parameters: 17.1B
• Weights: about 34.2 GB in F16; about 8.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 695
• Likes: 1,015
• License: mit
• Languages: en
• Quantized versions: 5
• Fine-tunes: 13
• Tags: diffusers, safetensors, image-generation, hidream.ai, text-to-image

Links

• Model: https://huggingface.co/HiDream-ai/HiDream-I1-Full
• Files: https://huggingface.co/HiDream-ai/HiDream-I1-Full/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download HiDream-ai/HiDream-I1-Full

Then follow the usage in its model card. Nothing in this kit runs by itself.