repository·main·Indexed 25 days ago
https://github.com/hidream-ai/hidream-i1An open-source 17B parameter image generative foundation model using a Sparse Diffusion Transformer. It includes three model variants—Full, Dev, and Fast—optimized for different balances of quality and speed, and is compatible with the diffusers library and Llama-3.1-8B-Instruct.
HiDream-I1 offers three model variants optimized for different balance points between quality and speed:
| Name | Script | Inference Steps | HuggingFace repo |
|---|---|---|---|
| HiDream-I1-Full | ./inference.py | 50 | HiDream-ai/HiDream-I1-Full |
| HiDream-I1-Dev | ./inference.py | 28 | HiDream-ai/HiDream-I1-Dev |
| HiDream-I1-Fast | ./inference.py | 16 | HiDream-ai/HiDream-I1-Fast |
You can use the inference.py script to generate images using different model versions by specifying the --model_type flag.
Available model types:
full: Uses the HiDream-I1-Full model (50 inference steps).dev: Uses the distilled HiDream-I1-Dev model (28 inference steps).fast: Uses the distilled HiDream-I1-Fast model (16 inference steps).# 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 fastTo set up the environment for HiDream-I1, you must install Flash Attention (recommended CUDA version 12.4) and the project requirements.
Note: The inference script automatically attempts to download meta-llama/Llama-3.1-8B-Instruct. You must agree to the Llama model license on Hugging Face and authenticate using huggingface-cli login before running.
pip install -r requirements.txt
pip install -U flash-attn --no-build-isolationTo launch an interactive web interface for image generation, run the gradio_demo.py script.
python gradio_demo.pyHiDream-I1 is supported by the diffusers library. It is recommended to install diffusers from source for optimal compatibility.
To use the pipeline, you must provide a tokenizer_4 and a text_encoder_4 initialized from the meta-llama/Meta-Llama-3.1-8B-Instruct model.
Parameter Notes:
guidance_scale: Use 5.0 for the full model. For dev and fast models, use 0.0.num_inference_steps: Use 50 for full, 28 for dev, and 16 for fast.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", # "HiDream-ai/HiDream-I1-Dev" | "HiDream-ai/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&Fast
num_inference_steps=50, # 28 for Dev and 16 for Fast
generator=torch.Generator("cuda").manual_seed(0),
).images[0]
image.save("output.png")