HiDream-I1

HiDream-I1 generates images from text prompts for creators and developers who want to work on their own hardware or add image generation to an application.

• Image styles: Create photorealistic images, cartoons, paintings and concept art from text descriptions.
• Model variants: Choose the Full model or the distilled Dev and Fast variants, which need fewer generation steps.
• Prompt following: Describe objects, attributes and relationships in a scene; reported DPG-Bench and GenEval results place it above Flux.1-dev and SD3-Medium overall.
• Interactive use and integration: Generate images through a Gradio interface or use Hugging Face Diffusers in a Python application.

Local inference uses CUDA on NVIDIA GPUs and requires Flash Attention. The supplied code downloads Llama 3.1 model files over the internet, though it can use those files from a local cache once downloaded.

The HiDream-I1 model is open source under the MIT license. Its FLUX.1 [schnell] VAE and google/t5-v1_1-xxl text encoder use Apache 2.0; its Llama text encoder uses the Llama 3.1 Community License Agreement. Generated images can be used for personal, research and commercial work, subject to the applicable license terms.