HiDream-I1

HiDream-I1 is a 17-billion-parameter text-to-image model open-sourced in April 2025 under the MIT license — free for personal, research and commercial use. It pairs a diffusion transformer with mixture-of-experts routing and leans on four text encoders (including T5-XXL and Llama-3.1-8B) for unusually good handling of complex prompts: colors, counts and spatial relationships come out right more often than with its open-source peers, and Chinese-language prompts work notably well. Three flavors ship — Full (50 steps), Dev (28) and Fast (16) — and ComfyUI workflows are first-class.

The honest picture: the full model is heavy, with VRAM demands that push it onto serious GPUs, and the distilled variants trade quality for speed. Like all open image models it inherits the usual failure modes — garbled text rendering on complex signage, occasional anatomy slips — and the ecosystem is community-driven, so you're relying on forums and Discord rather than a support desk when something breaks.

Clone the HiDream-ai GitHub repo and run inference.py, or use the native ComfyUI HiDream-I1 workflow. Weights download from Hugging Face (HiDream-ai/HiDream-I1-Full).