Run HiDream-I1 in ComfyUI with 4-Bit Models (Easy Guide)

Generating high-quality, photorealistic images in ComfyUI just got a serious upgrade with HiDream-I1—an open-source diffusion model built for speed, versatility, and realism. Whether you’re running on a powerful GPU or a lower-end setup, HiDream-I1 performs impressively well, especially when paired with 4-bit quantized weights for memory efficiency.

Thanks to the HiDream-Sampler extension by lum3on, integrating the model into ComfyUI is quick and straightforward. In this post, I’ll walk you through the setup, show how to use HiDream with 4-bit models inside ComfyUI, and share prompt examples to help you get the best results. Note that this is experimental and based on HiDream-Sampler released on 4/12/2025. (Update 4/17/2025: ComfyUI now supports HiDream-I1 natively. Please see this post for a workflow)

Installation

• Update ComfyUI to the latest version. If updating through ComfyUI manager doesn’t work, try running update\update_comfyui.bat script to update it.
• (For Windows) If you ar using Windows, you will have to install triton and SageAttention manually. This is probably the most difficult part of this guide. Please see this post about how to do it manually.
• Drop this picture to your ComfyUI canvas.
  
• Us ComfyUI Manager to install the missing nodes.
• (Update 4/15/2025) The new version of the custom node changed the requirement of triton to exactly 3.0.0. If you follow my guide to install 3.2, the node is not working correctly and you will lose the -nf4 models. Please modify this file \ComfyUI\custom_nodes\hidream_sampler\requirements.txt . Find this line triton==3.0.0 and change it to triton>=3.0.0. Restart ComfyUI a few times to get everything installed.
• Restart ComfyUI.

Nodes

There are two text to image nodes included. This is the advanced one. I disabled this in the workflow, but you are free to try it. The model will be downloaded when it’s specified here for the first time. Note that these are big models, so it takes time to download them. Both regular and the 4 bit versions are included in the drop down box for selection. The regular models use about 27 GB VRAM. The nf4 ones use about 15GB VRAM, but I noticed during running the usage of VRAM spiked to around 23GB VRAM. Therefore, you might still need a GPU with 24GB VRAM. (Update 4/15/2025: I tested this on a laptop with 16GB mobile 4090 and it works fine.) I like the images generated using full-nf4 model. All the examples in this post are generated using the full-nf4 model. In the advanced node, you can specify the prompts for each clip models. The llama_prompt is to instruct the system. You can enter it like “You are a creative AI assistant that helps create detailed, ultra realistic images based on user descriptions.”. You can modify the llama_prompt to fit your need. Users have been reporting that the use_uncensored_llm option might not work, so use it with caution. Currently, there is a limit of 128 tokens for the primary prompt. If you use a long prompt, the prompt will be truncated.

This is the simple version of the HiDream Sampler node.

This is the image to image node. Note that the negative prompt is only for full or full-nf4 models.

Samples

Prompt: ultra-detailed headshot of a beautiful East Asian woman in natural lighting, soft shadows on her face, expressive brown eyes, clean skin texture, black hair tied in a loose bun, minimal makeup, captured with a Leica 50mm lens, photorealism, soft background bokeh

Prompt: traditional Japanese temple surrounded by sakura trees in full bloom, petals falling in the breeze, wooden structure with intricate carvings, glowing lanterns, sunset sky, peaceful mood, high resolution, realistic lighting and textures

Prompt: a mysterious sorceress in a flowing dark robe with glowing blue runes, standing on a cliff with a magic staff, swirling arcane energy around her, night sky with stars, ethereal lighting, fantasy concept art, high detail, realistic style

Prompt: futuristic android woman with sleek silver armor and glowing blue eyes, standing in a cyberpunk lab, surrounded by holographic displays, dark ambient lighting with neon accents, ultra-detailed sci-fi environment, cinematic composition

Prompt: full-body fashion shot of an elegant woman in a red satin dress with high slit, posing in a luxury hotel lobby with marble floors and chandeliers, natural pose, high-end magazine lighting, professional studio photography vibe

Conclusion

The combination of HiDream-I1, ComfyUI, and the HiDream-Sampler extension offers an exciting and efficient way to create high-fidelity images—especially when paired with 4-bit quantized weights for better performance on consumer GPUs. Whether you’re into photorealistic portraits, fantasy concept art, or stylized compositions, HiDream-I1 delivers quality and versatility.

With minimal setup and smart optimizations, this workflow proves that cutting-edge generative models don’t have to be resource-heavy. Give it a try, experiment with your own prompts, and bring your ideas to life with powerful, streamlined control.

Further Reading

Simple ComfyUI Workflow for HiDream-I1 GGUF