README.md

metadata

tags:
  - art
  - image-generation
  - machine-learning
  - transformer

HiDream-I1_HQ-models

This repository collects high-quality models for generating the best images using HiDream-I1.

Left: HiDream-I1-dev, Right: i2i from SDXL Anime Illustration

Model Details

• Transformers (HiDream-I1-Full, Dev, Fast): 1.7 billion parameter transformer-based models using the Rectified Flow Transformer architecture, optimized for photorealistic, cartoon, and artistic image generation.
• Text Encoders: Process input text prompts to guide image generation. Includes Llama-3.1, flan-t5-xxl, and CLIP models.
• VAE: The Variational Autoencoder from FLUX.1 [schnell] for encoding and decoding images.

Model List

Model Name	License	Type	Commercial Use
HiDream-I1-Full-FP16	MIT	Transformer	✅
HiDream-I1-Dev-BF16	MIT	Transformer	✅
HiDream-I1-Fast-BF16	MIT	Transformer	✅
Llama-3.1-8B-Instruct-BF16	Llama 3.1 Community License	Text encoder	✅
flan-t5-xxl_TE-only_FP32	Apache 2.0	Text encoder	✅
CLIP-ViT-bigG-14-laion2B-39B-b160k-FP32	MIT	Text encoder	✅
CLIP-SAE-ViT-L-14-FP32	MIT	Text encoder	✅
FLUX1-schnell-AE-FP32	Apache 2.0	VAE	✅

• The Llama-3.1-8B-Instruct-BF16 model was created by combining the split-distributed original llama-3.1-8b-instruct model (Built with Meta Llama 3).
• If you need text encoders in FP8 format, please refer to the official ComfyUI repository.

Usage Instructions

Place the downloaded model files in following directories.

To use FP32 format text encoders, enabling the --fp32-text-enc setting at startup.

Transformers

• Models

  • HiDream-I1-Full-FP16
  • HiDream-I1-Dev-BF16
  • HiDream-I1-Fast-BF16

• Folder

  • models/StableDiffusion

Text encoders

• Models
  • Llama-3.1-8B-Instruct-BF16
  • flan-t5-xxl_TE_only_FP32
  • CLIP-ViT-bigG-14-laion2B-39B-b160k-FP32
  • CLIP-SAE-ViT-L-14-FP32
• Folder
  • models/text_encoder or models/clip

VAE

• Models

  • FLUX1-schnell-AE-FP32

• Folder

  • models/vae

ComfyUI Workflow

The FP16/BF16 models of HiDream-I1 require a minimum split capacity of 10.1 GB for the transformer in ComfyUI.

With 12GB VRAM and 64GB RAM, you can achieve the highest image quality using the following workflows.

HiDream-I1-Full

HiDream-I1-Dev

HiDream-I1-Fast

For loading text encoders, use the QuadrupleCLIPLoaderMultiGPU custom node from ComfyUI-MultiGPU. Explicitly specify device: cpu to load the text encoder into system RAM.

In my setup, using the Heun sampler and Beta scheduler generated higher-quality illustrations compared to the default settings.

License

This repository inherits the licenses of the original models:

• MIT License: Include the license text in your distribution.
• Apache 2.0 License: Include the license and provide notices.
• Llama 3.1 Community License: Include the statement "Built with Meta Llama 3" and adhere to the usage policy.

See the Model List for links to the full license texts.