This repository provides ComfyUI custom nodes to integrate a variety of advanced image diffusion models, including DiT, PixArt, HunYuanDiT, and MiaoBi, along with support for custom VAEs. It targets users looking to expand ComfyUI's capabilities beyond standard Stable Diffusion models, offering access to cutting-edge research models.

How It Works

The nodes facilitate loading and utilizing models with different architectures and conditioning mechanisms. This includes support for text encoders like Gemma2 and T5, handling compressed latent spaces, and enabling class-label conditioning for DiT. The project aims to provide seamless integration by offering specific loaders and workflow examples for each model type.

Quick Start & Requirements

• Installation: Clone the repository into your ComfyUI custom_nodes folder: git clone https://github.com/city96/ComfyUI_ExtraModels custom_nodes/ComfyUI_ExtraModels. Then, install requirements: pip install -r requirements.txt.
• Prerequisites: Python, ComfyUI, and potentially specific model weights downloaded separately. Some models may benefit from xformers.
• Links: Official Repo

Highlighted Details

• Supports DiT, PixArt (including Sigma and LCM variants), HunYuanDiT, and MiaoBi models.
• Integrates various text encoders like Gemma2 and T5.
• Includes support for custom VAEs, including Consistency Decoders.
• Provides specific nodes for model loading and conditioning.

Maintenance & Community

• The repository is actively maintained by city96.
• Links to relevant forks and original model repositories are provided for troubleshooting and alternative implementations.

Licensing & Compatibility

• The repository itself appears to be under an unspecified license.
• Compatibility with ComfyUI and its ecosystem is the primary focus. Model-specific licenses and compatibility notes are detailed within the README.

Limitations & Caveats

• Some models, like PixArt DPM Sampler, have limitations on prompt length and batch size.
• HunYuanDiT implementation is marked as WIP.
• DiT models are limited to specific resolutions and require specific conditioning inputs.
• Some model implementations may require specific VAEs or optimizations like xformers for optimal performance.