EricRollei Brings To Comfy_HunyuanImage3 Hunyuan Image 3.0

By vramkickedin | | 2 min read
A close up of an artists pallet with the text engraved Comfy_HunyuanImage3

Comfy_HunyuanImage3 is a set of ComfyUI custom nodes for a collection of quantized versions of the HunyuanImage-3.0 image model. The integration provides professional tools for text-to-image generation, image editing, and multi-image fusion directly within ComfyUI workflows.

EricRollei developed these nodes to make the massive HunyuanImage-3.0 model accessible to users with consumer-grade GPUs. Through quantization options and smart memory management, users with 24GB VRAM cards can run this model locally. The latest v1.3.0 update fixes critical bugs affecting INT8 model handling.

Model Size: 45GB (NF4) to 160GB (BF16) & VRAM GPU: 24GB minimum

What these image models and custom nodes offer

  • Built-in Chain-of-Thought prompt enhancement for Instruct models without external APIs.
  • Multi-image fusion combining 2-5 reference images into new compositions.
  • Image-to-image editing using natural language instructions.
  • High-resolution generation up to 4K on supported hardware.
  • NF4 and INT8 quantized models for reduced memory usage.
  • Smart memory management with automatic VRAM tracking and cleanup.
  • Nodes: Model loaders and Utilities .

Creative professionals and hobbyists working with limited hardware can generate high-quality images without cloud services. The quantized NF4 model runs on 24GB VRAM cards, making professional image generation practical on home workstations.

Development notes and Comfy_HunyuanImage3 Requirements

Version 1.3.0 addresses five bugs impacting INT8 models, including block swap issues, memory estimation errors, and VAE decode crashes. The release notes state that:

'all Instruct INT8 variants (Distil and full) are fully operational'

after these fixes. Users on Windows should be aware that INT8 block swap may encounter memory issues due to how the operating system handles pinned CPU memory. For optimal performance with INT8 models on 96GB GPUs, the documentation recommends setting blocks_to_swap=0.

This integration makes an 80B parameter model practical for local use through smart quantization and memory management.

Get Comfy_HunyuanImage3 on GitHub. Download pre-quantized models on Hugging Face.