ComfyUI v0.38.0 Released: Hunyuan Image 3.5 and More

ComfyUI v0.38.0 Released: Hunyuan Image 3.5 and More

At a Glance

Item Value
Repository Comfy-Org/ComfyUI
Version v0.38.0
Published 2026-09-30
License GPL-3.0
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

ComfyUI, a node-based GUI and inference engine for diffusion models, has released v0.38.0. ComfyUI is a tool built strongly for local inference (written in Python, GPL-3.0 license) that allows users to assemble and run image and video generation models like Stable Diffusion using a graph/node interface.

The most impactful change for users in this release is the expansion of supported models. Alongside the newly added nodes for Tencent’s Hunyuan Image 3.5, support for ming-image models and functional enhancements related to Qwen-Image 2.1 make this an update worth applying for those wanting to try new image generation models. Additionally, the torchaudio dependency has been removed, making the installation slightly lighter.

Key Changes

Addition of Hunyuan Image 3.5 Nodes: Text-to-image and editing nodes supporting Tencent’s Hunyuan Image 3.5 model have been added as partner nodes. This is intended for users who want to try Hunyuan-series text-to-image and image editing on ComfyUI, and after updating, the new nodes can be selected from the node list.

Support for ming-image Models: Support for ming-image models has been newly added. This is a feature addition for users wanting to run these models, alongside fixes related to detection during saving and loading.

Feature Enhancements for Qwen-Image 2.1: Support for tiny VAE and union fun controlnet has been added for Qwen-Image 2.1. Improvements to KV cache placement logic and transformer block compilation support have also been made, meaning users utilizing Qwen-Image 2.1 may benefit from reduced memory usage and faster speeds.

MiniMax-H3 Fun-Controlnet-Union 2.0 Support: Fun-Controlnet-Union 2.0 is now supported for the MiniMax-H3 model. In addition, a bug where VAE rms_rope crashed with offloaded qk_norm_scale has been fixed. This is an update for users employing MiniMax-H3.

ID-V2V (Wan 2.1/VACE) Support: ID-V2V (Image-to-Video conversion systems) powered by Wan 2.1/VACE-based models is now supported. This targets users of Wan-series video generation models.

Support for w6a8 Quantization Format: Support for the w6a8 quantization format has been newly added, broadening the choices for users working with quantized models.

Removal of torchaudio Dependency: The dependency on torchaudio has been removed from the main ComfyUI package. Since this reduces the dependent packages during installation, it makes setup slightly lighter for those doing fresh installations or running Docker.

Long-Context Acceleration for qwen3.5/3.8: Generation speed in long contexts has been improved. This is aimed at users running qwen3.5/3.8 series models with long prompts.

Seedvr2 Optimization: Seedvr2 processing has been optimized, delivering performance improvements for Seedvr2 users.

Addition of RDNA2 Architecture: The AMD GPU RDNA2 architecture has been added to the list, which is relevant for users running ComfyUI on matching AMD GPUs.

RGBA Image Crash Fix: A bug causing crashes when 4-channel (RGBA) images were passed to the ImageUpscaleWithModel node and Qwen VL image preprocessing has been fixed. Users inputting images with transparency information into workflows will see this error resolved after updating.

Supported Models and Hardware

This version adds support for the following models and formats:

  • Hunyuan Image 3.5 (Tencent): Text-to-image and editing nodes added, enabling text-to-image and image editing on ComfyUI
  • ming-image: Added support as a new model
  • Qwen-Image 2.1: Supported tiny VAE and union fun controlnet
  • MiniMax-H3: Supported Fun-Controlnet-Union 2.0
  • Wan 2.1/VACE-based models: Supported ID-V2V (Image-to-Video conversion)
  • Supported w6a8 quantization format
  • Added AMD GPU RDNA2 architecture to the list

As this is an addition of new model architectures and quantization formats themselves, there is no mention of existing workflows or checkpoints needing re-conversion. To try the relevant models, corresponding nodes and checkpoints must be prepared separately.

How to Get It

The documentation outlines manual installation steps for Windows/Linux. Python 3.13 is stated to be “very well supported", and the use of torch 2.7 or higher (as new a version as possible) is recommended. For Nvidia 20 series and above, PyTorch with cu130 or higher is required.

When using Manager, follow these steps:

pip install -r manager_requirements.txt

python main.py --enable-manager

Specific commands for updating an existing environment to v0.38.0 (such as git pull) are not mentioned in the documentation, so please check the release page for details.

Related Articles

What to Read Next

Sources

Engines and ToolsComfyUI,GUI,Hunyuan,MiniMax,Python,Qwen-Image,Verified