ComfyUI v0.39.0 Released with New CLI Flags and Partner Nodes

ComfyUI v0.39.0 Released with New CLI Flags and Partner Nodes

At a Glance

Item Value
Repository Comfy-Org/ComfyUI
Version v0.39.0
Published 2026-10-06
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 version v0.39.0 has been released. ComfyUI is a powerful, modular GUI, API, and backend for diffusion models developed in Python featuring a graph and node-based interface, and is considered one of the world’s fastest local inference engines. The most significant changes in this update include the addition of new startup options to control offline environments and external integration features, an expansion of various partner nodes, and automatic optimization for integrated GPU environments. These changes allow users to run and manage the software more flexibly and stably depending on their execution environment and use case.

Breaking Changes and Deprecations

In this version, some startup options have been deprecated and replaced with new ones.

Old Option New Option
--disable-api-nodes --offline or --disable-partner-nodes

Affected users: This affects users who previously used the --disable-api-nodes option to disable external API nodes. Going forward, please transition to the newly added --offline or --disable-partner-nodes options.

Additionally, among partner nodes, Luma’s “Ray 2" node has been deprecated. Workflows using these may be affected in future updates.

Key Changes

Addition of New Startup Options and Deprecations

To restrict connections to external networks and the loading of partner nodes, the new --offline and --disable-partner-nodes options have been added. These options are specified at startup when running in offline environments or when you do not wish to load specific partner nodes. Along with this, the legacy --disable-api-nodes has been deprecated.

Automatic Memory Optimization on Integrated GPUs

In environments using integrated GPUs, pinned memory is now automatically disabled. This is expected to improve stability and memory efficiency on integrated GPU environments that lack dedicated video memory. Users in target environments can benefit from this optimization simply by updating, without requiring any special configuration changes.

Improved Save Quality and 16-Bit Support

Default settings for encoding in the Save Video node now apply higher-quality defaults. Additionally, 16-bit float support has been added for Save EXR, and the default bit depth has also been updated to 16-bit float. This improves convenience for users prioritizing image quality or performing downstream compositing work.

Significant Expansion of Partner Nodes

Numerous partner nodes for integrating with external services and latest models have been added and updated. Specifically, nodes supporting models such as Anthropic’s “Sonnet 5.5", Ideogram’s “Ideogram 4.5", HeyGen’s “HeyGen Video 1.0", Black Forest Labs (BFL)’s “FLUX 3 Image", Grok’s “grok-imagine-video-1.5-lite", and ElevenLabs’ “Eleven v4" and “v4 Turbo" have been added. This is an important update for users wanting to use these models directly within ComfyUI.

Asset Scan Stability and Performance Improvements

Fixes to prevent database lock contention during model folder loading and asset scanning, as well as optimizations for pause and resume control of the scanning process, have been implemented. This improves stability in environments managing a large number of model files and during file loading at startup. Furthermore, CPU usage and pause durations during scans are now reported, making troubleshooting easier.

Bug Fixes and Support for Various Models

An issue causing crashes when selecting int8 or int4 caches with the Qwen image 2.1 model has been fixed. Furthermore, support for LynnReal light Minimax-H3 VAE has been added, and issues related to Minimax VAE offloading and temporary file release processing have been improved. In addition, supported models for comfy_attention and AttentionTensorContainer have been expanded, aiming to improve inference efficiency.

Supported Models and Hardware

This version adds numerous new model architectures and integration nodes for external services (partner nodes), greatly expanding the choices of available models. It also improves operational stability on specific hardware environments.

Newly Supported Models and Fixes

  • LynnReal light Minimax-H3 VAE Support: “LynnReal light Minimax-H3 VAE" is now supported. Along with this, issues regarding Minimax VAE offloading processing have been fixed, and optimizations to release MiniMax H3 embedded temporary files before blocks have been implemented to improve memory efficiency.
  • Qwen image 2.1 Cache Crash Fix: An issue where crashes occurred when selecting int8 or int4 cache in the Qwen image 2.1 model has been resolved, enabling stable inference using low-bit caches.
  • Addition of Various Partner Nodes: Numerous nodes for integrating with external services and latest models have been added. Specifically, nodes supporting models such as Anthropic’s “Sonnet 5.5", Ideogram’s “Ideogram 4.5", HeyGen’s “HeyGen Video 1.0", Black Forest Labs (BFL)’s “FLUX 3 Image", Grok’s “grok-imagine-video-1.5-lite", and ElevenLabs’ “Eleven v4" and “v4 Turbo" have been added. Additionally, an issue where backgrounds became transparent in OpenAI’s “GPT Image 2" has been fixed.

Hardware-Specific Optimizations

  • Automatic Pinned Memory Disabling on Integrated GPUs: Pinned memory is now automatically disabled in environments using integrated GPUs. This is expected to improve stability and memory efficiency in environments lacking dedicated video memory.

How to Get It

When updating ComfyUI or performing a fresh manual installation, please follow the recommended environments and procedures below.

Recommended Environment

  • Python Version:
  • Python 3.13 is very well supported. If you encounter issues with custom node dependencies, trying Python 3.12 is recommended. – It also runs on Python 3.14, but some custom nodes may experience issues. The free-threaded variant also works, but since the GIL remains enabled for some dependencies, it is not fully supported.
  • PyTorch Version:
  • PyTorch 2.7 has minimal support, but using a newer version is strongly recommended. – If you are using NVIDIA 20 series or newer GPUs, PyTorch version cu130 or later is required. Generally, it is recommended to use the latest major version of PyTorch corresponding to the latest CUDA version, excluding the newest releases under two weeks old. If your PyTorch is older than 6 months, please update.

Setup Procedure

  1. Run the following command to install the required manager dependencies:
pip install -r manager_requirements.txt
  1. Run ComfyUI with the --enable-manager flag at startup to enable the manager:
python main.py --enable-manager

Enabling TLS/SSL (Optional)

To conduct HTTPS communication in secure environments, you can generate and use a self-signed certificate with the following steps:

  1. Run the following command to generate a certificate (cert.pem) and private key (key.pem):
openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -sha256 -days 3650 -nodes -subj "/C=XX/ST=StateName/L=CityName/O=CompanyName/OU=CompanySectionName/CN=CommonNameOrHostname"

Note for Windows users: You can use alexisrolland/docker-openssl or third-party binary distributions to run this. If using a container, relative path volume mounts (e.g., .\\:/openssl-certs) can be used to generate files in the current directory.

  1. Run with the following options added at startup. This allows access via https://… instead of http://…:
python main.py --tls-keyfile key.pem --tls-certfile cert.pem

Related Articles

What to Read Next

Sources