ComfyUI v0.45.0 Released with Qwen 3.x Optimizations

ComfyUI v0.45.0 Released with Qwen 3.x Optimizations

At a Glance

Item Value
Repository Comfy-Org/ComfyUI
Version v0.45.0
Published 2026-11-11
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

Comfy-Org/ComfyUI v0.45.0 has been released. ComfyUI is a GUI, API, and backend that allows users to work with diffusion models using a graph/node-based interface (Python, GPL-3.0).

The change with the biggest impact on users in this version is the Qwen 3.x performance improvement (Part 4). With optimizations that combine parallel Linear weights, users running Qwen 3.x will find the update worthwhile. The location of portable version scripts has also changed, so affected users should check this before updating.

Breaking Changes and Deprecations

Item Old New
Portable version run_cpu script Previous location Under ./advanced
Portable version fast fp16 script Previous location Under ./advanced
local_web_sso Default off (unverified) Default on
partner_node_token Default off (unverified) Default on

The release notes only state “Move… to./advanced" and “Turn on… by default". Details of file name specifics and individual settings could not be confirmed from the documentation.

Users of the portable version utilizing CPU execution or fast fp16 scripts may need to update the paths used for startup. Since the defaults for local_web_sso and partner_node_token have changed, users wanting to keep fixed behavior should check their settings after updating.

Key Changes

Qwen 3.x Performance Improvements and Feature Additions

An optimization combining parallel Linear weights (Part 4) has been introduced. Concurrently, a fix was included to prevent mixing pre-fetched and non-pre-fetched usages in combined Linears. Qwen 3.x also gains video input support, and repetition/presence penalty is now applied during greedy decoding.

This targets users performing text generation with Qwen 3.x. The release notes do not specify whether reconversion or redownloading is required.

Newly Supported Models and Nodes

  • Added support for RelayVSR.
  • Added support for WeVisOne-VL.
  • Added support for MiniMax-H3-image.
  • Added support for texture rebaking and UltraTex.
  • Added solid mode to Extract Mesh from Splat.

All are mentioned only as added support, with no necessary files or settings provided in the documentation.

Addition of Partner Nodes

Partner Nodes, which call cloud APIs, now include Pika (H3 Flex, Soundtrack, SFX, Music), Tencent HY LLM, Kling 4.0 Flash, GPT Image 2.6, GPT Image V3/GIF/SVG, Gemini Omni Flash 1.2, Seedance 2.5 4k resolution, Seedream 5.5, Breeze TTS 2, and others. These are external service-mediated nodes and are not related to local inference.

Memory and Stability Fixes

  • SeedVR2 VAE slice length is now automatically adjusted based on available memory. This is relevant for users prone to running out of memory with SeedVR2.
  • Get Video Components no longer holds decoded frames in duplicate.
  • Error messaging when incorrectly selecting a pixel_space VAE has been improved.
  • Fixed a crash when using the UNetSelfAttentionMultiply node.
  • Fixed an issue causing crashes during text generation with Qwen in fp8.
  • Fixed Remesh Mesh for imported meshes.
  • Fixed megapixel precision in the resolution selector.
  • Fixes for “potential crash" and “potential database issue" are also included, though details are absent from the documentation.

Supported Models and Hardware

In this version, the following models and features are now handleable locally:

  • RelayVSR: Now available on ComfyUI.
  • WeVisOne-VL: Now available on ComfyUI.
  • MiniMax-H3-image: Now available on ComfyUI.
  • UltraTex: Now available in conjunction with texture rebaking.
  • Qwen 3.x: Video input can now be used.

However, the release notes merely state “Support…", omitting parameter counts, licenses, sources for required checkpoints, and VRAM requirements. As these could not be confirmed from the documentation, users should check each model’s distributor before use. New quantization formats and GPU support are not mentioned in this release’s notes.

How to Get It

The documentation’s README does not contain update commands. Notes on the execution environment are located in the manual installation (Windows, Linux) section.

  • Python 3.13 support is the most robust. Version 3.14 also runs, but issues may arise with some custom nodes.
  • torch 2.7 is minimally supported. Newer versions are strongly recommended.
  • Nvidia 20 series and above require pytorch with cu130 or higher.
  • Users with pytorch older than 6 months are advised to update.

The procedure when using manager is as follows:

pip install -r manager_requirements.txt
python main.py --enable-manager

If using the portable version, the locations of the run_cpu and fast fp16 scripts have changed as mentioned earlier. Check whether your startup procedures use these scripts before updating.

Since exact update procedures cannot be reliably determined from the documentation, please refer to the release page.

Related Articles

What to Read Next

Sources