ComfyUI v0.40.0 Released: Native LoRA Stack and More

ComfyUI v0.40.0 Released: Native LoRA Stack and More

At a Glance

Item Value
Repository Comfy-Org/ComfyUI
Version v0.40.0
Published 2026-10-14
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

The latest version “v0.40.0" of “ComfyUI", a powerful and modular GUI and inference engine for diffusion models utilizing a node-based interface, has been released. ComfyUI is written in Python, licensed under GPL-3.0, and features high-speed local inference.

The most significant impact of this update for users is the addition of the long-awaited native LoRA stack loader node. This makes it easier to build workflows that efficiently manage multiple LoRA models as a standard feature, replacing functionality that previously relied on external custom nodes with official features.

Breaking Changes and Deprecations

This update includes specification changes for some partner nodes and modifications to default system settings. Users running existing workflows should exercise caution.

Old New
Quiver SVG nodes (Originals) Quiver V2 SVG nodes (Old nodes deprecated)
Assets system (Disabled by default) Enabled by default in Windows standalone version / Enabled via --enable-assets elsewhere

Users utilizing Quiver’s SVG-related nodes are recommended to migrate to the new V2 nodes. Additionally, regarding the asset management system, it is now automatically enabled in the Windows standalone version, while other environments require explicitly specifying the enable flag.

Key Changes

Introduction of Native LoRA Stack Loader

The “LoRA stack" functionality, previously achieved by many users via custom nodes, has been implemented as a standard node. The process of batch-managing multiple LoRA models and their strengths and applying them to models is streamlined. This significantly improves the readability and maintainability of workflows for users who heavily rely on complex LoRA combinations.

Enhancement of Asset Management System and 3D Model Support

An asset export API has been implemented locally, and features such as video metadata extraction have been added. Furthermore, outputs from all 3D nodes are now reported as standard 3D output items, advancing the integration of workflows handling 3D models. This system is enabled by default in the Windows standalone version, though it can be disabled using the --disable-assets flag if necessary.

Qwen 3.5 LLM Optimization and Low VRAM Support

Performance in low VRAM environments has been improved for the Qwen 3.5 LLM. Specifically, this includes fixes for MTP (Multi-Token Prediction) embedding offloading and the enablement of image-based MTP. This makes it easier to incorporate more advanced language models into image generation workflows even with limited hardware resources.

Expansion of Partner Nodes

Partner nodes supporting many latest models have been added, such as Google’s “Nano Banana 2.1", Kling’s “Virtual Try-On", and Vidu’s “Q4 models". Additionally, audio- and video-generation-related features such as FishAudio’s “Drama 3" and Higgsfield’s “Genjutsu Motion Transfer" have been enhanced. Utilizing these nodes allows users to operate external services and the latest specialized models directly from ComfyUI.

Greater Flexibility in Model Settings and Attention Control

Support for comfy_attention has been expanded to other models, making kitchen sol attention available via settings within model files. Furthermore, a feature has been added to replace layer-by-layer attention settings per model using LoRA. This is an important improvement for advanced users who want finer control over model behavior.

Expansion of Hardware Support

Portable builds have been added for NVIDIA environments on the ARM64 architecture, facilitating deployment on specific embedded devices and server environments. Additionally, fixes optimized for specific hardware have been made, such as stabilizing SDPA (Scaled Dot Product Attention) operations when data does not fit into the L2 cache on AMD’s gfx1151 environments.

Supported Models and Hardware

This release adds a large number of models newly accessible via partner nodes. However, note that these are not open-weight models where the ComfyUI core loads weights locally, but rather partner integration nodes that call APIs provided by each company.

  • Google “Nano Banana 2.1" model
  • Kling “Virtual Try-On" node
  • Vidu “Q4" model node
  • FishAudio “Drama 3" model
  • Higgsfield “Genjutsu Motion Transfer" node
  • ClarityAI “Crystal Upscaler" node
  • HeyGen “Video 1.0" node 2k resolution support

Regarding locally run models, this release includes operational improvements for the Qwen 3.5 LLM in low VRAM environments, which is directly relevant to users already running the Qwen 3.5 series on ComfyUI.

How to Get It

These release notes do not mention any special migration commands required during an upgrade. Standard updates involve updating the repository via git pull and reinstalling dependencies.

According to the setup instructions in the README, enabling the Manager functionality requires the following commands:

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

Regarding PyTorch versions, while torch 2.7 is minimally supported, PyTorch with cu130 or higher is required for NVIDIA 20 series and newer GPUs. If your current PyTorch installation has not been updated for over 6 months, updating it simultaneously is recommended. As for the Python environment, 3.13 is reported to work stably, while 3.14 may cause issues with certain custom nodes.

For specific instructions on obtaining installers or portable versions, please refer to the release page.

Related Articles

What to Read Next

Sources