ComfyUI v0.36.0 Released: Generic Loops, Yue2 & Marigold v2

September 16, 2026September 18, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

comfyanonymous/ComfyUI, a powerful and modular Python-based GUI, API, and backend for diffusion models featuring a graph- and node-based interface, has released its latest version, “v0.36.0".

This version introduces many new features and optimizations, such as the implementation of Generic Loops, support for the music generation model “Yue2" and “Marigold v2", and faster RoPE processing for Llama-based models. In addition, various improvements in AMD environments and the addition/update of numerous partner nodes make this an update that further expands the possibilities of using generative AI locally.

Key Changes

The main changes affecting local LLM, image, and video generation users are as follows:

• Implementation of Generic Loops: “Generic Loops" (CORE-14) have been implemented, enabling loop processing within workflows. Different items in a list can now be processed within a loop.
• Support for Yue2 Music Model: Newly added support for the Yue2 music generation model. Maximum song length in nodes has been increased, issues in AMD environments have been fixed, and controls have been added to the ABC node for generation.
• Support for Marigold v2: Added support and blueprints for the Marigold v2 depth estimation model.
• Speedups for Llama-based Models: Llama-based models are now updated to use “apply rope" from comfy-kitchen (updated to v0.2.34), allowing faster kernels to be applied to more models.
• Addition of Video Concatenate Node: Newly implemented the “Video Concatenate" node (CORE-436) to combine videos.
• MiniMax-H3 VAE Optimization: Memory usage for the MiniMax-H3 VAE has been reduced and optimized. Fixes for h3 fun controlnet by the Comfy compiler are also included.
• AMD GPU Environment Improvements: The VA quota in Windows AMD environments has been raised to 4TB, and manual installation instructions have been updated.
• Partner Node Updates:
• BFL: Added the “Flux Video Edit" node. – Gemini-LLM: Added “Gemini 3.8 Flash" to text nodes and newly introduced “GeminiNodeV3" (V2 is deprecated). – OpenRouter: Added support for Microsoft’s “mai-image-2.6" model and introduced an auto aspect ratio feature for the MAI image node. – Tripo: Migrated to v3 API, added the Smart Segment node, and added “Tripo P2" (text, image, multiview) model nodes. – Bria: Added new image editing nodes and a “Video Eraser" node. – Pruna: Added text-to-video and image-to-video nodes using “P-Video-2".
• Other Adjustments: The Add Noise to Image node was changed so that noise is no longer added to the alpha channel. Additionally, linear was added to ImageColorSpace. The frontend package has been updated to v1.52.7 and workflow templates to v0.11.62.

Supported Models and Hardware

Models and hardware newly supported or optimized in this version are as follows:

• Supported Models:
• Yue2 music generation model – Marigold v2 depth estimation model – Llama-based LLMs (faster RoPE processing) – MiniMax-H3 (reduced VAE memory usage) – Microsoft mai-image-2.6 (via OpenRouter) – Gemini 3.8 Flash (via Gemini-LLM) – Tripo P2
• Supported Hardware and Environments:
• AMD GPUs (VA quota increased to 4TB in Windows environments, Yue2 operational bug fixes)

Hardware Requirements

Recommended environments and setup procedures for manual installation on Windows and Linux environments are as follows:

Recommended Environment

• Python: Python 3.13 is strongly recommended (Python 3.14 works, but some custom nodes may have issues; Python 3.12 is also usable).
• PyTorch: PyTorch 2.7 or higher is the minimum requirement, but the latest major version combined with the latest CUDA version is strongly recommended. Using NVIDIA 20 series or higher GPUs requires PyTorch with cu130 or later.

Setup Commands

1. Install dependencies:

pip install -r manager_requirements.txt

2. Command to start ComfyUI with the manager enabled:

python main.py --enable-manager

Releases Since Our Last Article

Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.

Related Articles

• Comfy-Org Releases YuE2 for ComfyUI: Open Music Generation

Sources

• comfyanonymous/ComfyUI v0.36.0 Release Notes