ComfyUI v0.35.0 Released with Comfy Compiler and Memory Fixes

September 10, 2026September 20, 2026

At a Glance

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

Overview

ComfyUI v0.35.0 has been released. ComfyUI is an extremely powerful and modular diffusion model GUI, API, and backend featuring a graph/node interface.

The most important changes in this update are improvements to memory management in containerized environments and the introduction of the new “Comfy Compiler". These are expected to improve stability in resource-constrained environments and optimize inference.

Key Changes

Improved Memory Management in Container Environments

It has been fixed to respect container cgroup memory limits. Previously, it operated based on the host’s RAM, but it will now operate based on the memory limits allocated to the container. For users running ComfyUI using container technologies like Docker, this is an important improvement to prevent crashes due to out-of-memory errors.

Introduction of Comfy Compiler

The new “Comfy Compiler" has been introduced. This is a feature aimed at optimizing inference. While it currently includes processing to temporarily pause the compiler for certain long-lived sparse allocations, it serves as a foundation for future optimizations.

Support for New Models and Features

Many new model architecture types and features have been added. In particular, support for the WAN3-Prime model has been added, alongside various extensions for MiniMax-H3 (such as PDD LoRA, Max model, controlnet patching, and VAE optionization). Support through partner nodes has also been expanded, including Recraft V4 Styles, Google-Omni 1.1, SenseNova U1.5, Meta’s Muse Image, and Anthropic’s Claude Fable 5.1.

Enhanced 3D Features

The File3DToMesh node has been added, allowing GLB, GLTF, OBJ, and STL files to be parsed into MESH. Additionally, temporary path annotations are now made in outputs such as Preview3DAdvanced.

Improvements to Video Editing and Image Output

VideoTrim and VideoCrop nodes have been added, making the VIDEO_EDIT widget input available. Furthermore, the avif format is now supported in the Save Image Advanced node, alongside video-related fixes such as remuxing HEVC as mp4/mov in hvc1 format.

Optimizations for AMD GPUs

PyTorch SDPA attention and FP8 ops have been enabled for AMD’s gfx1170 and gfx1171 devices. This change is expected to improve performance for users utilizing AMD hardware.

Supported Models and Hardware

• Models: WAN3-Prime, Google-Omni 1.1, SenseNova U1.5, MiniMax-H3 (with Max/Max Turbo/PDD LoRA support), Recraft V4 Styles, Meta Muse Image, Anthropic Claude Fable 5.1, OpenAI GPT 6 Astra, Pixal3D multiview model, and more.
• Hardware: AMD gfx1170 and gfx1171 (with PyTorch SDPA/FP8 support).
• Formats: avif (Save Image Advanced), GLB/GLTF/OBJ/STL (3D mesh conversion).

How to Update

If you have installed it manually, update the repository to the latest state.

git pull

Also, if you have installed manager dependencies, run the following command as needed.

pip install -r manager_requirements.txt

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.

Sources

• comfyanonymous/ComfyUI v0.35.0

Update History

• 2026-09-19: Rewrote the article from re-collected sources and restored it from draft to published.