ComfyUI 0.36.0: Generic Loops, a 4 TB AMD Ceiling Removal, and a Local Audio-Depth Pipeline

What happened

ComfyUI v0.36.0 shipped on 15 September 2026, and it is the densest feature drop in a while: Generic Loops as a core execution primitive, Marigold v2 depth estimation, Yue2 music generation, a 4 TB virtual-address quota for Windows AMD, and a Video Concatenate node. The partner-node surface grew across OpenRouter, Gemini, Bria, BFL, Pruna, and Tripo in the same release.

Context

ComfyUI has been the default local-node pipeline for image, video, and increasingly audio work. v0.35.0 landed the comfy-kitchen dependency chain and the assets-system skeleton that this release builds on. What is not new is the pattern: a point release that touches every layer simultaneously — execution engine, model support, partner integrations, frontend — rather than shipping a single headline feature. For a studio that runs ComfyUI on owned hardware, the question is never "does this work" but "what breaks and what finally stops being a workaround."

How it works

Three items change the execution layer rather than just adding a node. Generic Loops (CORE-14) introduces a first-class iteration primitive: loop boundaries are declared in node schema, the executionList input is marked strictly internal, and a heterogeneous list is carried through each iteration without type coercion. This replaces the manual subgraph-duplication pattern. The AMD Windows virtual-address quota was raised to 4 TB, letting the runtime reserve more space for model offloading on Windows + Radeon. On the inference side, comfy-kitchen's apply-rope now uses fast kernels on more llama-based models, and MiniMax-H3 VAE memory usage was reduced in two separate PRs. Neither is visible in the UI; both lower peak VRAM and per-token latency. Marigold v2 adds local metric depth-map generation. Yue2 adds music generation with an increased maximum duration and AMD-specific fixes. The Video Concatenate node (CORE-436) splices generated clips without an external mux step.

Our read

Generic Loops is not a node; it is a change to the execution graph's type system. By declaring loop boundaries in schema and marking executionList as internal, ComfyUI is moving from a flat DAG to a graph with scoped control flow. Every partner-node author will eventually have to say whether their outputs are loop-safe. That is a different architectural statement than "you can now loop over a list."

The 4 TB VA quota is the quietest item and probably the most consequential for a Windows AMD studio. It is a ceiling removal, not a performance number. A Radeon GPU can hold more model weights in virtual memory before the OS kills the process, meaning fewer OOM restarts on long video jobs. No launch post headlines this.

The assets subsystem tells its own story. Records were split from content, structured event logging was added, and then the --enable-assets test flag was disabled in the same cycle. The subsystem is not ready. Wiring it into production this month means you are testing ComfyUI's QA, not shipping work.

Yue2 and Marigold v2 together produce a local audio-plus-depth pipeline. That combination matters more than either model alone, and neither requires an API key.

What this changes

Monday: if you run Windows + Radeon, update and verify offloading no longer hits the old VA ceiling. If your workflow iterates over a folder of images, try Generic Loops in place of the manual-repeat subgraph; check that third-party nodes still resolve inputs inside a loop. If you composite over transparency, the Add Noise alpha fix removes a subtle corruption path. If you generate short music beds locally, Yue2 is available, but the release notes do not state whether weights ship with ComfyUI or must be sourced separately. If you use --enable-assets, stop until the flag is re-enabled. Partner-node additions (Pruna, BFL, Bria) are cloud off-ramps requiring API keys; they do not change local topology.

License

The release notes do not state or reaffirm a software licence. The upstream ComfyUI repository has previously carried a custom licence, but this changelog neither confirms nor changes it. Check the repository's LICENSE file before building anything commercial on this release.

Key takeaways

• Generic Loops changes the execution graph from a flat DAG to scoped control flow, which will eventually force third-party nodes to declare loop-safety.
• The 4 TB AMD VA quota is a ceiling removal, not a speedup; it reduces OOM restarts on Windows + Radeon video jobs.
• Marigold v2 plus Yue2 create a local depth-plus-music pipeline with no external API dependency, though weight sourcing for both is unconfirmed in the release notes.
• The assets subsystem shipped features and then disabled its own test flag in the same cycle; treat it as experimental.
• The llama fast-kernel rope and MiniMax-H3 VAE memory cuts are invisible in the UI but directly lower VRAM and latency for text-guided and LLM-assisted workflows.

Sources

1. v0.36.0 — tier 1