TECHNOLOGY

ComfyUI v0.21.0 Adds Joint Audio-Video Loading, Fixes Model Crash

ComfyUI v0.21.0 Adds Joint Audio-Video Loading, Fixes Model Crash

ComfyUI v0.21.0 is out, and it brings a practical upgrade for anyone working with video generation. The headline feature? The video loader node can now read audio and video at the same time. That means no more separate audio tracks to wrangle — the node handles both streams in one go. It's a small change that saves real time.

What Changed

The core update is in PR #13591: the video loader node now reads audio and video simultaneously. Previously, you'd have to load them separately or rely on external tools. That's gone. ComfyUI's dev team also squashed a nasty bug in models with un-initialized meta-tensors, which was causing CPU TensorRT engine crashes (CORE-67). That fix landed in PR #13578. A third change alters the "save 3d model" node's filename prefix to 3d/ComfyUI (CORE-106) — a tiny but welcome organizational tweak.

Why It Matters

For video creators using ComfyUI, this is a workflow smoother. Combining audio and video loading cuts steps and reduces the chance of sync errors. The crash fix is critical for anyone on CPU — TE models were randomly dying, and that's a stability nightmare. The 3D model prefix change? It keeps your exports tidy. None of these are flashy, but they're the kind of polish that separates a hobby tool from a production-ready one. ComfyUI keeps maturing, and v0.21.0 proves the team is listening to real-world pain points.

Official Source: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.21.0