Checked 27 September 2026 · ComfyUI H3 VAE
ComfyUI cut H3 VAE time. That does not make the whole render twice as fast.
Comfy Org says: its H3 video VAE encode/decode round trip fell from 24.3 to 12.7 seconds on an RTX 5090 at 1344×768 and 129 frames. Those are VAE-node times. The post does not give a complete H3 generation time.
What changed
The reported A/B saves 11.6 seconds in that stage. Comfy attributes the gain to a fused encoder, an optional faster accumulation path, and an int8 decoder option. It says to use ComfyUI v0.36.0 or later. The default fused encoder needs no extra flag; the faster encoder path needs --fast fp16_accumulation, and the fastest decode needs the int8 H3 VAE file.
The source reports small quality differences for the faster paths. The pool has not inspected matched outputs or repeated the timing. No result here measures denoising, model loading, queue time, MP4 delivery, or what a viewer sees.
Before you repeat the headline
Record the ComfyUI version, VAE file, flag, GPU, resolution, frame count, and time boundary. If your job spends minutes denoising, an 11.6-second VAE saving cannot be presented as a twofold full-render gain. Time the whole job before estimating usable output per hour.