ComfyUI reports faster encoding and decoding for MiniMax H3’s video VAE—the component that compresses frames for the model and reconstructs them afterwards. The improvement concerns this stage of the workflow. [1]
What did the test show?
In the developer’s RTX 5090 test, a 1344 × 768 video with 129 frames took 24.3 seconds before and 12.7 seconds after the update for an encode/decode round trip. These are developer measurements, not our own benchmark. [1]
Does the picture change?
The developer says the quality difference is unlikely to be visible by eye. That is an attributed assessment, not our independent finding. [1]
What it means for Rogue Signal
We use MiniMax H3 and ComfyUI to make Rogue. Claude’s production notes record ComfyUI 0.37.4, an fp16 video VAE and no fast fp16-accumulation setting. Some improvements may already apply; we have not run a controlled before-and-after test.
Our next useful measurement would separate time spent in each stage while comparing the resulting footage. We are not claiming a speed gain for our RTX 5080 setup.
Source
- ComfyUI — Making the MiniMax H3 Video VAE 2x Faster ↗
Jukka Seppänen · 22 September 2026. Developer report with test conditions and implementation details.
How this episode was made
ChatGPT researched the story and prepared the article and panels. Claude wrote the script and produced the video. Barry selected the story and approved the script and all three presenter clips. Rogue is an AI-generated presenter.
This episode trials three shots: front close-up, wide with hands out of frame, and three-quarter. The web version includes burned-in captions. Finishing combines 60% SeedVR2 output with 40% of a softer upscale, plus subtle grain, grading and room reflections in the audio.