VDN-H3 in ComfyUI: setup and Turbo LoRA comparison

This tutorial installs Video DeltaNet for MiniMax H3 in ComfyUI and compares its raw video output with FastVideo and Turbo LoRA workflows. The presenter describes hybrid attention that keeps full attention for nearby frames while compressing distant context. Claims of near-lossless quality and faster generation come from the project's benchmarks and the presenter's tests.

Setup uses the ComfyUI-VDN-H3 custom node and the stage-dmd-step-250 checkpoint. The presenter downloads the complete checkpoint folder into a VDN directory under ComfyUI's models folder, then connects the advanced patch node before sampling. The demonstration runs on a Pro 6000 GPU; it does not establish minimum hardware requirements for local AI video generation.

The comparisons use combat scenes with fast and slow motion, without upsampling or graphic enhancement. In the reference-to-video test, each branch uses the same prompt, a 10-second duration and 0.8 MP resolution. The presenter reports 2:22 for VDN without LoRA, 1:51 for Turbo LoRA with Sage Attention, and 1:54 for Turbo LoRA with Comfy Kitchen plus an attention patch.

The presenter finds less fast-motion pixelation with VDN in these examples. Adding Turbo LoRA to VDN saves only a few seconds in a later test, while VDN alone sometimes looks clearer. These results describe the tested scenes and settings rather than a universal speed or quality ranking.