MiniMax H3 VDN in ComfyUI: Hybrid Attention Acceleration, 8-Step Sampling and Consumer-GPU Reality

VDN is a recent MiniMax H3 acceleration technique built on Hybrid Attention. The official claim is dramatic: with 8 B200 GPUs and 8-step sampling, a 14.4-second video finishes denoising in 11.23 seconds. The practical question is what it does on consumer cards. The notes below come from the episode subtitles and the published mirror page; no workflow JSON was shipped or parsed, so no node-level parameters are claimed beyond what the episode states.

What VDN actually gives you

• On consumer GPUs the speed is roughly in the same class as the LightX2V stack — the headline B200 numbers are not what you get at home.
• Its real advantage is quality at the same length compared with other acceleration models, thanks to the attention change.
• The longer the video, the larger the acceleration gain becomes.

The episode's own measurements: a 10-second video rendered in 4 min 43 s, a 5-second video in 2 min 53 s.

The two workflows in the mirror

The mirror ships two graphs: T2V and R2V. VDN supports both the FL2VA and Ref2VA models. In the FL2VA graph, whether the first and last frames are connected determines the generation mode; the sample shown is effectively text-to-video.

The core node is the VDN node, with several adjustable parameters:

• VDN checkpoint — keep the default. The VDN MiniMax H3 INT8 ConvRot version is about 2.2 GB, which eases VRAM pressure.
• The BF16 version above it is about 4.3 GB; keep the default.
• Apply Turbo Adapter is the next step in the chain.

Hardware

The author recommends a 5090 or a 48 GB 4090 and above. A 24 GB card can still run with lower resolution and shorter duration. Preemptible GPU instances are cheaper but can be reclaimed at any time.

Deploying the mirror

1. Open the mirror page: https://www.compshare.cn/images/2TiHWAMekuFP?referral_code=A7fTQUYY9lCDFJwA6MMvoq
2. Deploy an instance and wait for the image to come up.
3. Open ComfyUI and pick the T2V or R2V workflow.

Links

• Mirror (with referral code): https://www.compshare.cn/images/2TiHWAMekuFP?referral_code=A7fTQUYY9lCDFJwA6MMvoq
• Video tutorial (YouTube): https://www.youtube.com/watch?v=PmhZzXGGcGU
• Chinese original (Bilibili): https://www.bilibili.com/video/BV1Fvbx6jEuv

Limits

• The official 8×B200 numbers are not consumer-GPU performance; the episode's timings are single-machine observations.
• VDN is a fast-moving community technique; node behaviour may change between versions.
• The mirror is a deployment convenience, not a quality guarantee.