Innovate Futures @ Benji

VDN-H3 in ComfyUI — Faster MiniMax H3 Without the LoRA (Patreon)

Published:

2026-09-04 15:42:58

Imported:

2026-09

Tags:

MiniMax

Downloads

Content

Video: https://youtu.be/kqyPTv77j3g

Hey everyone.

New video is up, and this one is for anyone running MiniMax H3 who's tired of fast motion turning into pixel soup.

Video DeltaNet just released VDN-H3 — a hybrid attention checkpoint that patches into your existing model pipeline and samples at 8 steps without a Turbo LoRA at all. I installed the ComfyUI node, wired it up, and ran it against the usual Turbo LoRA + Sage Attention combo on the same combat scenes.

Honest finding from the test: the LoRA route is a bit faster on the clock, but VDN kept the fast-motion areas clean where the other setup fell apart. And because it barely touches the original weights, the raw output looks like the base model, not a distilled version of it.

Both test workflows from the video are attached right here, free — VDN-H3-VS-fastvideoH3_t2v.json is the text-to-video A/B, and Minimax-H3VDN-R2V-testlab.json is the three-branch reference-to-video testlab. Open them, drop in your checkpoint, run them as-is.

Patreon Supporter , additional content will post to you soon.

Have fun!

— Benji

Resources:

- VDN weights + repo: https://github.com/OpenVDN/vdn-minimax-h3

- Weights on Hugging Face: https://huggingface.co/OpenVDN/vdn-minimax-h3

- ComfyUI node port: https://github.com/Saganaki22/ComfyUI-VDN-H3

- Install: clone into ComfyUI/custom_nodes, then hf download OpenVDN/vdn-minimax-h3 --include "stage-dmd-step-250/*" --local-dir <ComfyUI>/models/vdn

- Free test workflows attached on my Patreon (public post): VDN-H3-VS-fastvideoH3_t2v.json + Minimax-H3VDN-R2V-testlab.json

Files

Video DeltaNet Minimax H3 in ComfyUI — Faster Base Model Without Losing Quality

Video DeltaNet (VDN-H3) patches MiniMax H3 in ComfyUI with hybrid linear attention — 8-step sampling with near-lossless quality, no Turbo LoRA required. In this video I install the ComfyUI-VDN-H3 custom node, wire the stage-dmd-step-250 checkpoint, and run it head to head against FastVideo and the Turbo LoRA + Sage Attention setup everyone uses. You'll see raw first-sampling outputs with no upsampling or enhancement — fast motion, slow motion, and a seven-tail fox combat scene — so you can judge what each method actually produces. Then the reference-to-video testlab: three branches, same prompt, same 10 second duration, same 0.8 MP resolution. VDN alone hits 2:22 without any LoRA; Turbo LoRA with Sage Attention lands at 1:51; Comfy Kitchen plus Sol Attention at 1:54. The real finding: VDN keeps fast-motion areas clean where the LoRA route starts pixelating. This one is for ComfyUI users running MiniMax H3 who care about motion quality, not just benchmark speed. If you've been stacking turbo LoRAs and attention patches and wondering why fast action still breaks apart, this comparison shows where the loss actually comes from. Resources: - VDN weights + repo: https://github.com/OpenVDN/vdn-minimax-h3 - Weights on Hugging Face: https://huggingface.co/OpenVDN/vdn-minimax-h3 - ComfyUI node port: https://github.com/Saganaki22/ComfyUI-VDN-H3 Local Workstation GPU : https://amzn.to/3XfXsAO -------------------------------------------------------------------------------------------------------------------------------- If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech #MiniMaxH3 #ComfyUI #AIVideo #VDN #VideoDeltaNet