Skip to main content
Back

Ref2VA-VSA Cuts MiniMax H3 Video Generation 9x on a Single GPU

A new ComfyUI custom node brings MiniMax H3 reference-to-video generation down to 72 seconds on a single RTX 4090, with identity preservation intact.

Repo, Full Access
  • Ref2VA-VSA cuts MiniMax H3 character video generation from ~650s to ~72s warm on an RTX 4090.
  • Peak VRAM drops from ~22.5 GB to ~13.5 GB, bringing it into 16GB consumer GPU range.
  • Uses 4-step Euler sampling with 75% video sparsity top-k attention pruning.
  • Keeps reference image, audio, and text tokens 100% dense to preserve character identity.
  • Transplants FastVideo's learned VSA gate matrices onto the Ref2VA base, no retraining required.
  • Apache 2.0 licensed ComfyUI custom node with ready-to-use workflow included.

Ref2VA-VSA, an open-source ComfyUI custom node, cut the reported warm generation time for a 5.16-second MiniMax H3 reference-to-video clip from about 650 seconds to 72 seconds on one RTX 4090. Reference-to-video uses an image, text, and optional audio conditioning to animate a subject while retaining its appearance.

Peak video memory fell from roughly 22.5 GB to 13.5 GB. That measurement puts the pipeline within the capacity of some 16 GB GPUs, although the project currently lists GPUs with at least 24 GB and does not provide a validated 16 GB configuration. Drivers, memory allocation, model loading, and other active nodes can raise total usage.

The benchmark post reports no visible loss of character identity. The published evidence consists of visual comparisons from a single benchmark configuration rather than formal identity scores or a broader quality evaluation.

Inside the 72-second benchmark

The benchmark used one RTX 4090 and held the prompt, seed, resolution, frame count, and output duration constant. Sampling steps and samplers varied by pipeline, so the 9× figure measures the complete Ref2VA-VSA configuration against the 50-step native baseline.

  • Resolution: 1344 × 768
  • Duration: 5.16 seconds
  • Frame rate: 24 fps
  • Frame count: 124
  • GPU: NVIDIA RTX 4090
Author-reported results on a single RTX 4090
Pipeline Sampling Warm time Peak VRAM
MiniMax H3 native dense 50 steps, Euler About 650 seconds About 22.5 GB
H3 Turbo LoRA, dense 4 steps, Euler About 190 seconds About 21 GB
Video Delta Net, VDN-H3 8 steps, er_sde About 135 seconds About 18.5 GB
Ref2VA-VSA 4 steps, Euler