Skip to content

Dataset · hffordata

MiniMax-H3 audio-visual video generation benchmark

hffordata/h3-fewstep-benchmark-20260920

4,320 audio-visual videos from 15 accelerated versions of the MiniMax-H3 model for evaluating generation quality in few steps.

What's inside

Inside are 4,320 five-second video clips at 1344x768 resolution (24 frames per second) with a generated audio track. The clips were obtained by running 15 distilled checkpoints and adapters of the MiniMax-H3 model on 48 English text prompts with two seeds and a different number of generation steps (4, 8, and 32). With 288 videos per checkpoint, generation was performed using text only without input images.

What you can do with it

  • Comparing quality of distilled versions and adapters of MiniMax-H3 at low step counts
  • Evaluating video and audio degradation during text-to-video model inference acceleration
  • Training or calibrating automated quality assessment metrics for synthesized video
Tasks
Text → video
Downloads this month
30.5K
Likes
1

How to load it

from datasets import load_dataset

ds = load_dataset("hffordata/h3-fewstep-benchmark-20260920")