H3 SDG Prompt — 500 Video Examples
A random 500-sample example set of MiniMax-H3 text-to-video generations from the
prompts in zijianh/H3_SDG_Prompt
(9,097 prompts total).
Structure
videos/<video_id>.mp4 # generated video
prompts/<video_id>.txt # the exact prompt used to generate it
<video_id> is the source dataset's stable id field, so every example can be
traced back to its full SDG planning metadata (taxonomy, scene spec, episode
program, feasibility review) in the source repo.
Generation settings
| Model | MiniMax-H3 (FL2VA partition), served via vLLM-omni |
| Task | t2va (text → video + audio) |
| Resolution | 1344 × 768 |
| Duration | 15 s @ 24 fps |
| Inference steps | 50 |
flow_shift / audio_flow_shift |
12 / 3.0 |
| Seed | 1101 (fixed for all samples) |
Prompt format
The .txt files contain the prompt exactly as submitted to the model, which
wraps the source dataset's prompt in a structured template:
integrated_multimodal_description: <source prompt>
overall_soundscape: N/A
non_diegetic_music: N/A
The overall_soundscape / non_diegetic_music fields were set to N/A for this
run; audio was left to the model's own generation.
Sampling
500 of 9,097 drawn with Python random.seed(42); random.sample(range(9097), 500).
_sample_manifest.json records the selected indices and ids for reproducibility.
- Downloads last month
- 12