Small, reusable test-case packs for joint audio-video diffusion models (MiniMax H3 today).
| Pack | Contents | Location |
|---|---|---|
| H3 eval-58 | 58 case pack: core 36 (T2VA / FL2VA / REF2VA, 12 each) + hard 22 across seven difficulty categories. Chinese execution prompts with source text, media requirements, expected output, scoring dimensions and hard-fail rules for every case. | diffusion_testcase/h3-eval-58/ |
| Ref2VA pilot | The original five Ref2VA pilot cases (dog identity / scene separation / motion transfer / sound conditioning / multimodal composition) with CC-licensed media included. Kept standalone and unchanged. | diffusion_testcase/ref2va-pilot/ |
- 58 cases, one prompt sheet per case:
core/{t2va,fl2va,ref2va}/+hard/ - Machine-readable index:
cases.csv(Excel-friendly) andcases.jsonl(full fields) - 92 reference-asset URLs across 36 cases in
assets/assets.csv, fetched on demand byassets/fetch_assets.py; third-party media is not redistributed here - Generated from a single source workbook (
source/H3评测数据集_58条.xlsx) bytools/build_from_xlsx.py, so the workbook stays the source of truth
Start here: pack README · case index · how to run · coverage · sources and rights
Five MiniMax H3 Ref2VA cases covering image identity, subject/scene separation, image-plus-video motion, image-plus-audio, and multimodal reference composition, with prompts, assets, attribution and license details. See the Ref2VA pilot README.
- Rights: public accessibility is not redistribution permission. Several entries reference third-party media or real trademarks; see sources and rights before publishing any result.
- The two packs are independent:
ref2va-pilotis frozen,h3-eval-58is the maintained benchmark.