README.md

---
license: other
task_categories:
  - text-to-video
tags:
  - text-to-video
  - video-generation
  - benchmark
  - evaluation
  - world-model
size_categories:
  - n<1K
---

# omini_time_space_eval

Text-to-video benchmark suite probing whether a video world model represents
**time**, **space**, and **camera control** — 659 generated clips across four
benchmarks, with the prompts that produced them and the eval harness that
scores them.

| Bench | Items | What it tests |
|---|---:|---|
| `bench_t2v_time_comprehensive` | 200 | Time/weather transitions (20 variants incl. non-monotonic day→night→day) over outdoor scenes |
| `bench_t2v_space` | 162 | Grounded knowledge of specific real-world places, each anchored to a real Wikimedia reference photo (103 worldwide + 59 USA) |
| `bench_t2v_space_comprehensive` | 137 | Same space axis, hand-curated from 97 countries, no reference photo |
| `bench_t2v_camera_time` | 160 | Instructed camera motion (8 types) crossed with time/weather transition (20 variants) — full 8×20 grid, no confounds |

## Layout

```
bench_t2v_camera_time/
  manifest.json        prompts + per-item metadata
  results.jsonl        generation status/timing per item
  videos/*.mp4         the generated clips
  score_claude.py      the eval judge
  summarize.py         aggregation
bench_t2v_space/
  manifest_worldwide.json, manifest_usa.json
  {worldwide,usa}/results.jsonl
  {worldwide,usa}/videos/*.mp4
  score_claude.py, summarize.py
...
```

Note the eval scripts' default `--manifest` / `--videos-dir` values point at the
layout of the original working tree (`outputs/videos/`), not this repo's
flattened layout — pass both flags explicitly when running against these files.

## Evaluation

`score_claude.py` is the only judge: it samples 8 frames evenly across each clip,
downscales them to 768px on the long edge, and sends them to **Claude Opus 5 on
Amazon Bedrock** with a per-bench rubric (each axis scored 0–10).

| Bench | Axes | Pass criteria |
|---|---|---|
| `bench_t2v_camera_time` | `time_alignment`, `camera_motion`, `quality`, `smoothness` | ≥8, ≥6, ≥7 |
| `bench_t2v_space`, `bench_t2v_space_comprehensive` | `alignment`, `quality`, `smoothness` | ≥8, ≥7 |
| `bench_t2v_time_comprehensive` | `time_alignment`, `quality`, `smoothness` | ≥8, ≥7 |

```bash
pip install 'anthropic[bedrock]'
export AWS_REGION=us-east-1     # region where Claude Opus 5 is enabled

python score_claude.py --manifest manifest.json --videos-dir videos --out-dir .
python summarize.py --out-dir .
```

`score_claude.py` is resumable (skips ids already in `claude_scores.jsonl`) and
concurrent (`--concurrency`, default 4). `summarize.py` writes `scores.jsonl`
and `summary.json`.

**Score files are not yet included in this release** — the harness is published
here, the results will be added in a later revision.

## Provenance and licensing

Please read before reuse:

- **Videos** are generated by an internal `Cosmos3-Nano` checkpoint served
  through a modified vLLM; they are model outputs, not captured footage.
- **Prompts** in the time/camera benches derive scene descriptions from
  **PAI-Bench-G** captions (`t2v_prompts.json`, outdoor subset). The
  `image_name` field carries the source identifier. PAI-Bench-G's own license
  governs reuse of that derived text.
- **`ref_image` paths** in the space manifests point at Wikimedia Commons photos
  that are **not redistributed here** — only the relative filename and the
  `source_url` are included. Each such photo carries its own CC license; fetch
  and attribute individually if you need them.
- Clips depict real landmarks. They are synthetic generations, not photographs
  of those places.

The `license: other` tag reflects that the components above carry different
terms; there is no single blanket license for this repository.