---
title: Rollout vs Ground Truth
emoji: 🎬
colorFrom: green
colorTo: purple
sdk: gradio
sdk_version: 6.22.0
python_version: '3.12'
app_file: app.py
pinned: false
short_description: Frame-locked viewer for rollouts vs. ground truth
---

# Rollout vs. ground truth

Side-by-side viewer for forward-dynamics world-model rollouts. The page starts empty and assumes
no dataset — type a **dataset repo id** into the box, press **Load**, and pick an episode.

Accepted forms: `owner/name`, `owner/name@revision`, or a pasted
`https://huggingface.co/datasets/…` URL.

## What it shows

Ground truth and the generated rollout are composed into a **single video**, so the two panels
stay frame-locked; two independent players drift apart and cannot be scrubbed together. The strip
carries in-frame captions, a frame counter, and a timeline marking the critical frame.

Sidebar controls: search, sort (critical-frame PSNR, task, manifest order), the episode list, a
camera selector, an optional `|GT − generated|` difference panel, and playback fps — a 33-frame
window is only ~1 s at native 30 fps. PSNR and MSE for the selected episode are shown underneath.

## Expected dataset layout

```
windows_manifest.json                 # optional: list of window dicts (episode_id, task,
                                      #   edge_type, arm, start_frame, critical_frame, ...)
metrics_summary.json                  # optional: [{id, task, edge, arm, full_mse, full_psnr,
                                      #   crit_mse, crit_psnr}, ...]
windows/<episode_id>/gt.mp4           # required
windows/<episode_id>/generated.mp4    # required, same geometry and length as gt.mp4
```

Only the two mp4s are required. With no `windows_manifest.json` the episode list is discovered by
listing `windows/` one level deep. Videos may stack **N cameras vertically** — N is inferred from
the frame aspect ratio and the camera selector adapts. Black letterbox bars are trimmed
automatically, identically for both videos.

## Storage model

Nothing is downloaded at build time and no repo is ever pulled whole, so datasets far larger than
the Space still work.

- Loading a repo fetches only the manifest and metrics JSONs (a few hundred KB), or one directory
  listing when there is no manifest.
- An episode's mp4s are fetched the first time that episode is opened, never in bulk.
- Each browser session gets its own scratch directory. At most `MAX_WINDOWS_ON_DISK` episodes stay
  resident; beyond that the least recently used are evicted along with their rendered videos.
  Decoded frames are capped at 2 episodes in RAM and composed strips at 1.
- The directory is deleted when the session ends. A janitor also reaps sessions idle past
  `SESSION_TTL` and directories orphaned by an earlier process.

## Configuration

| Variable | Default | Purpose |
|---|---|---|
| `HF_TOKEN` | — | **Required for private datasets.** Set under *Settings → Variables and secrets*. |
| `MAX_WINDOWS_ON_DISK` | `12` | Episodes kept per session before LRU eviction. |
| `MAX_DISCOVER` | `5000` | Cap on episodes listed when a repo has no manifest. |
| `SESSION_TTL` | `3600` | Seconds of idleness before a session's storage is reclaimed. |

Note: on a public Space the `HF_TOKEN` is used to read whatever repo a visitor types, so scope it
to only the datasets you are willing to expose.

## Running locally

```bash
pip install -r requirements.txt
export HF_TOKEN=hf_...          # if the dataset is private
python app.py
```