---
pretty_name: SpatialMirage Stage 2 Training Data
license: apache-2.0
task_categories:
  - image-to-image
---

# SpatialMirage Stage 2 Training Data

This public release contains the SpatialEdit-500K camera-control subset used by
SpatialMirage Stage 2. It contains 243,338 training pairs and 4,096
held-out construction-time evaluation pairs. The 6,344 original upstream
WebDataset shards occupy 125,005,230,080 bytes and preserve the original source and
target JPEG bytes without re-encoding.

The upstream dataset is `EasonXiao-888/SpatialEdit-500K` at revision
`e54a9c870f98487d63f789a5e1bde4d35937d227` and is distributed under Apache 2.0.

## Contents

* `annotations/train_stage2_spatialedit_camera_243338.jsonl` is the exact row set
  used by the final Stage 2 recipe with portable image paths.
* `annotations/eval_stage2_spatialedit_camera_4096.jsonl` records the held-out
  construction split.
* `raw/camera_control/*.tar` are byte-identical original upstream shards.
* `scripts/materialize_images.py` extracts only the images referenced by a chosen
  annotation file.

## Materialization

```bash
python scripts/materialize_images.py \
  --annotation annotations/train_stage2_spatialedit_camera_243338.jsonl \
  --output-root /path/to/stage2_data
```

Resolve annotation paths against `/path/to/stage2_data` during training.

## License and citation

See `LICENSE`, `licenses/UPSTREAM_DATASET_CARD.md`, and
`licenses/THIRD_PARTY_NOTICES.md`.