README.md

# ScenePilot release bundle

This archive is a compact snapshot of the ScenePilot training, inference, and
evaluation code used for the reported experiments. It intentionally excludes
datasets, generated videos, experiment logs, temporary runtime scripts, caches,
and intermediate checkpoints.

## Included checkpoints

- `checkpoints/scenepilot_step60500.ckpt`: primary checkpoint used for the final
  ScenePilot camera, scene-time, joint-control, and memory evaluations.
- `checkpoints/scenepilot_step49500.ckpt`: frequently used 40K+ checkpoint for
  the AD-heavy qualitative evaluations.

The checkpoints are PyTorch zip-format files. Their SHA-256 hashes are listed
in `SHA256SUMS`.

## Base-model dependency

The Wan2.2 TI2V 5B base-model assets are not duplicated in this archive. The
training and inference launchers expect a separately downloaded
`Wan2.2-TI2V-5B` directory containing the DiT shards, VAE, UMT5 encoder, and
tokenizer. Set `WAN22_ROOT` to that directory before running the launchers.

## Main entry points

- Stage-1 training:
  `code/train_recammaster_stage_0_bidir_varhist_df_wan22_ti2v5b.py`
- Stage-2 memory training:
  `code/train_recammaster_stage_0_bidir_varhist_memory_static3drope_df_wan22_ti2v5b.py`
- Stage-2 launcher:
  `code/run_stage0_bidir_varhist_memory_static3drope_df_wan22_ti2v5b.sh`
- Camera and scene-time evaluation:
  `code/inference/static_cam_time_eval/`
- Joint camera-time evaluation:
  `code/inference/joint_camera_time_eval/`
- Final evaluation launcher:
  `code/tools/run_scenepilot_step60500_full_eval_2gpu_20260721.sh`

## Provenance

- Source working tree: `/home/hongbo.liu2/recam-helios`
- Git commit at packaging time: `cbe6d9c`
- Branch: `hongbo/dfhelios-cleanup-0520`
- Packaging date: 2026-08-03

The bundle captures the current working-tree files, including local changes
that were used during training and evaluation.