---
license: mit
tags: [world-model, dream-drift, atari, evaluation, dreamerv3]
pretty_name: dream-drift DreamerV3 eval frames
---

# dream-drift · DreamerV3 evaluation frames

Raw dream-vs-reality frame dumps behind the DreamerV3 rows of the [dream-drift](https://github.com/iwillsolvehardestproblem/dream-drift) benchmark. Re-score every published number without a GPU.

One file per game and protocol. `<game>id.npz` = in-distribution (the model's own replay trajectories); `<game>.npz` = cross-policy (expert DIAMOND-policy trajectories, the "confident healing" regime).

Arrays per file:
- `real` uint8 (32, 50, 64, 64, 3): ground-truth frames (HWC), 32 start states x 50 steps
- `dream` uint8 (32, 4, 50, 64, 64, 3): 4 dream samples per start, same recorded actions, open loop
- `starts` (32,): start indices into the source trajectory; `game`, `burnin` (=4), `seed` (=0) scalars
- `rrew`, `drew` (32, 50) / (32, 4, 50): real and predicted rewards (predicted are raw; threshold |r|<0.5 to 0, then sign)
- `rend`, `dend`: real and predicted termination flags

Scoring: `python harness/dreamer_score.py <dir-with-raw_frames.npz>` in the dream-drift repo reproduces `summary.json`, `drift.png`, and the videos.