Doom 2-Player PvP Latents: 167 Hours of Pre-Encoded Gameplay for World Models

2,606 episodes of 2-player PvP Doom gameplay, totaling 167.1 hours and 21,060,484 frames, pre-encoded into latent representations. The latents were created by author chrisxx using a DC-AE-Lite f32c32 encoder and were last updated on March 7, 2026. Each frame is compressed to a (32, 15, 20) float16 latent from an original 480x640 resolution at 35 frames per second.

Use Cases

• Training video prediction models based on pre-encoded latent representations.
• Developing reinforcement learning agents for competitive gameplay based on the 2-player PvP structure.
• Benchmarking world model architectures using the specified 32x spatial compression and 32 latent channels.
• Analyzing agent behavior and strategies from the extensive 167-hour gameplay duration.

Strengths

• Large scale with 2,606 episodes and 21,060,484 frames.
• Pre-processed for efficiency, with frames compressed to a (32, 15, 20) latent shape.
• Specific technical details provided, including original resolution (480x640), frame rate (35 fps), and encoder type (DC-AE-Lite f32c32).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is large (756 GB), which may require significant storage and computational resources.

Provenance

Source

huggingface, author chrisxx

Collection Method

Pre-encoded from 2-player PvP Doom gameplay using a DC-AE-Lite f32c32 encoder.

Time Range

The temporal coverage of the source gameplay is not specified.

Freshness

Last updated 2026-03-07 05:14:11.

Geography

Spatial coverage is not applicable for this synthetic game data.