chrisxx/doom-2players-latents

Doom 2-Player PvP Latents (DC-AE)

Pre-encoded latent representations of 2-player PvP Doom gameplay, ready for training video world models. Encoded using DC-AE-Lite f32c32 (32x spatial compression, 32 latent channels).

Dataset Details

Property	Value
Episodes	2,606
Total duration	167.1 hours (both perspectives)
Total frames	21,060,484
Latent shape per frame	(32, 15, 20) float16
Original resolution	480x640
Frame rate	35 fps
Size	756 GB
Format	WebDataset tar shards (~4 GB each)
VAE	mit-han-lab/dc-ae-lite-f32c32-sana-1.1-diffusers

Actions: [ "MF", # MOVE_FORWARD "MB", # MOVE_BACKWARD "MR", # MOVE_RIGHT "ML", # MOVE_LEFT "W1", # SELECT_WEAPON1 "W2", # SELECT_WEAPON2 "W3", # SELECT_WEAPON3 "W4", # SELECT_WEAPON4 "W5", # SELECT_WEAPON5 "W6", # SELECT_WEAPON6 "W7", # SELECT_WEAPON7 "ATK", # ATTACK "SPD", # SPEED "TURN",# TURN_LEFT_RIGHT_DELTA ]

VAE Details

DC-AE-Lite with f32c32 configuration:

• Spatial compression: 32x (480x640 pixels → 15x20 latent spatial dims)
• Latent channels: 32
• Precision: float16
• Compression: 3 × 480 × 640 RGB → 32 × 15 × 20 latent (42.7x compression ratio)

Data Structure

Each episode is stored as a group of files inside WebDataset tar shards:

{episode_key}.latents_p1.npy    # Player 1 latents (N, 32, 15, 20) float16
{episode_key}.latents_p2.npy    # Player 2 latents (N, 32, 15, 20) float16
{episode_key}.actions_p1.npy    # Player 1 actions (N, 14) float32
{episode_key}.actions_p2.npy    # Player 2 actions (N, 14) float32
{episode_key}.rewards_p1.npy    # Player 1 rewards (N,) float32
{episode_key}.rewards_p2.npy    # Player 2 rewards (N,) float32
{episode_key}.meta.json         # Episode metadata

Usage

Training Loader (streaming, all clips)

from doom_arena.latent_loader import LatentTrainLoader

loader = LatentTrainLoader(
    "path/to/latent/shards",
    clip_len=16,       # frames per clip
    batch_size=64,
    num_workers=4,
)

for batch in loader:
    latents_p1 = batch["latents_p1"]   # (B, T, 32, 15, 20) float16
    latents_p2 = batch["latents_p2"]   # (B, T, 32, 15, 20) float16
    actions_p1 = batch["actions_p1"]   # (B, T, 14)
    actions_p2 = batch["actions_p2"]   # (B, T, 14)
    rewards_p1 = batch["rewards_p1"]   # (B, T)
    rewards_p2 = batch["rewards_p2"]   # (B, T)

Random-access Dataset

from doom_arena.latent_loader import LatentDataset

ds = LatentDataset("path/to/latent/shards")
ds.summary()

ep = ds[42]
print(ep)                  # LatentEpisode(dwango5_5min PvP, 10467 frames)
print(ep.latents_p1.shape) # torch.Size([10467, 32, 15, 20])

# Index into frames
clip = ep[100:116]         # dict with 16-frame slices of all arrays

Encoding frames to latents

import torch
from diffusers.models.autoencoders.autoencoder_dc import AutoencoderDC

# Load VAE
vae = AutoencoderDC.from_pretrained(
    "mit-han-lab/dc-ae-lite-f32c32-sana-1.1-diffusers",
    torch_dtype=torch.float16,
).cuda().eval()

# Encode: (B, 3, 480, 640) float16 RGB [-1, 1] → (B, 32, 15, 20) float16
frames = torch.randn(4, 3, 480, 640, dtype=torch.float16, device="cuda")
with torch.no_grad():
    latents = vae.encode(frames).latent  # (4, 32, 15, 20)

Decoding latents to frames

# Decode: (B, 32, 15, 20) float16 → (B, 3, 480, 640) float16 RGB [-1, 1]
with torch.no_grad():
    reconstructed = vae.decode(latents)  # (4, 3, 480, 640)

# Convert to uint8 for display
pixels = ((reconstructed.clamp(-1, 1) + 1) / 2 * 255).byte()

Performance

Throughput measured with clip_len=70, batch_size=64, num_workers=4:

Storage	Frames/s	Batches/s	Seconds/batch
NFS	3,356	0.75	1.33s
Local NVMe	22,371	4.99	0.20s

Recommendation: Copy shards to local NVMe before training:

rsync -av path/to/shards/ /tmp/pvp_latents/

Source

• Encoded from chrisxx/doom-2players-mp4
• VAE: mit-han-lab/dc-ae-lite-f32c32-sana-1.1-diffusers
• Project: doom-arena

Citation

If you use this dataset, please cite the DC-AE paper:

@article{chen2024dcae,
  title={Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models},
  author={Chen, Junyu and Cai, Han and Chen, Junsong and Xie, Enze and Yang, Shang and Tang, Haotian and Li, Muyang and Lu, Yao and Han, Song},
  journal={arXiv preprint arXiv:2410.10733},
  year={2024}
}