Karajan42/diamond-doom-hg-v5.3

Diamond Doom Health Gathering v5.3

A diffusion-based world model trained on Doom Health Gathering maps with zero entity persistence regression through 490 effective epochs.

Key Features

• 69.8M parameter EDM (Elucidated Diffusion Model) with AR4 autoregressive conditioning
• 6-channel minimap conditioning: disparity, walls, items top-down, items FOV, enemies top-down, enemies FOV
• Zero disappearances across all persistence tests (items AND enemies)
• 5-action space: NOOP, FORWARD, TURN_LEFT, TURN_RIGHT, ATTACK
• GameNGen noise augmentation (10 buckets, alpha_max=0.7)

Training Data

• 1400 episodes on 100 procedurally generated Health Gathering WADs
• Unkillable DoomPlayers as enemies (100% visibility in non-combat episodes)
• 42.5% combat, 42.5% normal (SF agent), 15% edge-case episodes
• 723,800 total training steps

Persistence Results

Epoch	Items Disappearances	Enemies Disappearances	Denoiser Loss
50	0	0	0.01123
100	0	0	0.01069
200	0	0	0.01011
300 (this ckpt)	0	0	0.00973

Usage

Controls: Z/UP=Forward, A/LEFT=Turn Left, R/RIGHT=Turn Right, SPACE=Attack

Architecture

Based on DIAMOND with minimap conditioning extensions. Architecturally equivalent to the Memory/Observation/Dynamics decomposition in MultiGen (Stanford/Google, 2026).

Files

• agent_epoch_00300.pt — Model checkpoint (267MB, effective epoch 490)
• config/ — Hydra configs for agent, env, and trainer
• data/v5_3_init/ — 3 initialization episodes with weapon + enemies