DreamDoom

Frames witnessed: 0 · Imps harmed: 0

Action labels

Histogram

Bot speed 4× Target samples Bot personality

Training tensor
64×48×3 · K context frames + 1-hot action

Import

Conditional UNet (DDPM)

input = [ noisy x_t ‖ K context frames ] → conv ↓ ↓ ↑ ↑ with skip links
action one-hot + sin(t) embedding → FiLM-ish shift at every scale
predicts ε · cosine schedule T=1000 · context noise augmentation · 10% action dropout (CFG)

Learning rate Batch size Epochs Context length K Context noise aug Base channels

Waiting for data. The model is staring at nothing.

step 0 · epoch 0 · loss — · backend —

Loss

Preview: context · target · prediction

last ctx

target

denoised

Scrub denoising final

←/→ turn · ↑/↓ walk · Shift+↑/↓ run · Space attack · no key = NOOP

Neural-only loop

NOOP

— ms/frame · — fps

Hallucination drift

Reality integrity: —

Sampler steps (DDIM) 8 CFG scale 1.5 Bilinear upscale Show denoising

REAL ENGINE

NEURAL DREAM

PSNR — dB · SSIM —

Blind test (sides shuffled)

Score 0/0

Best Reality Detectors

Loading…

Explain like I'm an imp

1. A bot plays a real raycaster and we record (last K frames, action) → next frame.

2. A diffusion model learns to remove noise from the next frame, given the past frames and what button you pressed.

3. In Dream Mode the real engine is OFF. We start from pure static, denoise it over several steps into a frame, push it into the context, and repeat — the game is pure hallucination, steered by you.

4. Errors pile up ("drift"). Context noise augmentation in training teaches the model to tolerate its own mush.