# nous-diffusion-world

Diffusion-based world models (Phase 176.6), PyTorch, CPU-runnable with
overfit-verified sampling.

## Components

- **EDM core** (`edm.py`) — Karras preconditioning
  (`c_skip/c_out/c_in/c_noise`), log-normal noise sampling, EDM loss weighting,
  the ρ=7 σ-schedule, and a deterministic 2nd-order **Heun** sampler (DIAMOND
  runs it in 3 steps).
- **DIAMOND** (`diamond.py`) — a convolutional FiLM denoiser predicting the next
  frame conditioned on past frames (concatenated channels) and the action
  (embedded), with autoregressive rollout. (176.6.1.1)
- **GameNGen** (`gamengen.py`) — DIAMOND + **context-frame noise augmentation**:
  training corrupts the past frames at a random level and conditions the
  denoiser on it, the trick that keeps long interactive rollouts from drifting.
  (176.6.1.2)
- **Hybrid pipeline** (`hybrid.py`) — cheap recurrent latent dynamics (Genie's
  long-range role) advanced every step + an EDM **diffusion decoder** rendering
  photorealistic frames every `decode_every` steps: the throughput ⇄ fidelity
  knob. (176.6.1.4)
- **Oasis open-domain model** (`oasis.py`) — continuous multimodal-action and
  text-conditioned EDM dynamics with autoregressive history plus the specified
  consistency/FID/enjoyability comparison scorecard.

## Not included (external — left unchecked in TODOS)

The Minecraft-YouTube + Maya-alpha training and empirical comparison against a
trained Genie-Sovereign-3 required by `176.6.1.3` need full datasets and
accelerators; its locally actionable architecture and evaluation contract are
implemented above.

## Tests

```
cd libs/nous/diffusion-world && python3 -m pytest tests/ -q
```

Verifies EDM preconditioning limits + schedule, that the denoiser reconstructs a
constant target through sampling, that DIAMOND/GameNGen learn action-conditioned
next frames (correct frame per action after training), and that the hybrid
decoder reconstructs a frame from its latent while the rollout decodes only
every K steps.