README.md

metadata

title: World Memory Deck
emoji: 🧠
colorFrom: green
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
license: mit

World Memory — Presentation

A reveal.js slide deck for World Memory: Consistent Long-Video Generation via Knownness-Conditioned Diffusion Forcing (working title, CoRL 2026).

The deck walks through the full pipeline:

• Phase 1 — Offline caching: posed video → persistent 3D PointMemory → reprojected z_mem, per-region knownness k, per-token 3D address Xtok, and a compact memory slot bank.
• Phase 2 — Model: a frozen Wan Video DiT with four small zero-init hooks (per-token noise from k, 3D positional encoding, geometry-keyed cross-attention with Rope3D, confidence-gated shortcut).
• Phase 3 — Training: anchor-mixture Diffusion Forcing — repair memory, never copy it; the target is always the real frame.
• Phase 4 — Inference & rollout: knownness sets a per-token start-noise level; known regions reuse-and-repair, unknown regions generate, and new frames are written back into the world memory for consistent revisits.

View

Open the Space — index.html renders the deck directly. Navigate with arrow keys, press F for fullscreen, S for speaker notes, O for the slide overview, and Esc to zoom out.

Built with reveal.js (loaded from CDN). The centerpiece pipeline figure is a hand-authored inline SVG.