Add discussion: Physics-Consistent Video Generation

discussions/2026-08-24-yuxin-physics-consistent-video-generation/action-items.md

@@ -0,0 +1,5 @@
+# Action Items
+
+- [ ] Yuxin: present the method roadmap and clarify model capability versus system capability.
+- [ ] All participants: identify one claim whose current evidence does not support its strongest interpretation.
+- [ ] All participants: propose one controlled comparison that could distinguish physical generalization from template matching.

discussions/2026-08-24-yuxin-physics-consistent-video-generation/idea-map.md

@@ -0,0 +1,28 @@
+# Idea Map
+
+```mermaid
+graph TD
+    A[Physics-consistent video generation] --> B[External physics and control]
+    A --> C[Generator learning]
+    A --> D[Inference-time alignment]
+    A --> E[Evaluation evidence]
+    B --> B1[Infer initial state]
+    B --> B2[Compute motion]
+    B2 --> B3[Render or complete interactions]
+    C --> C1[Simulation data]
+    C --> C2[Physical attributes]
+    C --> C3[Representation alignment]
+    C --> C4[Physics-aware rewards]
+    C --> C5[Latent physical dynamics]
+    D --> D1[World-model scoring]
+    D --> D2[Trajectory search and sampling guidance]
+    E --> E1[Open-domain human or VLM judgment]
+    E --> E2[Comparison with real futures]
+    E --> E3[Equations and physical invariants]
+    B3 --> F[System capability]
+    C5 --> G[Model capability]
+    D2 --> F
+    F --> H[Single-variable counterfactual tests]
+    G --> H
+    E3 --> H
+```

discussions/2026-08-24-yuxin-physics-consistent-video-generation/index.html

@@ -0,0 +1,41 @@
+<!doctype html>
+<html lang="en">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <meta name="description" content="How simulators, training signals, latent dynamics, and inference-time alignment shape physics-consistent video generation.">
+  <title>Physics-Consistent Video Generation · CausalVideoGroup</title>
+  <link rel="stylesheet" href="../../styles/main.css">
+</head>
+<body>
+  <header class="site-header"><div class="nav-wrap">
+    <a class="brand" href="../../">CausalVideoGroup</a>
+    <nav class="site-nav" aria-label="Primary navigation">
+      <a href="../../projects/">Projects</a>
+      <a aria-current="page" href="../">Discussions</a>
+      <a href="../../people/">People</a>
+      <a href="../../about.html">About</a>
+    </nav>
+  </div></header>
+  <main>
+    <header class="page-header">
+      <p class="eyebrow">Topic discussion · 2026-08-24</p>
+      <h1>Physics-Consistent Video Generation: Who Teaches the Generator Physics?</h1>
+      <p class="lede">Led by Yuxin Wang</p>
+      <p><span class="tag">before-meeting</span> <span class="tag">physics-consistency</span> <span class="tag">benchmark</span></p>
+    </header>
+    <article class="prose">
+      <p>This discussion maps how physical knowledge enters a video-generation system through explicit simulation, training signals, latent dynamics, and inference-time alignment, while separating improvements to the generator from improvements to the full system.</p>
+      <div class="actions">
+        <a class="button" href="presentation.html">Open presentation</a>
+        <a class="button secondary" href="summary.html">Read topic summary</a>
+        <a class="button secondary" href="references.html">References</a>
+        <a class="button secondary" href="meeting-note.html">Meeting note</a>
+      </div>
+    </article>
+  </main>
+  <footer class="site-footer"><div class="site-footer-inner">
+    <span>© CausalVideoGroup</span><a href="../">All discussions</a>
+  </div></footer>
+</body>
+</html>

discussions/2026-08-24-yuxin-physics-consistent-video-generation/meeting-note.html

@@ -0,0 +1,21 @@
+<!doctype html>
+<html lang="en">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <meta name="description" content="Meeting Note for Physics-Consistent Video Generation: Who Teaches the Generator Physics?.">
+  <title>Meeting Note · Physics-Consistent Video Generation: Who Teaches the Generator Physics? · CausalVideoGroup</title>
+  <link rel="stylesheet" href="../../styles/main.css">
+</head>
+<body>
+  <header class="site-header"><div class="nav-wrap">
+    <a class="brand" href="../../">CausalVideoGroup</a>
+    <nav class="site-nav" aria-label="Discussion navigation"><a href="./">Overview</a><a href="presentation.html">Presentation</a><a href="summary.html">Summary</a><a href="references.html">References</a><a href="meeting-note.html">Meeting note</a></nav>
+  </div></header>
+  <main>
+    <header class="page-header"><p class="eyebrow">2026-08-24 · Yuxin Wang</p><h1>Meeting Note</h1><p class="lede">Physics-Consistent Video Generation: Who Teaches the Generator Physics?</p></header>
+    <article class="prose markdown-body"><p>The discussion is scheduled for 2026-08-24. Replace this notice after the presentation with changed views, evidence boundaries, critiques, research opportunities, concrete experiments, and explicitly assigned action items.</p></article>
+  </main>
+  <footer class="site-footer"><div class="site-footer-inner"><span>© CausalVideoGroup</span><a href="./">Back to discussion</a></div></footer>
+</body>
+</html>

discussions/2026-08-24-yuxin-physics-consistent-video-generation/meeting-note.md

@@ -0,0 +1,3 @@
+# Meeting Note
+
+The discussion is scheduled for 2026-08-24. Replace this notice after the presentation with changed views, evidence boundaries, critiques, research opportunities, concrete experiments, and explicitly assigned action items.

discussions/2026-08-24-yuxin-physics-consistent-video-generation/metadata.yaml

@@ -0,0 +1,42 @@
+title: "Physics-Consistent Video Generation: Who Teaches the Generator Physics?"
+date: 2026-08-24
+topic_slug: physics-consistent-video-generation
+leader:
+  name: "Yuxin Wang"
+  short_name: yuxin
+status: before-meeting
+summary: >-
+  A design-space view of how simulators, explicit controls, training signals,
+  latent dynamics, and inference-time alignment can make generated videos obey
+  real-world physical laws, and what current benchmarks can actually establish.
+tags:
+  - video-generation
+  - physics-consistency
+  - world-model
+  - controllable-generation
+  - benchmark
+  - evaluation
+related_projects: []
+papers:
+  - title: "PhysGen: Rigid-Body Physics-Grounded Image-to-Video Generation"
+    url: https://arxiv.org/abs/2409.18964
+    role: system
+    note: Adds explicit rigid-body simulation before diffusion refinement.
+  - title: "PISA Experiments: Exploring Physics Post-Training for Video Diffusion Models by Watching Stuff Drop"
+    url: https://arxiv.org/abs/2503.09595
+    role: method
+    note: Tests physics post-training and extrapolation through controlled free fall.
+  - title: "VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models"
+    url: https://arxiv.org/abs/2505.23656
+    role: method
+    note: Distills spatiotemporal relations into the generator.
+  - title: "What about gravity in video generation? Post-Training Newton's Laws with Verifiable Rewards"
+    url: https://arxiv.org/abs/2512.00425
+    role: method
+    note: Turns Newtonian motion proxies into verifiable rewards.
+  - title: "Inference-time Physics Alignment of Video Generative Models with Latent World Models"
+    url: https://arxiv.org/abs/2601.10553
+    role: system
+    note: Guides inference-time search with a frozen latent world model.
+external_videos: []
+slides: presentation.html