Part IV · Ch. 10 — Video Model

draft

Connectionist · Video · Spatio-temporal DiT (diffusion). Generates video coherent in time from text/image. Card: ../02-types-of-ai/02-connectionist.kmd.

🎨 Figure F-IV.10.0Coherence in time. Brief: sequence of generated frames, with temporal attention arrows linking objects across frames (identity/motion consistency).

Video Model

1. Definition and short history

Extends diffusion to the temporal dimension, generating coherent clips. Lineage: video GANs → video diffusion → Sora, Veo, Kling, Runway (2024+). It is the most compute-expensive generation modality.

2. Foundations

  • Physics / optics — motion, lighting, physical consistency.
  • Probability / stochastic processes — diffusion (ch. 02) in time.
  • DSP / compression — codecs and spatio-temporal latents.
  • Geometry — implicit 3D coherence, camera.

3. Algorithms and architectures

  • Spatio-temporal latent — 3D VAE compresses video.
  • Spatio-temporal DiT — diffusion with temporal attention + spatial.
  • Conditioning — text, initial image, keyframes, camera.
  • Coherence — attention across frames; identity/motion consistency.

4. Inputs

  • Hardware: heavy GPU/TPU; memory is the bottleneck (long sequences).
  • Data: video + captions; licensing and curation are expensive.
  • Data structures: 5D tensors (time), spatio-temporal latents.
  • Systems: video pipelines; generation by windows/chunks.

5. Specialized life cycle

Stage Specialization
0 Problem Duration, resolution, FPS, control (textimagecamera)
1 Data Licensed video+text, deduplication, filtering
2 EDA Scene/motion distribution, quality, biases
3 Modeling 3D VAE + temporal DiT; conditioning; temporal window
4 Training Extremely costly; checkpointing; staged training
5 Evaluation Quality, temporal coherence, prompt alignment, physics
5.5 Acceptance Deepfake safeguards, watermark (C2PA), content filters
6 Production Generation by chunks; high cost/latency; queues
7 Monitoring Quality, cost, abuse (disinformation)
8 Retraining New styles, longer duration/higher resolution
9 Governance Disinformation, deepfakes, image rights, consent

6. Capabilities, modes and modalities

Visualtemporalartistic: clips from text/image, animation, effects, storyboard. Converging with world models (dynamics simulation).

7. Limits, risks and ethics

Very high compute cost; physical/temporal incoherences; video deepfakes and disinformation; image rights and data. Provenance watermarking is critical.

8. State of the art and examples

Sora, Veo, Kling, Runway Gen-3; trend: longer clips, camera control, synchronized audio, and a bridge to world models (related chapter: robotics).