Summary: Video generation models synthesize temporally coherent video sequences from text, images, or actions. The 2024 breakthroughs (Sora, Genie, VideoPoet) achieved realistic motion and long-range consistency by scaling diffusion transformers on massive video datasets. These models are converging with world models for robotics — video generation provides the visual fidelity, while world model structure provides action-conditioning and temporal coherence.


Architecture Evolution

Diagram: (Mermaid diagram - view source for diagram code)

timeline
    title Video Generation Architecture Timeline
    2018 : GAN-based (MoCoGAN, TGAN) — short clips, low fidelity
    2020 : Autoregressive (VideoGPT) — pixel-level, slow, limited length
    2022 : Latent Diffusion (Video LDM) — compressed space, better quality
    2023 : DiT for Video (MagicVideo, Emu Video) — Transformer backbone
    2024 : Sora (OpenAI) — Scaling DiT, patch-based, world sim
    2024 : Genie (DeepMind) — Action-controllable from unlabeled video
    2024 : UniSim (DeepMind) — Universal simulator from 1.6M trajectories
    2024 : VideoPoet (Google) — Unified LLM for video generation/editing

Core Approaches

Approach Architecture Temporal Modeling Control Key Example
Pixel Autoregressive Transformer/VQ-VAE Next-frame prediction Text/conditional VideoGPT (2021)
Latent Diffusion U-Net/ViT in VAE latent Multi-frame denoising Text/image Video LDM (2022)
Diffusion Transformer (DiT) DiT blocks with space-time attention Joint space-time denoising Text/image/action Sora, Genie, VideoPoet
Consistency/One-Step Distilled ODE trajectory Single/multi-step Text/image LCM-Video, AnimateLCM
World Model / Simulator Latent action + dynamics Controllable rollout Actions/latent actions Genie, UniSim, DreamerV3+Diff

Sora Architecture (OpenAI, 2024)

Diagram: (Mermaid diagram - view source for diagram code)

flowchart TB
    subgraph INPUT["Input Processing"]
        TEXT[Text Prompt] --> T_ENC[Text Encoder T5/CLIP]
        IMG[Optional: Image/Video] --> V_ENC[VAE Encoder]
    end
subgraph PATCH["Patch Embedding"]
    V_ENC --> PATCH[Space-Time Patches]
    T_ENC --> COND[Conditioning]
end

subgraph DIT["DiT Backbone (Scaled)"]
    PATCH --> BLOCK1[DiT Block 1]
    BLOCK1 --> BLOCK2[DiT Block 2]
    BLOCK2 --> BLOCKN[DiT Block N...]
    COND --> BLOCK1
    COND --> BLOCK2
    COND --> BLOCKN
end

subgraph OUTPUT["Decoding"]
    BLOCKN --> UNPATCH[Unpatchify]
    UNPATCH --> V_DEC[VAE Decoder]
    V_DEC --> VIDEO[Generated Video]
end

classDef proc fill:#e3f2fd,stroke:#90caf9;
classDef data fill:#fff3e0,stroke:#ffcc02;
class T_ENC,V_ENC,PATCH,BLOCK1,BLOCK2,BLOCKN,UNPATCH,V_DEC proc;
class TEXT,IMG,VIDEO,COND data;

Key innovations:

  • Space-time patches: Video tokenized as 3D patches (like ViT but with temporal dimension)
  • Scaling laws: Quality improves predictably with compute — GPT-style scaling for video
  • Native resolution: Variable resolution/duration/aspect ratio in single model
  • Physical understanding: Emerges from scale — object permanence, physics, causality

Genie Architecture (DeepMind, 2024) — Action-Controllable World Model

Diagram: (Mermaid diagram - view source for diagram code)

flowchart TB
    subgraph VIDEO_TOK["Video Tokenizer (VQ-VAE)"]
        VID[Video Frames] --> VQ_ENC[Encoder]
        VQ_ENC --> LATENTS[Discrete Latents]
        LATENTS --> VQ_DEC[Decoder]
        VQ_DEC --> RECON[Reconstruction]
    end
subgraph LATENT_ACTION["Latent Action Model"]
    LATENTS --> LAM[Latent Action Encoder]
    LAM --> ACTIONS[Discrete Latent Actions]
    %% Learned from video WITHOUT action labels!
end

subgraph DYNAMICS["Dynamics Transformer (11B)"]
    ACTIONS --> DYN[Dynamics Transformer]
    LATENTS --> DYN
    DYN --> NEXT_LATENTS[Next Latents]
end

NEXT_LATENTS --> VQ_DEC

classDef tok fill:#e8f5e9,stroke:#c8e6c9;
classDef act fill:#fff3e0,stroke:#ffcc02;
classDef dyn fill:#f3e5f5,stroke:#ce93d8;
class VQ_ENC,VQ_DEC,LATENTS,RECON tok;
class LAM,ACTIONS act;
class DYN,NEXT_LATENTS dyn;

Breakthrough: Learns latent actions from unlabeled video only — no action annotations needed. The latent actions correspond to meaningful controllables (move left, jump, interact).


UniSim (DeepMind, 2024) — Universal Simulator

Component Specification
Data 1.6M trajectories (robotics, driving, gaming, internet video)
Architecture Masked Autoencoder + Diffusion Transformer
Control Action-conditioned generation
Transfer Policies trained in UniSim → real robot with minimal fine-tuning

Diffusion Policy for Video/Action Generation

Diagram: (Mermaid diagram - view source for diagram code)

flowchart LR
    subgraph COND["Conditioning"]
        OBS[Observation] --> ENC[Encoder]
        GOAL[Goal / Task] --> ENC
    end
subgraph DIFF["Diffusion Transformer (DiT)"]
    NOISE[Noise ε] --> DIT[DiT Block]
    ENC -->|Condition| DIT
    TIME[Timestep t] --> DIT
    DIT --> PRED[Predicted ε / v]
end

subgraph DECODE["Action/Video Decoding"]
    PRED --> DENOISE[Denoising Step]
    DENOISE --> ACTION[Action Sequence]:::action
    DENOISE --> VIDEO[Video Frames]:::video
end

classDef action fill:#e3f2fd,stroke:#90caf9;
classDef video fill:#fce4ec,stroke:#f8bbd0;


Temporal Consistency Challenges & Solutions

Challenge Solution Example
Flickering Latent diffusion + strong temporal attention Video LDM, Sora
Drifting identity Reference frame attention, identity loss VideoPoet, Gen-2
Physics violation World model structure, action conditioning Genie, UniSim
Long duration Hierarchical generation, Mamba/SSM dynamics Genie (1000+ frames)
Compute cost Consistency distillation (1-4 steps) LCM-Video, AnimateLCM

Evaluation Metrics

Metric What It Measures SOTA (2024)
FVD (Frechet Video Distance) Visual quality + temporal coherence Sora: ~100, Genie: ~150
FID (Frame-level) Per-frame quality Sora: ~15
CLIP Score Text-video alignment Sora: ~0.32
Action Accuracy Control fidelity (for world models) Genie: high, UniSim: policy transfer
Physics Benchmarks Physical plausibility Physics-IQA, CLEVRER

Applications in Embodied AI

Diagram: (Mermaid diagram - view source for diagram code)

graph LR
    VIDEO_GEN[Video Generation Models] --> WORLD_MODEL[World Models for Robotics]
    VIDEO_GEN --> SIM[Simulators / Digital Twins]
    VIDEO_GEN --> DATA_AUG[Data Augmentation for RL]
WORLD_MODEL --> PLANNING[Planning in Latent Space]
WORLD_MODEL --> POLICY[Policy Training]
WORLD_MODEL --> SIM2REAL[Sim-to-Real Transfer]

SIM --> DOMAIN_RAND[Domain Randomization at Scale]
SIM --> POLICY

DATA_AUG --> RL_TRAINING[More Diverse RL Training]


Compute Requirements (2024)

Model Params Training Compute Inference (16f, 480p)
Sora ~3-10B Massive (100K+ GPU-hours) ~30s on A100
Genie 11B Large ~5s/frame on A100
VideoPoet ~10B Large ~10s on TPU v5
Stable Video Diffusion 1.1B Moderate ~2s on RTX 4090
AnimateDiff / LCM-Video ~500M Low (distilled) Real-time on 4090

Future Directions

  1. Unified World Model / Video Generator — Single model for generation, planning, simulation
  2. Real-time Interactive Video — Consistency models + streaming for game engines
  3. 3D-Consistent Video — Gaussian splatting / NeRF integration for multi-view consistency
  4. Physics-Grounded Generation — Differentiable physics + diffusion for guaranteed plausibility
  5. Foundation World Models — "GPT for physics" pre-trained on all robotics + video data

Related Wiki Pages


Sources

  • world-models-diffusion-ai-robotics — Blog post: convergence deep-dive
  • source-genie — Genie paper (Bruce et al., DeepMind 2024)
  • source-unisim — UniSim paper (Yang et al., DeepMind 2024)
  • source-sora — Sora technical report (OpenAI 2024)