diffusion model video generation

**Video Generation with Diffusion Models: Temporal Coherence and Scaling — generating minutes of high-quality video via latent diffusion** Video generation extends image diffusion models to spatiotemporal domains, enabling minute-long generation with consistent characters and physics. Sora (OpenAI, 2024) demonstrates billion-parameter diffusion transformers for video. **Spatiotemporal Diffusion Architecture** 3D U-Net/3D attention: extend 2D convolutions to 3D by adding temporal dimension (depth). Spatiotemporal attention: attend across spatial + temporal dimensions jointly (expensive—quadratic in resolution and frames). Factorized attention: alternately apply spatial (per-frame) and temporal (frame-to-frame) attention, reducing complexity. Timestep conditioning: denoise-step t guides generation—gradually refining video from noise. **Sora: Scaling to Videos** Sora (OpenAI, 2024): diffusion transformer (DiT) architecture. Key insights: (1) Video tokenizer compresses video to lower-dimensional latent space (VQ-VAE-style compression—96x reduction: from 1280×720 pixels to 16×9 tokens, key missing detail: temporal compression factor); (2) Large transformer (billions of parameters) denoises latent video representation; (3) Training on vast video dataset (proprietary); (4) Inference: iterative denoising generates consistent, hour-length videos (claimed, unverified). User prompts: text→video via text conditioning (CLIP embeddings or similar). **Temporal Consistency Challenge** Naive frame-by-frame generation lacks temporal consistency (flicker, jitter, physical implausibility). Solutions: (1) optical flow guidance (enforce consistency with flow), (2) temporal attention (attending to previous frames), (3) latent diffusion (compression reduces high-frequency flicker artifacts), (4) world model pre-training (learn persistent object representations). **Video Tokenizers and Compression** MAGVIT (Masked Generative Video Tokenization): tokenizes video frames + temporal differences into discrete tokens (vocabulary size 4096+). CogVideoX (THUDM) uses similar compression. Compression: 1280×720×48 frames (RGB 8-bit) → 64×40×48 tokens (16-bit indices) = 1000x reduction. Decompression: token→VAE decoder→RGB video. **Open Models** HunyuanVideo (Tencent), CogVideoX (Tsinghua), Wan 2.1 (Microsoft/Alibaba) provide open alternatives to Sora. Evaluation: FVD (Fréchet Video Distance, temporal-aware FID), FID on key frames, human preference studies. Training compute: 10-100 PFLOP-days for billion-parameter models—accessible only to large labs. Inference: ~1 minute per 10-second video on single GPU (slow, suggests deployment challenges).