test-time-training/ttt-video-dit

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

2448 stars · 9 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Official PyTorch implementation of 'One-Minute Video Generation with Test-Time Training', which finetunes the CogVideoX 5B diffusion transformer with TTT layers for long-range video generation and style transfer. It includes training and inference code for generating 63-second videos from text storyboards.

Use cases

• generate one-minute videos from text storyboards
• finetune a diffusion transformer for video style transfer
• extend video generation context beyond a few seconds
• experiment with test-time training layers in transformers
• reproduce research results comparing TTT layers to Mamba 2 and sliding-window attention

When to choose

• you need to generate long, multi-scene videos with strong temporal consistency
• you want to experiment with TTT layers in diffusion transformers
• you have H100 GPUs and want to train or sample from the research model

When to avoid

• you lack H100-class GPUs or CUDA 12.3+ for the TTT-MLP kernel
• you need production-ready, artifact-free video generation
• you only need short 3-second clips from the pretrained CogVideoX model

Facets

library · maturity experimental

deep-learning machine-learning video-processing llm-training deep-learning machine-learning artificial-intelligence python diffusion-transformer text-to-video test-time-training cogvideox research-code pytorch video gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="test-time-training/ttt-video-dit")