⚡ ViRDM · 4-step · Causal Video

VBench Total 84.87 Peak training memory 48.3 GB/GPU Post-training ~2 h · 8 × A100 20 generator updates

Also supports training on a single 80GB GPU.

📄 Paper 🌐 Project ⌘ Code 🤗 Model
Examples

This demo uses a four-block startup buffer for smoother streaming on Hugging Face ZeroGPU; local deployments can use one block when generation keeps pace.