T2V-Turbo 🚀

Our model is distilled from VideoCrafter2.

T2V-Turbo learns a LoRA on top of the base model by aligning to the reward feedback from HPSv2.1 and InternVid2 Stage 2 Model.

T2V-Turbo-v2 optimizes the training techniques by finetuning the full base model and further aligns to CLIPScore

T2V-Turbo trains on pure WebVid-10M data, whereas T2V-Turbo-v2 carufully optimizes different learning objectives with a mixutre of VidGen-1M and WebVid-10M data.

Moreover, T2V-Turbo-v2 supports to distill motion priors from the training videos.

Project page for T2V-Turbo 🥳

Project page for T2V-Turbo-v2 🤓

Running on CUDA 😀
2 14
0 0.5
4 50
16 48
0 2147483647
torch.dtype

Dtype for inference. Default is bf16.
0 100
Examples
Guidance scale Percentage of steps to apply motion guidance (v2 w/ MG only) Number of inference steps Number of Video Frames Seed Randomize seed torch.dtype