README.md

metadata

tags:
  - video
  - text-to-video
  - video-diffusion
  - latents

revt2v-data

Precomputed latent dataset for revT2V (reverse-time video diffusion distillation). Each .pt file is one (prompt, seed) clip generated from the frozen ali-vilab/text-to-video-ms-1.7b teacher pipeline. Every record is a dict with:

• prompt -- the text prompt
• latents -- forward-time latents, shape (C, F, H, W)
• reversed_latents -- the same latents flipped along the frame axis (the training target)
• prompt_embeds / negative_prompt_embeds -- cached text-encoder outputs

Structure

• dataset1/ -- 154 shards (the original dataset)
• dataset2/ -- 416 shards + manifest.jsonl (prompt, seed, motion_score, kept)

Loading

from revt2v.data import LatentDataset

ds = LatentDataset("path/to/dataset1")  # or dataset2
record = ds[0]