jason-yoo-108/lifelong-drive

Lifelong Drive

One of the five single-video-stream datasets introduced in Lifelong Learning of Video Diffusion Models From a Single Video Stream (arXiv:2406.04814). Code: https://github.com/plai-group/lifelong-vdm.

Continuous dashcam footage of a 1,675 km drive from Chongqing to Shanghai. The drive spans mountains, cities and tunnels, sunny, cloudy and rainy weather, and other vehicles moving at different speeds: real-world perceptual data from a single correlated and non-stationary stream. The drive is nearly continuous; rare, subtle discontinuities occur where the driver took breaks (the video resumes at the same or a nearby highway location).

The repository holds both the 512 x 512 RGB video and the pre-computed Stable Diffusion latents the paper's models were trained on.

	Train stream	Test stream
Frames	1,000,000	100,000
Source video	512 x 512, 20 FPS (~13.9 h), train/LifelongDriveTrain.mp4	test/LifelongDriveTest.mp4
Latents	4 x 64 x 64 per frame, 2,000 x 500 frames	200 x 500 frames

The test stream is the 100,000 frames that follow the train stream in the same video.

Files

train/LifelongDriveTrain.mp4
train/encoded_video/batch_0000.pt … batch_1999.pt
test/LifelongDriveTest.mp4
test/encoded_video/batch_0000.pt … batch_0199.pt

Each batch_XXXX.pt is a torch.saved dict covering 500 consecutive frames (batch_i = stream frames i*500 … i*500+499):

key	shape	dtype	meaning
quantized_latents	(500, 4, 64, 64)	uint8	per-frame 8-bit quantized VAE latents
min_vals	(500,)	float16	per-frame dequantization offset
scales	(500,)	float16	per-frame dequantization scale

Latents were produced by datasets/stable_diffusion_encoder.py in the code repository: frames are mapped to [-1, 1], encoded with the SDXL VAE madebyollin/sdxl-vae-fp16-fix (latent = vae.encode(x).latent_dist.sample() * 0.13025), then quantized per frame with q = round((latent - min) / scale), scale = (max - min) / 255.

Loading

import torch
d = torch.load("train/encoded_video/batch_0000.pt", map_location="cpu")
q, lo, sc = d["quantized_latents"], d["min_vals"], d["scales"]
latents = q.float() * sc.float()[:, None, None, None] + lo.float()[:, None, None, None]  # (500, 4, 64, 64)

# optional: decode back to RGB with diffusers
from diffusers import AutoencoderKL
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix")
rgb = (vae.decode(latents[:4] / 0.13025).sample.clamp(-1, 1) + 1) / 2  # (4, 3, 512, 512) in [0, 1]

To use the dataset with the paper's code (it expects the directory at datasets/drive):

pip install -U huggingface_hub
hf download jason-yoo-108/lifelong-drive --repo-type dataset --local-dir datasets/drive   # add --exclude "*.mp4" to skip the video

Provenance and license

The footage comes from the YouTube video "1674.8 kilometers of long-distance travel across half of China - driving from Chongqing to Shanghai" by 中国街景 ChinaStreetView (2023). It was cropped and downsampled to 512 x 512 at 20 FPS; the first 1,000,000 frames form the train stream and the next 100,000 the test stream (recipe in the datasets/mp4_to_npy.py docstring of the code repository).

The derived video and latents are released under CC BY 4.0 for research use. The underlying footage belongs to its original creator; please credit the source channel when redistributing.

Citation

@article{yoo2024lifelong,
  title   = {Lifelong Learning of Video Diffusion Models From a Single Video Stream},
  author  = {Yoo, Jason and He, Yingchen and Naderiparizi, Saeid and Green, Dylan and van de Ven, Gido M. and Pleiss, Geoff and Wood, Frank},
  journal = {arXiv preprint arXiv:2406.04814},
  year    = {2024}
}