BLOBER

The official code for the paper:

GLOBER: Coherent Non-autoregressive Video Generation via Global Guided Video Decoder

ARXIV DEMO

Conda Environment

conda env create -f environment.yaml
pip install -r requirements.txt

Train Script

# Train scripts for both the auto-encoder and generator are the same
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python main.py --base $CFG --logdir experiments/ # --ckpt path/to/ckpt

Sample&Evaluation Scripts

We follow the implementation of StyleGAN-V(https://github.com/universome/stylegan-v) for evaluation.

# AutoEncoder
bash scripts/script_for_sample_3c.sh $CUR $CUDA $TOTAL $CFG $EXP $PTH $UC_FRAME $UC_VIDEO $UC_DOMAIN
bash scripts/script_for_fvd_3c.sh $EXP $UCFRAME $UCVID $UCDOMAIN $PTH $CUDA

# Generator
bash scripts/script_for_sample.sh $CFG $EXP $PTH $UC $CUR $TOTAL $CUDA
bash scripts/script_for_fvd.sh $EXP $UC $PTH $CUDA

Checkpoints

Will be released soon.

Test generation speed of prior methods

VIDM: \url{https://github.com/MKFMIKU/vidm}

VDM: \url{https://github.com/lucidrains/video-diffusion-pytorch}

VideoFusion: \url{https://huggingface.co/docs/diffusers/main/en/api/pipelines/text_to_video}

TATS: \url{https://github.com/SongweiGe/TATS}