febf3ad4b2
* upload v2.0 * update docs * [hotfix] fit latest fa3 (#802) * update readme * update readme * update readme * update train readme * update readme * update readme: motion score * cleaning video dc ae WIP * update config * add dependency functions * undo cleaning * use latest dcae * complete high compression training * update hcae config * cleaned up vae * update ae.md * further cleanup * update vae & ae paths * align naming of ae * [hotfix] fix ring attn bwd for fa3 (#803) * train ae default without wandb * update config * update evaluation results * added hcae report * update readme * update readme demo * update readme demo * update readme gif * display demo directly in readme * update paper * delete files --------- Co-authored-by: Hongxin Liu <lhx0217@gmail.com> Co-authored-by: Shen-Chenhui <shen_chenhui@u.nus.edu> Co-authored-by: wuxiwen <wuxiwen.simon@gmail.com>
1.8 KiB
10× inference speedup with high-compression autoencoder
The high computational cost of training video generation models arises from the large number of tokens and the dominance of attention computation. To further reduce training expenses, we explore training video generation models with high-compression autoencoders (Video DC-AEs). As shown in the comparason below, by switching to the Video DC-AE with a much higher downsample ratio (4 x 32 x 32), we can afford to further reduce the patch size to 1 and still achieve 5.2× speedup in training throughput and 10x speedup during inference:
Nevertheless, despite the advantanges in drastically lower computation costs, other challenges remain. For instance, larger channels low down convergance. Our generation model adapted with a 128-channel Video DC-AE for 25K iterations achieves a loss level of 0.5, as compared to 0.1 from the initialization model. While the fast video generation model underperforms the original, it still captures spatial-temporal relationships. We release this model to the research community for further exploration.
Checkout more details in our report.
Inference
To inference on our fast video generation model:
torchrun --nproc_per_node 1 --standalone scripts/diffusion/inference.py configs/diffusion/inference/high_compression.py --prompt "The story of a robot's life in a cyberpunk setting."
Training
Follow this guide to parepare the DATASET for training. Then, you may train your own fast generation model with the following command:
torchrun --nproc_per_node 8 scripts/diffusion/train.py configs/diffusion/train/high_compression.py --dataset.data-path datasets/pexels_45k_necessary.csv