search

G-U-N / AnimateLCM

AnimateLCM: Let's Accelerate the Video Generation within 4 Steps!
https://animatelcm.github.io
MIT License
567 stars 41 forks source link

Asking about SVDSolver #33

Open dreamyou070 opened 1 month ago

dreamyou070 commented 1 month ago

when you train LCM_svd, you set svd_solver like,

svd_solver = SVDSolver(args.N, noise_scheduler.config.sigma_min, noise_scheduler.config.sigma_max, 7,0.7, 1.6)

why you change training timestep to args.N (for example 40), can I get advice about the SVDSolver ..?

G-U-N commented 1 month ago

The SVD is built with EDM framework, which is a continuous diffusion format. Yet the consistency distillation requires us to set discrete timesteps for distillation. You might refer to https://github.com/luosiallen/latent-consistency-model/blob/main/LCM_Training_Script/consistency_distillation/train_lcm_distill_sd_wds.py if you are building things with DDPM framework.