TeaCache
TeaCache copied to clipboard
ali-vilab
Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model
I see the examples on how to run it on one GPU but when I try something like this: `torchrun --nproc_per_node=8 teacache_generate.py --task t2v-14B --size 1280*720 --ckpt_dir ./Wan2.1-T2V-14B --dit_fsdp --t5_fsdp...
In teacache_forward(): Here is the accumulated metric in code: `self.accumulated_rel_l1_distance_even += rescale_func(((modulated_inp-self.previous_e0_even).abs().mean() / self.previous_e0_even.abs().mean()).cpu().item())` The modulated input is defined here: `modulated_inp = e0 if self.use_ref_steps else e` Where e and...
> Hi, [@zishen-ucap](https://github.com/zishen-ucap). Your observation is right. For some model, timestep embedding shows better correlation, e.g., Wan2.1 and CogVideoX. With timestep embedding before 'time_projection', it is unnecessary to to set...
