TIGER-AI-Lab / ConsistI2V

ConsistI2V: Enhancing Visual Consistency for Image-to-Video Generation (TMLR 2024)
https://tiger-ai-lab.github.io/ConsistI2V/
MIT License
217 stars 15 forks source link

memory #15

Closed 1585511010 closed 5 months ago

1585511010 commented 5 months ago

hallow,could you please tell me how much memory of cuda i should to prepare training

wren93 commented 5 months ago

Hi, for training 256x256 videos with 16 frames, we were able to use batch size=3 on a single A100 80G gpu. It takes about 71G memory during training for this setting.