search

IanYeung / MGLD-VSR

Code for ECCV 2024 Paper "Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution"
Other
84 stars 1 forks source link

Question about Generating Training Data #12

Open Ree1s opened 4 months ago

Ree1s commented 4 months ago

Thank you for your impactful work. I'm interested in learning more about how the training data is generated. Could you explain which vqgan model is used to encode and decode low-quality frames in the script vsr_val_ddpm_text_T_vqganfin_w_latent.py? The video_vae_cfw.ckpt can only be saved after the second training stage?

IanYeung commented 4 months ago

Thank you for your impactful work. I'm interested in learning more about how the training data is generated. Could you explain which vqgan model is used to encode and decode low-quality frames in the script vsr_val_ddpm_text_T_vqganfin_w_latent.py? The video_vae_cfw.ckpt can only be saved after the second training stage?

I use the original one. The video_vae_cfw.ckpt is obtained after second stage training.

Ree1s commented 4 months ago

Thank you for your reply. Which is the original one? Is it saved during the first stage training?