In
https://github.com/RockeyCoss/SPO/blob/main/step_aware_preference_model/trainer/tasks/spm_task.py #L370-L421
Why do we need to do a denoise step using UNET when we try to get a noisy version of the image?
Isn't
image -> vae encode -> add noise -> vae decode
better than
image -> vae encode -> add noise ->unet denoise -> vae decode ?
In
https://github.com/RockeyCoss/SPO/blob/main/step_aware_preference_model/trainer/tasks/spm_task.py #L370-L421
Why do we need to do a denoise step using UNET when we try to get a noisy version of the image?
Isn't
image -> vae encode -> add noise -> vae decode
better than
image -> vae encode -> add noise ->unet denoise -> vae decode ?