Plachtaa / seed-vc

State-of-the-Art zero-shot voice conversion & singing voice conversion with in context learning
GNU General Public License v3.0
637 stars 70 forks source link

Why "range covered by prompt are set to 0" in noised latents? #15

Open FacePoluke opened 1 month ago

FacePoluke commented 1 month ago

Thank you for your open-source work. I have a question: why are the parts of the noisy latents that are covered by the prompt set to 0? https://github.com/Plachtaa/seed-vc/blob/46f5848e266281501941aa78ea849012c43f5c65/modules/flow_matching.py#L133

Plachtaa commented 1 month ago

prompt component should be kept constant along different diffusion steps

FacePoluke commented 1 month ago

prompt component should be kept constant along different diffusion steps

you mean no noise should be added to the latents of the prompt component ?