aljaz/generative-models
1
0
Fork 0
mirror of https://github.com/Stability-AI/generative-models.git synced 2025-12-19 22:34:22 +01:00
Code Issues Packages Projects Releases Wiki Activity

Stable Video Diffusion

Browse Source
This commit is contained in:
Tim Dockhorn
2023-11-21 10:40:21 -08:00
parent 477d8b9a77
commit 059d8e9cd9
59 changed files with 5463 additions and 1691 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
sgm/modules/diffusionmodules/denoiser_scaling.py
View File

@@ -1,11 +1,24 @@
from abc import ABC, abstractmethod
from typing import Tuple
import torch
class DenoiserScaling(ABC):
@abstractmethod
def __call__(
self, sigma: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
pass
class EDMScaling:
def __init__(self, sigma_data=0.5):
def __init__(self, sigma_data: float = 0.5):
self.sigma_data = sigma_data
def __call__(self, sigma):
def __call__(
self, sigma: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
c_skip = self.sigma_data**2 / (sigma**2 + self.sigma_data**2)
c_out = sigma * self.sigma_data / (sigma**2 + self.sigma_data**2) ** 0.5
c_in = 1 / (sigma**2 + self.sigma_data**2) ** 0.5
@@ -14,7 +27,9 @@ class EDMScaling:
class EpsScaling:
def __call__(self, sigma):
def __call__(
self, sigma: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
c_skip = torch.ones_like(sigma, device=sigma.device)
c_out = -sigma
c_in = 1 / (sigma**2 + 1.0) ** 0.5
@@ -23,9 +38,22 @@ class EpsScaling:
class VScaling:
def __call__(self, sigma):
def __call__(
self, sigma: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
c_skip = 1.0 / (sigma**2 + 1.0)
c_out = -sigma / (sigma**2 + 1.0) ** 0.5
c_in = 1.0 / (sigma**2 + 1.0) ** 0.5
c_noise = sigma.clone()
return c_skip, c_out, c_in, c_noise
class VScalingWithEDMcNoise(DenoiserScaling):
def __call__(
self, sigma: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
c_skip = 1.0 / (sigma**2 + 1.0)
c_out = -sigma / (sigma**2 + 1.0) ** 0.5
c_in = 1.0 / (sigma**2 + 1.0) ** 0.5
c_noise = 0.25 * sigma.log()
return c_skip, c_out, c_in, c_noise