Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

Merge branch 'vae-clean-new' into dev/v1.2

Shen-Chenhui8e37e7d76ec0
diff --cc opensora/models/vae/vae_temporal.py
index 473d057,fd5c7e8..bdf336f
--- a/opensora/models/vae/vae_temporal.py
+++ b/opensora/models/vae/vae_temporal.py
@@@ -369,20 -368,13 +369,20 @@@ class VAE_Temporal(nn.Module)
)
def get_latent_size(self, input_size):
- for i in range(len(input_size)):
- assert input_size[i] % self.patch_size[i] == 0, "Input size must be divisible by patch size"
- input_size = [input_size[i] // self.patch_size[i] for i in range(3)]
- return input_size
+ latent_size = []
+ for i in range(3):
+ if input_size[i] is None:
+ lsize = None
+ elif i == 0:
+ time_padding = self.time_downsample_factor - input_size[i] % self.time_downsample_factor
+ lsize = (input_size[i] + time_padding) // self.patch_size[i]
+ else:
+ lsize = input_size[i] // self.patch_size[i]
+ latent_size.append(lsize)
+ return latent_size
def encode(self, x):
- time_padding = self.time_downsample_factor - x.shape[2] % self.time_downsample_factor
+ time_padding = 0 if (x.shape[2] % self.time_downsample_factor == 0) else self.time_downsample_factor - x.shape[2] % self.time_downsample_factor
x = pad_at_dim(x, (time_padding, 0), dim=2)
encoded_feature = self.encoder(x)
moments = self.quant_conv(encoded_feature).to(x.dtype)