Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

debug

Shen-Chenhuif867ebc8198f
diff --git a/configs/vae_magvit_v2/train/16x128x128.py b/configs/vae_magvit_v2/train/16x128x128.py
index 722f5cb..5be8214 100644
--- a/configs/vae_magvit_v2/train/16x128x128.py
+++ b/configs/vae_magvit_v2/train/16x128x128.py
@@ -30,7 +30,7 @@ model = dict(
kl_embed_dim = 64,
activation_fn = 'swish',
separate_first_frame_encoding = False,
- disable_space = True,
+ disable_space = False,
custom_conv_padding = None
)
diff --git a/configs/vae_magvit_v2/train/pipeline_16x128x128.py b/configs/vae_magvit_v2/train/pipeline_16x128x128.py
index e86c2cc..722f5cb 100644
--- a/configs/vae_magvit_v2/train/pipeline_16x128x128.py
+++ b/configs/vae_magvit_v2/train/pipeline_16x128x128.py
@@ -30,6 +30,7 @@ model = dict(
kl_embed_dim = 64,
activation_fn = 'swish',
separate_first_frame_encoding = False,
+ disable_space = True,
custom_conv_padding = None
)
diff --git a/opensora/models/vae/vae_3d_v2.py b/opensora/models/vae/vae_3d_v2.py
index 0ff219b..05ee2a8 100644
--- a/opensora/models/vae/vae_3d_v2.py
+++ b/opensora/models/vae/vae_3d_v2.py
@@ -617,7 +617,7 @@ class Decoder(nn.Module):
self.num_groups = num_groups
# self.upsample = upsample
- self.s_stride = 1 if self.disable_spatial_upsample else 2 # spatial stride
+ self.s_stride = 1 if disable_spatial_upsample else 2 # spatial stride
self.custom_conv_padding = custom_conv_padding
# self.norm_type = self.config.vqvae.norm_type
# self.num_remat_block = self.config.vqvae.get('num_dec_remat_blocks', 0)