Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

register module

Shen-Chenhuid8666a73c41e
diff --git a/configs/vae_magvit_v2/train/17x128x128.py b/configs/vae_magvit_v2/train/17x128x128.py
index e315b3e..ba08169 100644
--- a/configs/vae_magvit_v2/train/17x128x128.py
+++ b/configs/vae_magvit_v2/train/17x128x128.py
@@ -23,14 +23,12 @@ perceptual_weight = 0.1 # TODO: need to change this to 0.1 !!! according to MAGV
model = dict(
type="VAE_MAGVIT_V2",
in_out_channels = 3,
- latent_embed_dim = 16,
+ latent_embed_dim = 256,
filters = 128,
num_res_blocks = 4,
channel_multipliers = (1, 2, 2, 4),
temporal_downsample = (False, True, True),
num_groups = 32, # for nn.GroupNorm
- conv_downsample = False,
- upsample = "nearest+conv", # options: "deconv", "nearest+conv"
kl_embed_dim = 64,
custom_conv_padding = None,
activation_fn = 'swish',
diff --git a/opensora/models/vae/__init__.py b/opensora/models/vae/__init__.py
index 9eed101..1dbf956 100644
--- a/opensora/models/vae/__init__.py
+++ b/opensora/models/vae/__init__.py
@@ -1,3 +1,4 @@
from .vae import VideoAutoencoderKL, VideoAutoencoderKLTemporalDecoder
-from .vae_3d import VAE_3D
+from .vae_3d import VAE_3D, VAE_3D_B
+from .vae_3d_v2 import VAE_3D_V2, VAE_MAGVIT_V2
diff --git a/opensora/models/vae/vae_3d_v2.py b/opensora/models/vae/vae_3d_v2.py
index e72f196..a067304 100644
--- a/opensora/models/vae/vae_3d_v2.py
+++ b/opensora/models/vae/vae_3d_v2.py
@@ -8,7 +8,6 @@ from numpy import typing as nptyping
from opensora.models.vae import model_utils
from opensora.registry import MODELS
from opensora.utils.ckpt_utils import load_checkpoint
-from einops.layers.torch import Rearrange # NOTE: this is a torch module
from einops import rearrange
"""Encoder and Decoder stuctures with 3D CNNs."""