Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

lecam config

Shen-Chenhui16b0149fc4e7
diff --git a/configs/vae_magvit_v2/train/16x128x128_z4_LC.py b/configs/vae_magvit_v2/train/16x128x128_z4_LC.py
index 81da8d5..9310257 100644
--- a/configs/vae_magvit_v2/train/16x128x128_z4_LC.py
+++ b/configs/vae_magvit_v2/train/16x128x128_z4_LC.py
@@ -55,7 +55,7 @@ generator_factor = 0.1 # SCH: generator adversarial loss, MAGVIT v2 uses 0.1
lecam_loss_weight = 0.001 # NOTE: MAVGIT v2 use 0.001
discriminator_loss_type="non-saturating"
generator_loss_type="non-saturating"
-discriminator_start = 1000 # 50000 NOTE: change to correct val, debug use -1 for now
+discriminator_start = -1 # 50000 NOTE: change to correct val, debug use -1 for now
gradient_penalty_loss_weight = None # 10 # SCH: MAGVIT uses 10, opensora plan doesn't use
ema_decay = 0.999 # ema decay for LeCam
diff --git a/opensora/models/vae/README.md b/opensora/models/vae/README.md
index 73099ac..daa16f9 100644
--- a/opensora/models/vae/README.md
+++ b/opensora/models/vae/README.md
@@ -1,5 +1,11 @@
## Commands
+## 0. References
+
+* https://github.com/google-research/magvit
+* https://github.com/CompVis/taming-transformers
+* https://github.com/adobe/antialiased-cnns/pull/39/commits/3d6f02b6943c58b68c19c07bc26fad57492ff3bc
+
## 1. VAE 3D
### 1.1 Train