Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

LTXV: Accumulate VAE decode results on intermediate_device (#12955)

Jukka Seppänen0904cc3fe5a5
diff --git a/comfy/ldm/lightricks/vae/causal_video_autoencoder.py b/comfy/ldm/lightricks/vae/causal_video_autoencoder.py
index 5b57dfc5e..9f14f64a5 100644
--- a/comfy/ldm/lightricks/vae/causal_video_autoencoder.py
+++ b/comfy/ldm/lightricks/vae/causal_video_autoencoder.py
@@ -11,6 +11,7 @@ from .causal_conv3d import CausalConv3d
from .pixel_norm import PixelNorm
from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
import comfy.ops
+import comfy.model_management
from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed
ops = comfy.ops.disable_weight_init
@@ -536,7 +537,7 @@ class Decoder(nn.Module):
mark_conv3d_ended(self.conv_out)
sample = self.conv_out(sample, causal=self.causal)
if sample is not None and sample.shape[2] > 0:
- output.append(sample)
+ output.append(sample.to(comfy.model_management.intermediate_device()))
return
up_block = self.up_blocks[idx]