From 1abec3e6ab422580139d2bdf44b73d43189059a9 Mon Sep 17 00:00:00 2001
From: Rattus <rattus128@gmail.com>
Date: Thu, 19 Mar 2026 10:29:51 +1000
Subject: [PATCH] ltx: vae: add cache state to downsample block

---
 .../vae/causal_video_autoencoder.py           | 42 +++++++++++++++----
 1 file changed, 33 insertions(+), 9 deletions(-)

diff --git a/comfy/ldm/lightricks/vae/causal_video_autoencoder.py b/comfy/ldm/lightricks/vae/causal_video_autoencoder.py
index f7aae26da..035f1ff36 100644
--- a/comfy/ldm/lightricks/vae/causal_video_autoencoder.py
+++ b/comfy/ldm/lightricks/vae/causal_video_autoencoder.py
@@ -737,12 +737,25 @@ class SpaceToDepthDownsample(nn.Module):
             causal=True,
             spatial_padding_mode=spatial_padding_mode,
         )
+        self.temporal_cache_state = {}
 
     def forward(self, x, causal: bool = True):
-        if self.stride[0] == 2:
+        tid = threading.get_ident()
+        cached, pad_first, cached_x, cached_input = self.temporal_cache_state.get(tid, (None, True, None, None))
+        if cached_input is not None:
+            x = torch_cat_if_needed([cached_input, x], dim=2)
+            cached_input = None
+
+        if self.stride[0] == 2 and pad_first:
             x = torch.cat(
                 [x[:, :, :1, :, :], x], dim=2
             )  # duplicate first frames for padding
+            pad_first = False
+
+        if x.shape[2] < self.stride[0]:
+            cached_input = x
+            self.temporal_cache_state[tid] = (cached, pad_first, cached_x, cached_input)
+            return None
 
         # skip connection
         x_in = rearrange(
@@ -757,15 +770,26 @@ class SpaceToDepthDownsample(nn.Module):
 
         # conv
         x = self.conv(x, causal=causal)
-        x = rearrange(
-            x,
-            "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
-            p1=self.stride[0],
-            p2=self.stride[1],
-            p3=self.stride[2],
-        )
+        if self.stride[0] == 2 and x.shape[2] == 1:
+            if cached_x is not None:
+                x = torch_cat_if_needed([cached_x, x], dim=2)
+                cached_x = None
+            else:
+                cached_x = x
+                x = None
 
-        x = x + x_in
+        if x is not None:
+            x = rearrange(
+                x,
+                "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
+                p1=self.stride[0],
+                p2=self.stride[1],
+                p3=self.stride[2],
+            )
+
+        cached = add_exchange_cache(x, cached, x_in, dim=2)
+
+        self.temporal_cache_state[tid] = (cached, pad_first, cached_x, cached_input)
 
         return x