vae fix

Author: a-r-r-o-w

src/diffusers/models/autoencoders/autoencoder_kl_ltx.py

@@ -992,7 +992,9 @@ def __init__(
         # timestep embedding
         self.time_embedder = None
         self.scale_shift_table = None
+        self.timestep_scale_multiplier = None
         if timestep_conditioning:
+            self.timestep_scale_multiplier = nn.Parameter(torch.tensor(1000.0, dtype=torch.float32))
             self.time_embedder = PixArtAlphaCombinedTimestepSizeEmbeddings(output_channel * 2, 0)
             self.scale_shift_table = nn.Parameter(torch.randn(2, output_channel) / output_channel**0.5)
 
@@ -1001,6 +1003,9 @@ def __init__(
     def forward(self, hidden_states: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
         hidden_states = self.conv_in(hidden_states)
 
+        if self.timestep_scale_multiplier is not None:
+            temb = temb * self.timestep_scale_multiplier
+
         if torch.is_grad_enabled() and self.gradient_checkpointing:
             hidden_states = self._gradient_checkpointing_func(self.mid_block, hidden_states, temb)