Reduce artifacts on Wan by doing the patch embedding in fp32.

Author: comfyanonymous

comfy/ldm/wan/model.py

@@ -18,7 +18,7 @@ def sinusoidal_embedding_1d(dim, position):
     # preprocess
     assert dim % 2 == 0
     half = dim // 2
-    position = position.type(torch.float64)
+    position = position.type(torch.float32)
 
     # calculation
     sinusoid = torch.outer(
@@ -353,7 +353,7 @@ def __init__(self,
 
         # embeddings
         self.patch_embedding = operations.Conv3d(
-            in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+            in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=torch.float32)
         self.text_embedding = nn.Sequential(
             operations.Linear(text_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
             operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
@@ -411,7 +411,7 @@ def forward_orig(
                 List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
         """
         # embeddings
-        x = self.patch_embedding(x)
+        x = self.patch_embedding(x.float()).to(x.dtype)
         grid_sizes = x.shape[2:]
         x = x.flatten(2).transpose(1, 2)