Rather interpolate any Uni3C input frame count differences

2025-09-10 16:47:13 +03:00
parent dd36ab46c6
commit c254ae4235
1 changed files with 1 additions and 1 deletions
@@ -1975,7 +1975,7 @@ class WanModel(torch.nn.Module):
            if hidden_states.shape[1] == 16: #T2V work around
                hidden_states = torch.cat([hidden_states, torch.zeros_like(hidden_states[:, :4])], dim=1)
            if render_latent.shape[2] != hidden_states.shape[2]:
-                render_latent = torch.cat([render_latent, render_latent[:, :, :1]], dim=2)
+                render_latent = torch.nn.functional.interpolate(render_latent, size=(hidden_states.shape[2], render_latent.shape[3], render_latent.shape[4]), mode='trilinear', align_corners=False)
            render_latent = torch.cat([hidden_states[:, :20], render_latent], dim=1)

        # embeddings