From 4fc4159c88c60703d2f7b4af7f031511cc6bc5f8 Mon Sep 17 00:00:00 2001
From: kijai <40791699+kijai@users.noreply.github.com>
Date: Sun, 27 Apr 2025 17:22:02 +0300
Subject: [PATCH] fix up default rope dtype

---
 wanvideo/modules/model.py | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/wanvideo/modules/model.py b/wanvideo/modules/model.py
index bd3d287..da05ead 100644
--- a/wanvideo/modules/model.py
+++ b/wanvideo/modules/model.py
@@ -118,7 +118,7 @@ def rope_apply(x, grid_sizes, freqs):
 
         # append to collection
         output.append(x_i)
-    return torch.stack(output).float()
+    return torch.stack(output).to(x.dtype)
 
 
 class WanRMSNorm(nn.Module):
@@ -1021,7 +1021,8 @@ class WanModel(ModelMixin, ConfigMixin):
         camera_embed=None,
         unianim_data=None,
         fps_embeds=None,
-        fun_ref = None
+        fun_ref = None,
+        fun_camera=None,
     ):
         r"""
         Forward pass through the diffusion model
@@ -1073,6 +1074,10 @@ class WanModel(ModelMixin, ConfigMixin):
             for u in x
             ]
 
+        if self.control_adapter is not None and fun_camera is not None:
+            fun_camera = self.control_adapter(fun_camera)
+            x = [u + v for u, v in zip(x, fun_camera)]
+
         grid_sizes = torch.stack(
             [torch.tensor(u.shape[2:], dtype=torch.long) for u in x])