diff --git a/trellis/pipelines/trellis_image_to_3d.py b/trellis/pipelines/trellis_image_to_3d.py
index a3a7e46..49dff10 100644
--- a/trellis/pipelines/trellis_image_to_3d.py
+++ b/trellis/pipelines/trellis_image_to_3d.py
@@ -985,8 +985,10 @@ class TrellisVGGTTo3DPipeline(TrellisImageTo3DPipeline):
         else:
             raise ValueError(f"Unsupported type of image: {type(image)}")
         
-        if self.low_vram:
-            self.VGGT_model.to(self.device)
+        # FIX: always move VGGT to the pipeline device before use. Previously this
+        # was gated on self.low_vram, so with low_vram=False the model stayed on CPU
+        # while images were on cuda -> "tensors on cuda:0 and cpu" device mismatch.
+        self.VGGT_model.to(self.device)
         with torch.no_grad():
             with torch.cuda.amp.autocast(dtype=self.VGGT_dtype):
                 # Predict attributes including cameras, depth maps, and point maps.