Frame Interpolate can fail with CUDA input / CPU weight error

[nomadoor]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

Using the Frame Interpolate node should correctly interpolate frames (e.g. doubling FPS) without errors.

Actual Behavior

When using the Frame Interpolate node added in PR #13258, it can fail with this error:

RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same

It looks like the input tensor is on CUDA, while the interpolation model may be offloaded to CPU (possibly due to free_memory()).

I am not sure this is the correct fix, but the error goes away for me if the interpolation model is explicitly kept on device and moved to the correct dtype/device before inference:

-        model_management.free_memory(activation_mem, device)
+        model_management.free_memory(activation_mem, device, keep_loaded=[interp_model])
+        inference_model.to(device=device, dtype=dtype)

So this may be a device placement issue in the Frame Interpolate node.

Steps to Reproduce

FILM.json

Debug Logs

Requested to load FILMNet
loaded completely; 8128.68 MB usable, 65.68 MB loaded, full load: True
Frame interpolation:   0%|                                                                     | 0/120 [00:00<?, ?it/s]
!!! Exception during processing !!! Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same
Traceback (most recent call last):
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\execution.py", line 534, in execute
    output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                                                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\execution.py", line 334, in get_output_data
    return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\execution.py", line 308, in _async_map_node_over_list
    await process_inputs(input_dict, i)
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\execution.py", line 296, in process_inputs
    result = f(**inputs)
             ^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
    return method(locked_class, **inputs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\comfy_api\latest\_io.py", line 1826, in EXECUTE_NORMALIZED
    to_return = cls.execute(*args, **kwargs)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\comfy_extras\nodes_frame_interpolation.py", line 154, in execute
    feat_cache["img0"] = feat_cache.pop("next") if "next" in feat_cache else inference_model.extract_features(img0_single)
                                                                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\comfy_extras\frame_interpolation_models\film_net.py", line 221, in extract_features
    feature_pyramid = self.extract(image_pyramid)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\comfy_extras\frame_interpolation_models\film_net.py", line 99, in forward
    sub_pyramids = [self.extract_sublevels(image_pyramid[i], min(len(image_pyramid) - i, self.sub_levels))
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\comfy_extras\frame_interpolation_models\film_net.py", line 85, in forward
    head = layer(head)
           ^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\container.py", line 240, in forward
    input = module(input)
            ^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\comfy_extras\frame_interpolation_models\film_net.py", line 24, in forward
    x = self.conv(x)
        ^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\ComfyUI\comfy\ops.py", line 428, in forward
    return super().forward(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\conv.py", line 554, in forward
    return self._conv_forward(input, self.weight, self.bias)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\conv.py", line 549, in _conv_forward
    return F.conv2d(
           ^^^^^^^^^
RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same

Other

ComfyUI 0.19.3
ComfyUI_frontend v1.42.15

System Info

OS: win32
Python Version: 3.12.10 (tags/v3.12.10:0cc8128, Apr 8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
Embedded Python: true
Pytorch Version: 2.7.1+cu128
Arguments: ComfyUI\main.py --disable-auto-launch --reserve-vram 2.0 --disable-smart-memory
RAM Total: 95.87 GB
RAM Free: 40.42 GB
Templates Version: 0.9.62

Devices

• cuda:0 NVIDIA GeForce RTX 4070 Ti : cudaMallocAsync (cuda)
  VRAM Total: 11.99 GB
  VRAM Free: 10.74 GB
  Torch VRAM Total: 32 MB
  Torch VRAM Free: 23.88 MB

[meimei-arch]

Pytorch Version: 2.7.1+cu128 - you should be running cu130 , not cu128..... on that gpu, and don't just take my word for it - the main page to this very project dictates for Nvidia users to use cu130 now. or cu132 for nightly if you're feeling adventurous.

[tuolaku]

Pytorch 版本：2.7.1+cu128——你应该运行的是 cu130，而不是 cu128.....关于那块显卡，别只听我说——这个项目的主页明确要求Nvidia用户现在使用cu130。如果你想冒险，晚上也可以用cu132。

Environment:

• torch: 2.12 nightly
• torchvision: 0.27.0
• torchaudio: 2.11.0
• Python: 3.13
• CUDA: 13.2

I have the same problem.

[meimei-arch]

Pytorch 版本：2.7.1+cu128——你应该运行的是 cu130，而不是 cu128.....关于那块显卡，别只听我说——这个项目的主页明确要求Nvidia用户现在使用cu130。如果你想冒险，晚上也可以用cu132。

Environment:

* torch: 2.12 nightly

* torchvision: 0.27.0

* torchaudio: 2.11.0

* Python: 3.13

* CUDA: 13.2

I have the same problem.

Python 3.13 can sometimes act funny at least in my experience. would recommend you drop down to Python 3.12 . Try that - if it still acts funny - then swap out nightly cuda for stable . (currently i personally run stable) cu130. and have no issues in combination with Python 3.12.

[rattus128]

you should only need the first line of your change. Can you try with that second line dropped?

I cant repro this but the logic is sound. Just that first line should fix it for you and if you can confirm ill submit the fix.

[nomadoor]

@rattus128

I tried changing only the first line:

-        model_management.free_memory(activation_mem, device)
+        model_management.free_memory(activation_mem, device, keep_loaded=[interp_model])

However, after restarting ComfyUI, I still get the same error:

Requested to load FILMNet
loaded completely; 8128.68 MB usable, 65.68 MB loaded, full load: True
RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same

At least on my setup, the second line still seems to be required:

inference_model.to(device=device, dtype=dtype)

With both lines applied, the workflow runs successfully.

[tuolaku]

Pytorch 版本：2.7.1+cu128——你应该运行的是 cu130，而不是 cu128.....关于那块显卡，别只听我说——这个项目的主页明确要求Nvidia用户现在使用cu130。如果你想冒险，晚上也可以用cu132。

环境：

* torch: 2.12 nightly

* torchvision: 0.27.0

* torchaudio: 2.11.0

* Python: 3.13

* CUDA: 13.2

我也有同样的问题。

至少以我的经验，Python 3.13 有时会表现得怪怪的。建议你降级到Python 3.12。试试这个方法——如果它还是表现怪异——那就把每晚的 cuda 换成稳定的 。（我目前个人运行稳定版）Cu130。并且与 Python 3.12 结合使用时没有问题。

The latest version of PyTorch, combined with CUDA 13.2, offers a significant boost in inference speed. Given the substantial improvement, I probably won't sacrifice that performance just because of this issue. ：D

[kijai]

This should be the proper fix for the root of the issue: #13698

I admit it's on me to not properly test without dynamic VRAM initially.

[nomadoor]

@kijai
I tested this on my setup, and it seems to be working fine now.
Thank you for the fix!

[kijai]

@kijai I tested this on my setup, and it seems to be working fine now. Thank you for the fix!

Thank you for testing and bringing this to our attention!