June 10, Update: stmfnet
Just to save folks a headache the file is available for download as a model now.
June 06, Update: V4.0
Just a minor improvement, using a double interpolation of SFTMVFI results in higher fidelity of dynamic or fast motion scenes. Video is manually loaded to reduce time when generating high quantities of test runs. Now you can select which videos you want to upscale, will copy the name of the original input video.
Included the context framing node from Kijai's wrapper. Usually leave this off will exponentially increase generation times if on; however, can let you generate much longer videos beyond what your vram will allow you.
Interpolation requires higher VRAM if scaling if greater than 1.0. Changed the scaling factor to float for more control.
Updated nodes if there is any issues feel free to let me know.
https://drive.google.com/file/d/1s5JJdt5X69AO2E2uuaes17aPwlWIQagG/view
Place it in the folder: Custom_nodes\ComfyUI-Frame-Interpolation\ckpts\stmfnet
April 14, Update: V3.3
Updated Nodes.
March 27, Update: V3.2
Small changes to the Torch Compile and Tea Cache for more dynamic outputs using my Lora's.
Workflow is suitable for a 24GB card using the 14B models and a 12GB card using the 1.3B model.
Updates:
***Error within the naming of outputs corrected.
Uploaded a very simple and straightforward Wan I2V workflow. The workflow includes an upscaler with frame interpolation to bring your videos to a cool and smooth 60FPS. No BS just a simple addition. Original WanI2V is based on the workflow example of the amazing work from Kijai linked here: https://github.com/kijai/ComfyUI-WanVideoWrapper/tree/main/example_workflows.
Additionally, I've left the annotations within.
Prev Updates:
Compatible with 12GB memory cards. Tested using a 3060 RTX 12GB using this workflow can comfortably go up to 85 frames. 97 Frames is pushing the absolute limits will take about an hour to generate.
Due to demand this workflow is a small improvement to the Super Simple HV2V for those who don't do postprocessing outside of Comfyui.
Designed for use with Hunyuan Video and trained Lora models, it can upscale videos to true 60 FPS. It combines detailed notes on denoising strength, step optimization, aspect ratio adjustments, frame interpolation, and upscaling to create high-quality, smooth videos.
Order matters:
Generate → Upscale → Interpolate. Upscaling first preserves detail for interpolation.
Adjust parameters based on your Lora’s behavior and hardware limits.
Description
stmfnet file folder.
FAQ
Comments (8)
I keep getting this error when trying to run the workflow. I use latest version of ComfyUI, all nodes installed and updated and I use the same models as you. Using an RTX 3090.
WanVideoSampler
Which node is stmfnet used in? I couldn't find a way to select it.
STEMFNet VFI
if you are using comfyUI manager, search for and install this custom node:
ComfyUI-Frame-Interpolation
or git clone it
https://github.com/Fannovel16/ComfyUI-Frame-Interpolation
after successful installation, put the file here (stmfnet.pth is the file, not a folder obviously)
custom_nodes\comfyui-frame-interpolation\ckpts\stmfnet\stmfnet.pth
in comfy load this node: STEMFNet VFI
Getting an error when it hits the sampler:
Attempted to inline function marked as skipped Explanation: Dynamo developers have intentionally marked that the function WanAttentionBlock.get_mod should not be traced. Hint: Avoid calling the function WanAttentionBlock.get_mod.
i am getting an error in the workflow for WanVideoSampler
Invalid context window indices [7, 9, 11, 13, 15, 17, 7, 9, 11] for latent_model_input with shape torch.Size([16, 12, 40, 32])
I don't get it where do I put the stmfnet?
if you are using comfyUI manager, search for and install this custom node:
ComfyUI-Frame-Interpolation
or git clone it
https://github.com/Fannovel16/ComfyUI-Frame-Interpolation
after successful installation, put the file here
custom_nodes\comfyui-frame-interpolation\ckpts\stmfnet\stmfnet.pth
(stmfnet.pth is the file, not a folder obviously)
in comfy load this node: STEMFNet VFI
There isn't a 1.3B model for i2v?