RIFE 4.7 / 4.9 (Practical-RIFE IFNet) for gufo
What this is
The IFNet checkpoints rife47 and rife49 (architecture "4.7") as FP32 safetensors for
gufo-studio-windows (gufo serve video
with rife-model, gufo video-interpolate). gufo runs the network natively on HIP
(AMD Strix Halo, Windows) to raise the frame rate of finished videos, following the defaults
of the ComfyUI "RIFE VFI" node.
Source and licence
- Model: Practical-RIFE by hzwer (github.com/hzwer/Practical-RIFE),
MIT License, "Copyright (c) 2021 hzwer" (included as
LICENSE). Paper: RIFE, Huang et al., ECCV 2022. - The
.pthfiles were obtained from the download mirrors that ComfyUI-Frame-Interpolation (MIT) uses, e.g.https://huggingface.co/marduk191/rife/resolve/main/rife49.pth. Those mirror repositories carry no licence metadata of their own; the weights are the Practical-RIFE releases. SHA-256 of the source files:rife47.pth6a8a825a...,rife49.pthe55fd00f....
How it was converted
tools/rife/convert_rife_pth.py from gufo-studio-windows (standard library only; a restricted
unpickler that knows only the state_dict globals, so no checkpoint code runs). A module.
prefix is stripped; 124 tensors each, FP32, metadata format=rife-ifnet, source=<pth>.
rife47.convert.txt / rife49.convert.txt list the tensor names and shapes.
How to use with gufo
[external/gufo-rife]
command = <gufo.exe>
args = serve video
model = <video model folder>
root = <job folder>
rife-model = <this folder>\rife49.safetensors
ttl = 86400
CLI: gufo video-interpolate --model rife49.safetensors --input in.mp4 --output out.mp4 --multiplier 4
or --target-fps 60.
Files
| File | Size |
|---|---|
rife49.safetensors |
21.3 MB |
rife47.safetensors |
21.3 MB |
rife49.convert.txt, rife47.convert.txt |
5 KB each |
AI-assistance note
Converted and documented with AI assistance (Claude Code).