LOCAL VIDEO UPSCALING TEST
Original Clout synthetic source: 640x360, 24 fps, 3 seconds, 72 frames, no audio
The ZIP contains decoded PNG source frames; the companion MP4 is an unchanged-frame encoding roundtrip. Neither is AI enhancement or a provider benchmark.

In a new test folder, create an empty frames directory and place the source beside it
Extract:
ffmpeg -i video-upscale-test-source.mp4 -fps_mode passthrough frames/frame-%04d.png
Reassemble unchanged frames:
ffmpeg -framerate 24 -i frames/frame-%04d.png -c:v libx264 -crf 18 -pix_fmt yuv420p -movflags +faststart roundtrip.mp4
Inspect:
ffprobe -v error -show_streams -show_format -of json roundtrip.mp4

Confirm all 72 frames, dimensions, frame order and three seconds of continuous playback
Only then add a supported local model stage into a separate output directory
Check matching numbered filenames, consistent dimensions, no missing frames and unchanged timing
Record exact release, model, device, processing time, dimensions and review observations
Reassemble at the original rate and inspect motion, identity, textures and product detail
The supplied rate applies only to this constant-frame-rate test; variable-frame-rate footage needs a separate timing plan
For footage with audio, preserve the original and verify duration and synchronization across the full finished export
Colab is hosted processing, not local inference
Sources checked October 2, 2026:
https://github.com/k4yt3x/video2x
https://github.com/xinntao/Real-ESRGAN-ncnn-vulkan
https://ffmpeg.org/ffmpeg-formats.html#image2