---

name: interpolation description: "Use ECCV2022-RIFE image-pair, video, and PNG-sequence interpolation CLIs safely, including checkpoints, scales, ratios, and output handling." disable-model-invocation: true metadata: disco-role: operating license: MIT

ECCV2022-RIFE interpolation

Use this sub-skill when the user asks to interpolate two frames, run RIFE on a video, make 2X/4X/16X slow motion, use an arbitrary --ratio, process a numbered PNG frame directory, tune UHD/4K --scale, create montage output, preserve or explain audio transfer, or diagnose RIFE checkpoint layout for inference.

Do not use this sub-skill for official benchmark metrics, dataset evaluation, training, Vimeo checkpoint production, or architecture changes. Route benchmarks/metrics to the repository evaluation guidance and route training/checkpoint production to the repository training guidance.

Operating procedure

1. Confirm the user has a source checkout with dependencies available. The inference scripts are source scripts, not installed console entry points.

2. Confirm a checkpoint directory. The default is train_log; a non-default directory must be passed with --model. For the current source fallback, expect flownet.pkl inside the checkpoint directory.

3. Pick the workflow:

   • image pair: inference_img.py with --img left right, --exp or --ratio;
   • video file: inference_video.py with --video input.mp4;
   • numbered PNG frames: inference_video.py with --img frames_dir and numeric filenames such as 0.png, 1.png, ...;
   • UHD/4K: add --UHD or an explicit --scale such as 0.5.

4. Use the bundled command builder for safe command construction and optional input/checkpoint validation before proposing a long inference run:

   python sub-skills/interpolation/scripts/interpolation_command_builder.py --help
   

5. Before running inference, tell the user where outputs will be written. Image interpolation writes output/img*.png or output/img*.exr relative to the current working directory. Video PNG output writes vid_out/*.png; video-file output writes the requested --output or a derived *_2X_*fps.<ext> / *_4X_*fps.<ext> style path.

6. If the user requests CUDA, FP16, or speed claims, state that functional inference can fall back to CPU but is much slower. --fp16 is only appropriate on CUDA devices with suitable Tensor Core support.

Required references

• CLI details, option meanings, checkpoint behavior, and output naming: references/cli-reference.md
• End-to-end recipes for image pairs, ratios, video files, PNG sequences, UHD, montage, and audio handling: references/workflows.md
• Troubleshooting for checkpoints, missing dependencies, frame ordering, scale/shape, EXR/PNG, CPU/CUDA/FP16, and audio/ffmpeg: references/troubleshooting.md
• Safe command-building helper: scripts/interpolation_command_builder.py

Key boundaries and caveats

• Do not promise verified HD-model variants beyond the current source evidence. The current inference scripts attempt HD imports first, but this checkout does not provide the active model.RIFE_HD* import paths; ordinary use falls back to model.RIFE.
• Do not claim pretrained weights are bundled. Checkpoints are external assets and must be provided by the user.
• Do not run long video interpolation automatically. Build and validate the command first, then ask for approval if execution is potentially expensive or mutating.
• Do not use --skip as an active static-frame optimization; the current script prints that this flag is abandoned.