Prep Video For Extend - Klinter

This is the second half of the Klinter video-extension pair, and it does one very specific, very important thing: it takes a stack of video frames and cuts out exactly 5 seconds' worth (120 frames at 24 fps) starting from a point you choose. Video extension models are famously finicky about context length - they want a specific number of frames - and this node exists to give them precisely that, nothing more, nothing less.

The inputs are frames (an IMAGE tensor, typically straight from the pack's Load Video For Extending node), cut_point (a FLOAT, default 0), and cut_point_type (a dropdown: seconds or frames). Pick your cut point in whichever unit you think in, and the node converts: if you say seconds, it multiplies by 24 (hardcoded assumption that frames came in at 24 fps from the loader); if you say frames, it uses the number as-is. Either way it slices out 120 frames starting there and returns them as processed_frames.

Two constraints come straight from the source and will frustrate you if you don't plan around them. First, your source must be at least 5 seconds (120 frames) long - shorter clips error out immediately. Second, the cut point must leave at least 5 seconds of video after it, because the node always takes a full 120-frame slice. So on a 6-second clip you can only cut at 0 or 1 second. That's not a bug, it's the "extension needs a full context window" contract, but it means you can't use this node to trim the tail of a long video - you'd cut the start of the window you want to extend from.

It's also not a general-purpose trimmer: no resize, no fps change, no looping or interpolation. Feed it frames in 24 fps and it assumes as much. If your frames come from a different source at a different rate, the math silently misaligns - keep the pack's own loader upstream and you're fine.

Install is the pack standard: search "Klinter_nodes" in ComfyUI Manager, or cd ComfyUI/custom_nodes && git clone https://github.com/klinter007/klinter_nodes then restart. Pure torch, no extra deps, no ffmpeg needed here (that's the loader's job).

Use it as the bridge between "video I loaded" and "video the model can extend" - it makes the frame-count contract a solved problem.