Scene Cuts JK🐉

This is the node behind the pack's "Cuts" series of video workflows - the ones that generate multiple videos, one per scene in your edit, instead of one long generation that drifts off-model halfway through. Scene Cuts JK takes a list of cut times or per-scene durations and turns it into frame-accurate scene segments that downstream samplers can loop over. You give it a timeline; it gives you the per-scene frame math.

The two modes (the mode toggle)

• mode = True (duration mode, default) - your segments values are durations: "3.0, 6.2, 9.875" means scene one is 3 seconds, scene two is 6.2, scene three is 9.875. It chains them end to end.
• mode = False (cut-point mode) - your segments values are absolute cut points on the timeline: where each scene ends.

Either way, segments accepts comma- or newline-separated values in second.millisecond format. And there's an optional audio input: feed it an AUDIO tensor and the node reads its actual duration, clamping the total to the audio length - which is how the "video based on scene & audio cuts" workflows stay in sync with the soundtrack.

The frame-count knobs

• fps (default 16) - frames per second for the conversion.
• loop_frame_count (default 81) - the per-scene generation length the sampler will use.
• overlap_frame_count (default 10) - the overlap between consecutive long-video segments, so cuts blend instead of jumping.
• warmup_frame_count (default 0) - Wan-only. Some Wan Wrapper models need a few warmup frames before output is usable, especially ref2v; this subtracts them.
• compatibility - toggles frame-count rounding between Wan-style (multiple of 4 + 1) and LTXV2-style (multiple of 8 + 1). This is the knob that makes the same node drive both model families.
• long_vid_method - Wan-only; switches to the for-loop long-video stitching strategy.
• min_loop_frame_count - deactivated in v2.1.8; leave it alone.
• select_cuts - filter which cuts come out (e.g. 0, 0-2, 0,2,4). Handy when you only want to regenerate scene three.

The outputs

The one you actually feed downstream is scene_cuts - a structured dict {cuts: [[start, end], ...], count, warmup} that the video module's cut-driven nodes consume. Alongside it you get a bunch of readable summaries: scene_count, cut_frame_counts, loop_frame_counts, total_duration and total_frame_count (each showing both the scene timeline and the actual generated frames), select_cuts, fps, and echoes of seg_frame_count, overlap_frame_count, and long_vid_method. Preview those string outputs while tuning and they tell you exactly what the graph will generate.

Why the "Cuts" approach wins

Long video generation fails because 500+ frames of a single diffusion run drifts - color shifts, identity changes, coherence collapse. Cutting the video into scenes and generating each one (with overlap and consistent stitching) keeps every segment on-model. It's the same philosophy as the pack's long-video workflows: chunk it, then sew it. This node is the seamstress's ruler.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/jakechai/ComfyUI-JakeUpgrade
pip install -r requirements.txt

It reads audio duration straight from the AUDIO tensor, so no extra audio deps. The most common setup error is feeding duration-mode values into cut-point mode (or vice versa) and getting weird scene counts - if scene_count looks wrong, flip the mode toggle before touching your timings.