TS Video Loader

Most video loaders in ComfyUI are a file picker and a prayer: load the whole thing, then trim with numbers. TS Video Loader is the version you'd build if you had time - it reads a video into frames, audio and a compact metadata bundle, and lets you pick the piece visually, on a timeline that lives inside the node.

The node's body holds a player and a timeline with a filmstrip. Drag the handles to set in and out, zoom into a single second of an hour-long take (Ctrl/Cmd + wheel), loop the selection while you judge it, type an exact timecode when the mouse isn't precise enough. If the file has audio, the waveform is drawn under the filmstrip - a beat or a spoken word is far easier to hit by the wave than by the picture. And the player follows whichever handle you're dragging, so the exact frame that becomes the first or the last is on screen while you're still choosing it.

Fast on purpose

This is where the loader separates from the pack. All resizing, rotation and colour conversion happen inside the decoder's own filter graph, and reading stops at the end of your selection. The README's numbers: a two-second cut from an hour-long 4K take costs a seek plus two seconds, not an hour. Measured at 0.43 seconds where the naive path - decode the whole thing at full resolution, then resize - takes 5.9 seconds and 6.4 GB of memory. The frame_rate resampling works by real timestamps, so a variable-frame-rate source comes out evenly spaced instead of drifting.

The inputs that matter

• source_path - the file. Drag and drop from the file manager or another node's preview, paste, or type a path anywhere on the ComfyUI machine.
• start_seconds / end_seconds - the selection. end_seconds of -1 means "until the end of the file."
• frame_rate - output rate; 0 keeps the source rate. Fractional rates like 23.976 work.
• longer_side / shorter_side - sizing. Set one, the other derives from it (0 means "derive me"), so one graph fits landscape and portrait footage alike. divisible_by rounds down to what video models want (8, 16, 32).
• max_frames - hard cap on loaded frames, 0 means no cap; the memory guard.
• resize_filter - area (default) suits strong downscaling, lanczos keeps edges sharp, neighbor is for pixel art.
• when_too_large - what to do when the frames won't fit in RAM (the ceiling is 60% of the machine's memory, at least 8 GB): stop refuses with a message naming the size, or use disk puts frames in a memory-mapped file in the temp folder so the allocation can't fail - at the price of disk traffic (measured: 31.9 GB took 92 s on disk against 51 s in RAM).

Outputs

• images - the decoded frames, float32 in 0..1.
• audio - the same time range's audio, or nothing at all when the file has no track (silence would quietly overwrite a real track downstream).
• video_info - the metadata bundle; feed it to TS Video Info to get the numbers.
• video - the trimmed source file itself, no resizing or rate change, as a cheap VIDEO reference for nodes that take video.

Installing it

Part of comfyui-timesaver: ComfyUI Manager → search Timesaver, or

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

then restart ComfyUI. No system ffmpeg needed - the pack ships imageio-ffmpeg's static binary.

Gotchas

One security note worth reading: on a localhost ComfyUI (127.0.0.1), this node and its preview can read any path you give it - that's the point, nothing gets copied into input. But if you run ComfyUI open to a network (--listen 0.0.0.0), the preview is served over HTTP to whoever can reach that port, so the loader then stays inside your home folder and ComfyUI's own directories. Owners of the machine can widen that with TS_MEDIA_EXTRA_ROOTS or lift it with TS_MEDIA_ALLOW_ANY_PATH=1 - set on the machine, not in a workflow, since workflows can arrive from anyone.