Video Folder Cowboy

If you've ever written a video batch pipeline by hand, you know the pain that this node exists to kill: you point a loader at a folder, and vid10.mp4 loads before vid2.mp4 because that's what plain lexicographic sort gives you. Video Folder Cowboy is a directory iterator that sorts filenames naturally - splitting text and numeric chunks so vid1 < vid2 < vid10 actually holds - then hands you one video's frames as an IMAGE batch, indexed by a simple counter.

This is the load-once-iterate-many node that sits at the front of per-clip render pipelines. You wire the counter into a queue loop, and each run loads the next clip. It's also the reason the pack has a whole cowboy family: wire this node's filename output into TextFileCowboy and RefFolderCowboy and clip 000.mp4 automatically pairs with prompts/000.txt and refs/000/ - a whole keyed batch system with zero naming pain.

How it works

The node globs your folder (default pattern **/*.mp4, **/*.avi, **/*.mov, **/*.mkv, **/*.webm), sorts the matches, then loads frames with OpenCV: BGR→RGB conversion, float normalization to 0–1. The frame-skip and limit controls do the heavy lifting for big files - you can load every 4th frame or cap at 30 frames, which is how you preview a long clip without importing it all.

The inputs that matter

• video_index - which file in the sorted list to load. This is the knob your loop turns.
• sort_by - name (natural sort) is the default and usually what you want; you can also sort by date_modified, size, random, or none.
• index_mode - what happens when the index overruns the file count: wrap (default, back to the start), clamp (stick at the last file), or error (raise). Wrap is perfect for infinite queue loops; error is what you want when a missing clip should stop the run, not silently repeat.
• frame_skip, max_frames, start_frame - skip N frames between loads (0 = every frame), cap total frames (0 = all), and skip to a starting frame index.

Outputs

Six of them, and they're all actually useful: frames (the IMAGE batch), filename (no extension - this is what the cowboy pairings key on), total_videos, file_path (full path), frame_count, and fps. The last one matters more than beginners expect, because downstream VHS and Combine nodes want real FPS rather than a guess.

Troubleshooting

Where people get burned: forgetting that frame_skip counts between loaded frames, so frame_skip=3 loads every 4th frame, not every 3rd. And if you're feeding these frames straight into a Wan 2.1 sampler and hitting errors, the frame count is your problem, not the loader - Wan wants 4x+1 frames (81, 97…), so run the frames through the pack's Wan2.1 Frame Adjuster first. randomize_final is there if you want a shuffled deck; sort_subdirs (on by default) keeps subdirectory processing deterministic.

Installing

TrentNodes installs once and gives you the whole Trent/ menu. Easiest: ComfyUI Manager → search "Trent Nodes" → Install. If Manager complains about the pack being flagged (a day-one repo rename left a duplicate in the registry that still trips the unsafe check), clone it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes
cd TrentNodes
pip install -r requirements.txt

Restart ComfyUI and you're done. This node only really needs OpenCV, which the pack pulls in anyway.