Load Images/Videos From Directory

Single-file loaders are easy. The moment you have a folder of 200 renders, 50 shot files, or a frame sequence scattered across subdirectories, you find out which loader nodes actually pull their weight. Load Images/Videos From Directory is the heavyweight of SineSwiper's LoadAnim-Adv pack: point it at a folder and it loads every image or video in it, lets you pick which frames to keep, and hands you the file paths so you know what you got.

What it's for

This is the batch engine of the pack. It runs the same loader that backs Load Image/Video, but in a loop over a directory scan. Typical jobs: feeding a whole shot's worth of frames into a vid2vid pipeline, ingesting a folder of GIFs for frame extraction, or preprocessing an image sequence before a batch loop. The README even calls out the killer trick - image_file_start_index exists specifically so you can increment it in a batch loop and chew through files sequentially without reloading everything.

The inputs that matter

Most are shared with the single-file loaders: directory, RGBA, width/height (0 = original), keep_aspect_ratio. The interesting ones are the batch controls:

• valid_extensions - comma-delimited list, default jpg,jpeg,png,webp,gif,tga. Add mp4,mov,avi,mkv if you want videos in the scan.
• sort_method - None, Alphabetical (ASC/DESC), Numerical (ASC/DESC), or Datetime (ASC/DESC). Numerical is the one you want for frame sequences - it sorts file2 before file10, which plain alphabetical sorting won't do.
• image_file_load_cap - max number of files to load. 0 = no limit.
• image_file_start_index - skip this many files from the front. This is your batch-loop hook.
• frame_indexes_to_select - the frame-selection syntax, applied to each file. Default 0: means every frame.
• total_frame_load_cap - a global cap on frames across all files, for when the folder is bigger than your VRAM patience.
• flatten_frames - when on, all files' frames get concatenated into one IMAGE tensor instead of staying as per-file sets.
• include_subfolders - recurse. Off by default.

Outputs

Everything comes out as a list: image, mask, audio, frame_counts, fps, and image_path (the filenames, in the same order - invaluable for debugging what actually loaded). If flatten_frames is on, image collapses to a single combined set.

The frame selection syntax

It's Python's range() with the parts separated by colons, first frame = index 0:

• 0:5 - frames 0 through 4
• :-5 - all but the last 5
• ::2 - every second frame
• 0,5:10,-1 - a mix of singles, ranges, and the last frame

Gotchas - read these before you run it

Two real traps, both in the author's own tooltips:

1. Audio is not sliced by frame selection. If you select a subset of frames from a video, the audio output still carries the full track - so it'll drift out of sync with your shortened clip. The README flags it as a known issue.
2. Flattened FPS is a lie. If you flatten files with different frame rates, the fps output is just the average, which the tooltip warns "may have unpredictable results." Keep your sources at matching FPS if you need the number to mean anything.

Also worth knowing: the node raises an error if the directory doesn't exist or has no matching files - it fails loudly rather than silently feeding your pipeline nothing.

Installing

ComfyUI Manager → search "LoadAnim", or:

cd ComfyUI/custom_nodes
git clone https://github.com/SineSwiper/ComfyUI-LoadAnim-Adv.git

Restart, done. Dependencies are just torch, numpy, Pillow - no models to download.

For a one-shot single file this node is overkill. For a folder, it's the difference between writing 200 Load Image nodes and writing one.