Load Image Path (FFMPEGA)

Six megabytes per image doesn't sound like anything until you're stacking thirty of them into a grid, a slideshow, or a watermark pass. Every standard Load Image node decodes its file into a full IMAGE tensor in GPU memory, and the bloat multiplies fast. Load Image Path (FFMPEGA) sidesteps the whole problem with a dumb-simple trick: it gives you the file path as a string and lets ffmpeg read the file directly when the work happens. Zero decoding, ~0 MB, any number of images.

That makes it a featherweight companion to the pack's heavier nodes. The pack README frames it as the no-brainer for FFMPEG Agent workflows, and it is: pick an image with the image file picker, and the output is image_path, which you wire into the Agent's image_path_a / image_path_b slots for overlay, grid, and slideshow skills. One node per image, or chain them - the Agent auto-expands the slots.

The inputs that matter

Honestly, there's basically one: image, the picker (select from ComfyUI's input directory or upload). The rest are plumbing for chainers:

• image_path - an upstream path string that, when connected, overrides the file picker. Handy when you're passing an image path from another node rather than picking fresh.
• images - an upstream IMAGE tensor, passed straight through to the images output. It's how you keep a tensor-based workflow happy while also handing the Agent a path.
• mask_points - pass-through from the point selector, forwarded for SAM3 guidance.

Outputs mirror that: image_path (the one you actually want), images (the pass-through), and mask_points.

Installing it

It's part of ComfyUI-FFMPEGA, so you get it with the pack: ComfyUI Manager → search ComfyUI-FFMPEGA → Install, or manually:

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

Restart ComfyUI, and make sure ffmpeg is on your PATH - the pack leans on a system ffmpeg for everything that actually touches media.

Where people get burned

Same class of mistake as its video sibling: it outputs a path, so it only feeds nodes that accept paths. Don't try to plug image_path into an image-input slot - that's what the images pass-through output is for. And the zero-memory pitch cuts both ways: because nothing is loaded, there's no validation error if the file goes missing between picking and running - the failure shows up later as a no-op or a failed ffmpeg run. Keep the source file where you picked it from, and this node is the quietly efficient workhorse of the pack.