Load Video Path (FFMPEGA)

Most video loaders in ComfyUI decode every frame into GPU memory the moment they run. That's fine for a 30-frame test clip and a quiet way to kill a machine when you point it at a 4K travel video. Load Video Path (FFMPEGA) takes the opposite approach: it validates the file exists, shows you an inline preview and the metadata, and then hands the path downstream as a string. Zero frames loaded, zero VRAM spent. ffmpeg reads the file directly later, when the actual work happens.

That's the entire design philosophy of this pack, and it matters most here. The outputs tell the story: video_path (a validated string you connect to the FFMPEG Agent's video_a/video_b/video_c slots), frame_count, fps, and duration - the metadata you need for decisions, without the payload.

The inputs that matter

• video - a file picker. Select from ComfyUI's input directory or upload a new one. The file is not loaded into memory; only the path is forwarded.
• force_rate - override the source FPS (0 = keep source). Use it to normalize mismatched clips before you concat them - a classic source of "the output is all over the place" failures.
• skip_first_frames / frame_load_cap / select_every_nth - VHS-style trim parameters: skip the intro, cap how many frames count as usable, or sample every Nth frame. These affect the frame_count/duration metadata, so they're useful for making downstream nodes behave on a cropped or subsampled input.

There are pass-throughs if you need them - images, audio, and video_path inputs so you can chain from Save Video or another loader (a connected video_path overrides the picker). And like the rest of the pack, it carries mask_points and crop_data from its point/crop selectors, which feed SAM3-guided masking downstream.

Installing it

It ships in ComfyUI-FFMPEGA, so the whole pack installs at once: ComfyUI Manager → search ComfyUI-FFMPEGA → Install, or:

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

Restart ComfyUI. FFMPEG on PATH is the one hard dependency (ffmpeg -version should print a version, not "command not found").

Where people get burned

The trap is assuming it works like the standard VHS loaders. Because it emits a path, it only connects to nodes that speak path - in this pack, that's the FFMPEG Agent's video_a/video_b/video_c inputs, not an images slot. Wire it into something expecting a tensor and nothing connects. And remember the trim knobs exist: if you're stacking several clips for a grid or concat, set force_rate on all of them to the same value first, or your mixed-frame-rate output will look like a flipbook. For single-clip work with an LLM in the loop, frame_load_cap is also your "don't let the agent analyze the whole 20-minute video" lever.