🎬AGSoft LastFrameExtractor

Every video pipeline needs "the last frame" at some point - as the starting frame for a continuation, as a thumbnail, or as the img2img base for what comes next. Stock ComfyUI doesn't hand you that on a plate; you'd load all the frames and index into the batch. AGSoft LastFrameExtractor pulls a single frame from the end of a video file, at whatever offset you want, and returns it with the video's metadata attached.

The offset is the feature. frame_offset_from_end = 1 gives you the very last frame, 2 the second-to-last, and so on. That matters more than it sounds: video files often end on a fade, a black frame, or a credit card, and being able to reach one or two frames back means you don't have to eyeball a frame count or trim the file first.

How it works

The node takes three possible sources, in priority order:

1. video - a VIDEO object from Video Helper Suite's Load Video node or standard ComfyUI's loader. Highest priority; if it's connected, the other sources are ignored.
2. video_frames - pre-loaded frames as an IMAGE tensor. Medium priority; used when no VIDEO object is attached.
3. custom_path - a path to the video file. Lowest priority, so it's the fallback when nothing's connected.

It reads the source, counts the frames, and extracts the one frame_offset_from_end frames from the end. If you've already got a VIDEO object in your graph from a VHS loader, this node is basically a zero-setup drop-in.

Outputs

• frame (IMAGE) - the extracted frame, ready for a VAE encode or as an image.
• width / height (INT) - the frame's dimensions.
• total_frames (INT) - the video's frame count. Often the more useful number for scheduling.
• video_metadata_json (STRING) - metadata as a JSON string, which you can pipe into a text display or parse if you're automating.

Installation

It's part of comfyui-AGSoft:

cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft.git
# restart ComfyUI

Or ComfyUI Manager → search "comfyui-AGSoft". The pack pulls in numpy and opencv-python; video reading leans on whatever video backend ComfyUI already has (VHS installs its own), so no extra model files.

Where people get burned

The number one confusion is the priority order - people wire up a custom_path and a video at the same time and wonder why the path is being ignored. The tooltip is explicit: the VIDEO object wins. If you're testing with a path, leave the video inputs unplugged. Second, remember this extracts one frame. If you need a spread of frames or a true thumbnail strip, the pack's VideoFrameExtractor is the sibling that does ranges. And if the last frame comes back black, that's usually the video's own last frame (fades) - bump frame_offset_from_end to 2 or 3 and move on.