TS SAM Media Loader

SAM3 (Meta's promptable segmentation line) is one of the best things to happen to masking in ComfyUI - you give it a point or two and it returns a clean mask for that object, zero-shot, on anything. The catch is that the native SAM3 Detect node wants its prompts as JSON strings, and typing coordinates by hand for a video frame is miserable. This node is the clicky front end: load an image or video, click green (positive) and red (negative) points right on a first-frame preview, and get back the exact prompt JSON the native SAM3 nodes expect.

How it works

It's a loader with a point-picker bolted on. Drop an image or a video file in, and the node's body shows the first frame. Every click records a point in pixel space; positive (green) points say "this is the thing," negative (red) points say "this is NOT the thing." That's the same interaction you'd have in the original SAM demo - a few clicks to steer the mask, and SAM3 does the rest.

The outputs are what make it play nicely with the native stack:

• images - the loaded frames as an IMAGE batch (frame_stride decodes every N-th frame for video; max_frames caps the count).
• positive_coords and negative_coords - STRING JSON in the SAM3 point-prompt format. Wire those straight into SAM3 Detect's point-prompt inputs.
• fps and audio - the video's frame rate and audio track, pulled out for free.
• initial_mask - if you also connect a SAM3 model, the node renders the first-frame mask itself, ready to feed SAM3 Video Track for tracking.

One honest caveat: the optional model input is preview-only. It lets the node draw a blue mask overlay so you can see what your points did, but the actual SAM3 Detect in your workflow should be wired to the model loader directly. Don't connect this node's model and expect it to do the segmentation for you.

The settings that matter

The top four inputs (source_path, media_type, coordinates, neg_coordinates, sam3_checkpoint) are annotated as internal - the node writes them as you click; you don't. The two you might touch are frame_stride (1 = every frame, which is what you want for tracking; higher for a quick look) and max_frames (a memory guard for long clips).

Install

No extra dependencies. It's part of comfyui-timesaver, and SAM3 itself ships with native ComfyUI - so this is the rare node that installs and just works. Video frames come through the pack's bundled imageio-ffmpeg, so you don't even need a system ffmpeg.

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

Then restart ComfyUI, add the node, and drop a file in.

Common issues

• "My points do nothing." Check that you're feeding positive_coords/negative_coords into the point-prompt sockets of a real SAM3 Detect node - and that Detect is wired to its own model loader. This loader only formats the prompts.
• Strange frame count for video. frame_stride > 1 means you're sampling frames, not skipping them - a 60 fps clip at stride 2 gives you 30 frames at real-time speed. That's usually what you want, just don't be surprised.
• No audio output for a clip. The audio output is empty for images by design; for video, check max_frames and frame_stride aren't set so aggressively that the decoded selection has nothing to carry the audio. Frames and audio come from the same decode.

Where this shines is the SAM3 Video Track workflow: points → initial_mask on frame one → tracking across the clip. Without this node, that's a wall of JSON you'd be typing by hand.