(Down)Load SAM2Model

This is the "load the thing" node for kijai's SAM2 pack, and every other node in the pack starts here. Segment Anything 2 is Meta's universal segmentation model - point at an object, or hand it a box, and it hands back a precise mask. SAM2 is the successor to the original SAM: more accurate, and, the big one, it does video - it can track an object across frames instead of re-guessing every frame. This node grabs the weights and gives you a sam2_model to plug into whichever segmentor you're using.

The (Down)Load in the name is literal: on first run it downloads the checkpoint you picked, so the model dropdown is populated before you've manually fetched anything. Pick a model, pick a mode, wire it forward. That's the whole job.

The mechanism, and the one setting people get wrong

SAM2 comes in a family of sizes built on Meta's Hiera backbone, and the loader exposes all of them: sam2_hiera_tiny, small, base_plus, and large, plus the sam2.1 refresh of each. Tiny is fast and light and honestly fine for clean, obvious subjects; large is slower and sharper on fiddly edges. The 2.1 weights are the newer revision - reach for those unless you've got a reason not to.

The setting that actually trips people is segmentor. It has three choices - single_image, video, automaskgenerator - and it has to match the node you're feeding downstream. This isn't cosmetic: the loader builds a different predictor object for each mode.

• single_image → feeds Sam2Segmentation (mask one image from points or boxes)
• video → feeds Sam2VideoSegmentationAddPoints + Sam2VideoSegmentation (track across frames)
• automaskgenerator → feeds Sam2AutoSegmentation (segment everything, no prompts)

Wire a video model into the single-image node and it'll error or behave weird. If something in the pack is throwing a type or attribute error right out of the gate, this is the first thing to check.

Inputs and outputs that matter

Four required inputs, and you'll only think about two:

• model - which SAM2 checkpoint. Start with sam2.1_hiera_base_plus; it's the sensible middle. Downloads on first use.
• segmentor - the mode, per above. Match it to your downstream node.
• device (cuda / cpu / mps) and precision (fp16 default, bf16, fp32) - leave these. cuda + fp16 is what you want on any NVIDIA card. fp16 is plenty accurate here; there's no reason to burn VRAM on fp32 for a segmentation mask.

Output is a single sam2_model (type SAM2MODEL) - the handle every SAM2 node in this pack expects as its first input.

How to install it

Easiest path is ComfyUI Manager: open the Manager, search ComfyUI-segment-anything-2, install, restart. Or clone it yourself:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-segment-anything-2

Then restart ComfyUI. The model weights land automatically the first time you run the node (into ComfyUI/models/sam2), so budget a minute of download on that first execution - large is a couple gigs, tiny much less. If you're on ComfyICU it's already in the shared image, so you skip both the install and the wait.

Common issues

• Wrong segmentor for the node you're feeding. The number-one gotcha. video mode won't drive the single-image node and vice versa. Set the mode to match your workflow before you debug anything else.
• First run hangs "downloading." It's not frozen, it's pulling weights. Watch the console. If it fails, it's usually a network hiccup or a HuggingFace reach problem - re-run.
• Slow, or OOM on a big video. SAM2's video mode holds state across frames, so VRAM climbs with clip length. Drop to a smaller Hiera variant (small or tiny) or shorten the clip before you blame the loader.

Once you've got a sam2_model out of here, the interesting work happens downstream - masking for inpainting, regional prompting, compositing, or object tracking in video. This node just makes sure the model's loaded in the right mode for whatever that is.