SAM Multi-Mask Picker — 3 candidates + scores

Here's the SAM experience every newcomer has: you click a point, and the model returns three masks - one tight, one loose, one that grabbed half the background - ranked by confidence, and you're supposed to know which mask_index to take. The community's answer is exactly what this node is: stop guessing, show all three, let the human pick. SAM Multi-Mask Picker (MEC) runs the inference and then puts the three candidates in front of you as clickable thumbnails with their confidence scores, so the choice is visual instead of blind.

How the picking works

How it works is two parts. Underneath, it's a full SAM inference - image in, a point prompt in points_json (default [{"x":256,"y":256,"label":1}]) and/or a box in bbox_json, and SAM's mask decoder returns its standard three candidates. On top, a JS widget renders a thumbnail of each mask overlaid on your image; you click one or press 1/2/3 to select it, and selected_index updates to match. Press R and it re-runs with fresh prompts. The pick lands on selected_mask, and all_masks carries all three as a stack so you can inspect them downstream. scores is a comma-separated confidence list - visible proof of why you shouldn't have trusted mask 0 by default.

Where the model comes from

The model comes from a model_name dropdown (it auto-lists whatever SAM weights are in your models folders, with sam2.1_hiera_base_plus as the default - bigger models = better quality, more VRAM), or you can plug in a pre-loaded model from SAMModelLoaderMEC via the sam_model input, which overrides model_name. Same for prompts: bbox from an upstream BBox node overrides bbox_json. precision (default fp32, with fp16/bf16 for VRAM) is worth dropping on a tight card.

Which SAM generation you're on

One nuance: the brief and the source agree this node's lineage is the older SAM1/SAM2/HQ-SAM generation - it says it works with SAM1, SAM2, and HQ-SAM, which means it's the "pick visually" tool for the classic promptable SAM workflow, while the newer pack nodes (Mask Generator, the ViTMatte pipeline) are the SAM 2.1/3 line. Don't let that scare you: the candidate-picking problem is identical across generations, and if you're on the classic SAM weights this is the right UI. If you're on 2.1/3, the SamMultiMaskPickerMEC still accepts a pre-loaded sam_model from the loader, so you get the picker UI with modern weights.

Outputs and install

Outputs: selected_mask, all_masks, selected_index, scores, info (JSON of prompts, scores, selection). Feed selected_mask into an inpaint mask, a matting node, or a detail pass - the whole point is that downstream work starts from the candidate you actually meant.

Install is the pack: clone Code2Collapse/ComfyUI-CustomNodePacks into custom_nodes (or Manager → "CustomNodePacks") plus the SAM runtime and weights (models/sam2/, models/sams/) and pip install git+https://github.com/facebookresearch/sam2.git for the 2.1 line. It's the node the README frames as "tired of SAM giving 3 masks and not knowing which one to pick" - and honestly, that's most of us.