• Skip to main content
  • Skip to secondary menu
  • Skip to primary sidebar
  • Skip to footer
Best Free Web Resources — local AI tools and hardware guides

Best Free Web Resources

Local AI Tools & Digital Engineering Resources

What can I run?
  • AI Tools
    • Start: Local LLMs
    • Ollama Models & VRAM
    • LM Studio Models
    • GPU Comparison
    • AI Hardware Matrix
    • Local LLM Directory
    • VRAM Calculator
    • Hardware Compare
    • Local Model Recommender
    • API Pricing & ROI
  • Local AI Guides
  • Web Resources
    • Inspiration
    • Photography
    • Games
    • SEO
    • Tutorial
      • Coding
      • WordPress
    • Resume/CV
    • Graphics
      • Logo
    • Freebies
Cover illustration: SAM 3 in ComfyUI

SAM 3 Locally: Files, ComfyUI Nodes, Prompts and the Licence

November 22, 2026 by Rupesh Kumar

SAM 3 runs natively in ComfyUI from v0.20.0, and it needs one file: sam3.1_multiplex_fp16.safetensors, 1.75 GB, placed in ComfyUI/models/checkpoints/. That file is Comfy-Org’s repackage of Meta’s SAM 3.1 checkpoint, it carries the text encoder inside it, and unlike Meta’s own Hugging Face repos it is not gated. You load it with an ordinary Load Checkpoint node, type what you want masked into a CLIP Text Encode node, and the SAM3 Detect node returns masks and boxes.

This guide covers the files, the nodes and settings, the prompt syntax that sets how many objects you get, two open bugs, Meta’s own Python route, and the SAM License. BFWR has not run SAM 3. The steps and figures come from Meta’s repository and licence, ComfyUI’s source code and pull requests, the Hugging Face API, and the projects’ GitHub issues, all read on 15 November 2026.

Quick answer

  • Native support: ComfyUI has built-in SAM 3 nodes since v0.20.0 (April 2026). No custom node is needed. The current release is v0.45.0 (10 November 2026).
  • Files: one file, sam3.1_multiplex_fp16.safetensors (1,745,546,848 bytes) from Comfy-Org/sam3.1, in models/checkpoints/. Meta’s originals are 3.44–3.50 GB in FP32 and gated.
  • Nodes: Load Checkpoint, CLIP Text Encode, SAM3 Detect. For video tracking: Run SAM3 Video Track, SAM3 Track Preview and SAM3 Track to Mask.
  • The common surprise: a text prompt returns one object by default. Write cat:3 to keep up to three, and separate several prompts with commas, such as cake:2, apron.
  • Memory: the weights are 1.75 GB at FP16. Neither Meta nor ComfyUI publishes a VRAM minimum.
  • Licence: the SAM License allows commercial use, with no revenue or user cap. It bans military, weapons, nuclear and espionage uses and reverse engineering, and Meta can change the terms with immediate effect.

What SAM 3 is, and what changed from SAM 2

Meta’s README describes it like this: “SAM 3 is a unified foundation model for promptable segmentation in images and videos.” SAM 2 needed you to click a point or draw a box for each object. SAM 3 adds “the ability to exhaustively segment all instances of an open-vocabulary concept specified by a short text phrase or exemplars”. You type red car and it masks every red car, in a still image or across the frames of a video.

The model is a detector and a tracker sharing one vision encoder. The README says “It has 848M parameters”, and the Hugging Face API counts 859,922,360 F32 parameters in facebook/sam3. The detector handles text, box and image-exemplar prompts. The tracker carries over the SAM 2 design for video and for refining a mask with points.

Two releases exist:

  • SAM 3 (facebook/sam3, November 2025): sam3.pt at 3,450,062,241 bytes, plus a model.safetensors of 3,439,938,512 bytes for the Transformers library.
  • SAM 3.1 (facebook/sam3.1, 27 March 2026): sam3.1_multiplex.pt at 3,502,755,717 bytes. Meta’s release notes call its Object Multiplex feature “a shared-memory approach for joint multi-object tracking that is significantly faster without sacrificing accuracy”, and claim a “~7x speedup at 128 objects on a single H100 GPU” over the November release. That is Meta’s own figure; its benchmark tables show small gains and losses, not a clean win.

Both Meta repos are gated with manual approval: a form asking for your name, date of birth, country and affiliation, then a wait.

Native ComfyUI support: files and nodes

Support arrived in ComfyUI pull request #13408, “feat: SAM (segment anything) 3.1 support”, by kijai, merged on 23 April 2026 and first included in the v0.20.0 tag of 27 April 2026. The PR describes it as follows: “Model code is a re-implementation for ComfyUI, extra dependency free and optimized for single GPU.” It loads both SAM 3 and SAM 3.1 weights, and kijai wrote that “in my testing I did not really see any reason to use the old SAM3 model weights”. That is his observation, not ours.

The file

Download checkpoints/sam3.1_multiplex_fp16.safetensors from the Comfy-Org/sam3.1 repo and put it here:

ComfyUI/models/checkpoints/sam3.1_multiplex_fp16.safetensors

It is 1,745,546,848 bytes (1.75 GB, where GB means 10^9 bytes throughout this guide), half the size of Meta’s FP32 originals because it is stored at FP16. ComfyUI’s model definition accepts FP16, BF16 and FP32. The text encoder lives inside the same file, and ComfyUI splits it off when it loads the checkpoint, so there is no separate CLIP or VAE file to find. That makes this the smallest working setup: one 1.75 GB download.

The image workflow

ComfyUI ships two templates, Image Segmentation: SAM3 (utility_image_segment_sam3) and Video Segmentation: SAM3 (utility_video_segment_sam3), both in the Image Tools and Video Tools groups and both marked as needing ComfyUI 0.20.0. Both are built on the same three nodes:

  1. Load Checkpoint (CheckpointLoaderSimple) with sam3.1_multiplex_fp16.safetensors. Its CLIP output feeds the text encoder and its MODEL output feeds the detector.
  2. CLIP Text Encode with your prompt, for example person:2, dog.
  3. SAM3 Detect (SAM3_Detect, under image/detection), which outputs masks and bboxes.

SAM3 Detect’s settings, from the node’s source:

  • threshold: 0.5 by default. Lower it if the object you want is missed.
  • refine_iterations: 2 by default, 0 to 5. Its tooltip reads “SAM decoder refinement passes (0=use raw detector masks)”.
  • individual_masks: off by default, which merges every detection into one mask. Turn it on to get one mask per object.
  • Optional bboxes, positive_coords and negative_coords inputs for box and point prompts. The points are a JSON list of x and y pixel coordinates in the KJNodes PointsEditor format.

Every input is resized to 1008 × 1008 pixels before the model sees it, and the masks are scaled back to your image size afterwards. A 4K frame costs the model no more than a small one.

Prompt syntax: why you only get one object

This is the question in ComfyUI issue #14087, “Sam3 Detect node is missing max_detections input”. There is no such input. Each text prompt keeps only its single highest-scoring detection unless you say otherwise in the prompt. kijai’s answer in that issue is to add a count after a colon, so cat:2 keeps up to two cats, and he added that “you can do multiple detection prompts by separating them with comma”. The parser in comfy/text_encoders/sam3_clip.py confirms the rules: split on commas, read an optional :N after each part, default to 1, and drop any brackets, so (cat:2) works too.

The note inside ComfyUI’s own SAM3 workflow adds a limit: “Max tokens for this model is only 32”. Keep each prompt to a short noun phrase, which is what SAM 3 was trained on anyway.

Two open bugs to work around

  • Do not write :1. Issue #15811 reports that person:1 is sent to the text encoder as the literal text person:1, suffix included. In the code on the master branch today, the single-prompt shortcut still passes the raw text through. Write person for one object; the count only needs writing when it is 2 or more.
  • A missing object can still get a mask. Issue #17105 is titled “SAM3_Detect ignores the presence score, so prompts for absent objects mask whatever is in the image”. Meta’s reference code multiplies each detection score by an image-level presence score before thresholding; ComfyUI’s node does not. A fix (PR #17106) was still open on 15 November. Until it lands, check the output when the object might not be there, or raise the threshold.

Video tracking

The Video Segmentation template runs SAM3 Detect on every frame separately. For tracking that keeps object identities across frames, use Run SAM3 Video Track (SAM3_VideoTrack). It takes the frames, the model, an optional initial_mask and optional text conditioning. Its defaults are detection_threshold 0.5, max_objects 4 (0 means the internal cap of 64), and detect_interval 1, which runs detection on every frame; its tooltip says “Higher values save compute”. SAM3 Track Preview writes an overlay video and SAM3 Track to Mask turns chosen object indices into ordinary masks.

Do you still need a custom node?

Mostly not. The best-known one, PozzettiAndrea/ComfyUI-SAM3 (GPL-3.0 wrapper, Meta’s licence for the model), predates native support, had its last commit on main on 19 May 2026, and its README warns that it “uses experimental package comfy-env to attempt a one click isolated install”. Its 24 August pushed_at date reflects a push elsewhere, not new work on main.

Running Meta’s own code

Meta’s repository (facebookresearch/sam3) is the reference implementation and the only route to fine-tuning. Its README lists Python 3.12 or higher, PyTorch 2.7 or higher and a CUDA-compatible GPU with CUDA 12.6 or higher. The install, from the README (it needs an NVIDIA GPU with CUDA):

conda create -n sam3 python=3.12
conda activate sam3
pip install torch==2.10.0 torchvision --index-url https://download.pytorch.org/whl/cu128
git clone https://github.com/facebookresearch/sam3.git
cd sam3
pip install -e .

The README says to “please request access to the checkpoints on the SAM 3 Hugging Face repo” and log in with hf auth login, because the image builder downloads sam3.pt from that gated repo by default. The repository is active: 12 commits on main in the 90 days to 15 November, the latest on 13 November. It has no tagged releases at all, so you install from the main branch.

Its most-discussed memory problem is issue #169: out-of-memory on long videos, which users in the thread traced to per-frame state kept on the GPU and eased by offloading frames and state to system RAM. Long video, not the weights, is what needs memory.

The SAM License, read closely

SAM 3 is not under an open-source licence. It uses the SAM License, last updated 19 November 2025, which is the same text in Meta’s GitHub repo, in both gated Hugging Face repos and in Comfy-Org’s repackage. Note one inconsistency: the repo’s pyproject.toml still carries an “OSI Approved :: MIT License” classifier. The LICENSE file is what governs.

What it grants:

  • A “non-exclusive, worldwide, non-transferable and royalty-free limited license” to “use, reproduce, distribute, copy, create derivative works of, and make modifications to the SAM Materials”.
  • No commercial restriction, no revenue threshold, no monthly-user cap and no excluded countries beyond sanctions.
  • You own the derivative works and modifications you make, subject to Meta’s ownership of the originals.

What it requires and forbids:

  • Redistribution under the same terms. If you share the weights or anything derived from them, “you may only do so under the terms of this Agreement and you shall provide a copy of this Agreement”. A fine-tuned SAM 3 stays under the SAM License. Comfy-Org’s repackage follows this: it ships the same LICENSE, and the PR that added support says “For the model itself, Meta’s original license applies.” Downloading the ungated copy does not get you out of the terms, because the licence binds anyone using any part of the materials.
  • Banned uses. Nothing subject to ITAR, and no end uses “related to military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapons”. You must not be a target of trade sanctions.
  • No reverse engineering. Your use “will not involve or encourage others to reverse engineer, decompile or discover the underlying components of the SAM Materials”.
  • Privacy law. Use must comply with “applicable privacy and data protection laws”, which matters for a model whose obvious job is finding people in photos and video.
  • Research credit. If you publish research results, “you must acknowledge the use of SAM Materials in your publication”.
  • Terms can change. Meta may modify the agreement, and “All such changes will be effective immediately.” Continued use counts as acceptance.
  • Litigation trigger. If you sue Meta claiming SAM infringes your IP, your licence ends on the day you file.

This is a reading of the published licence text, not legal advice. For a product, read the full LICENSE file yourself.

FAQ

Is SAM 3 free for commercial use?

Under the SAM License, yes. It is royalty-free and has no revenue or user limit. The conditions are the banned uses (military, weapons, nuclear, espionage), no reverse engineering, compliance with privacy law, and passing the same licence on if you redistribute the model.

Why does SAM3 Detect only find one object?

Each text prompt keeps one detection by default. Add a count: person:4. Separate different things with commas: person:4, bicycle:2. Turn on individual_masks if you want them as separate masks rather than one merged mask.

Does SAM 3 need a lot of VRAM?

The FP16 weights are 1.75 GB, and every image is processed at 1008 × 1008. Neither Meta nor ComfyUI publishes a minimum. Long videos are what push memory up in Meta’s own code (issue #169). In ComfyUI’s tracker, max_objects caps how many objects are carried, and a higher detect_interval saves compute.

Can I use SAM 3 masks for inpainting?

Yes. SAM3 Detect outputs a standard ComfyUI mask, which any inpainting or video-editing workflow can take. ComfyUI’s own “Change Any Objects” template pairs SAM3 with Wan2.1 VACE for exactly this.

Related guides and tools

  • Stable Diffusion inpainting in 2026: models, VRAM and a ComfyUI workflow
  • ControlNet models in 2026: which ones still ship, and what they weigh
  • ComfyUI install in 2026: Desktop vs Portable vs Manual
  • ComfyUI templates in 2026: 660 built in, and half need the cloud
  • Local AI video generation in ComfyUI
  • ComfyUI hub: models, VRAM and guides
  • Model licence checker

Sources

  • facebookresearch/sam3 on GitHub: README, installation, model description
  • SAM License (last updated 19 November 2025)
  • SAM 3.1 release notes
  • facebook/sam3 and facebook/sam3.1 on Hugging Face (gated; file sizes and parameter count from the Hugging Face API)
  • Comfy-Org/sam3.1: repackaged ComfyUI file
  • ComfyUI PR #13408: native SAM 3 and 3.1 support
  • ComfyUI nodes_sam3.py: node names, inputs and defaults
  • ComfyUI issues #14087, #15811 and #17105
  • SAM 3 issue #169: video out-of-memory
  • PozzettiAndrea/ComfyUI-SAM3
  • SAM 3: Segment Anything with Concepts (paper)

Updated 15 November 2026. Meta’s SAM 3 README, LICENSE, release notes and repository activity, the Hugging Face file listings for facebook/sam3, facebook/sam3.1 and Comfy-Org/sam3.1, ComfyUI’s SAM 3 source code, templates, pull requests and issues, and the ComfyUI-SAM3 custom node were all read on that date.

⚡ Free Interactive Tools

Work out what your hardware can actually run

Free calculators and directories, sized from real model weights — before you download gigabytes or order a card:

🎨 ComfyUI VRAM Peak memory per image model 🚀 Start: Local LLMs Hardware, runner and first model 📊 What Can I Run? Models by VRAM tier, 8GB–128GB 💻 Hardware Matrix 75 laptops & GPUs profiled

Filed Under: Technology, Tutorial

About Rupesh Kumar

Rupesh Kumar founded Best Free Web Resources in 2010 and edits the site. He writes about free web resources, web development and running open AI models locally. Learn more: About BFWR & our editorial standards → · Follow on X (@bestfreewebres) →

Primary Sidebar

⚡ FREE INTERACTIVE TOOLS

Local AI Hardware & VRAM Matrix

Benchmark your PC or Mac for local AI, calculate VRAM requirements, and compare cloud API costs vs local hardware ROI.

🚀 New here? Start with Local LLMs → 💻 Hardware Matrix (79 Devices) → 📚 Local LLM Directory (30 Models) → 🎮 AI GPU Comparison → ⚡ VRAM & KV Cache Calculator → 💰 API Pricing & Break-Even → 🍎 Mac VRAM Unlocker (90% Boost) →
📦 Ollama 🖥️ LM Studio ⚙️ llama.cpp 🚦 vLLM 💬 Open WebUI 🎨 ComfyUI 🔒 Self-Hosted 📊 VRAM Tiers 🎯 Recommender ⚖️ Comparisons

Search

  • Facebook
  • Instagram
  • RSS
  • Twitter
  • YouTube

Footer

Latest Guides

  • SAM 3 Locally: Files, ComfyUI Nodes, Prompts and the Licence
  • Jan With Local Models: Setup, the llama.cpp Engine and the Local API Server
  • Free Seamless Textures in 2026: CC0 Sources, Photoshop Use and What the Licences Allow
  • GPT4All in 2026: Is It Still Maintained, and What to Use Instead
  • ComfyUI Manager Not Showing: Why, and How to Fix It in 2026
  • MiniMax H3 Image in ComfyUI: Files, VRAM, Settings and the Licence

AI Tools & Calculators

  • What can I run?
  • VRAM Calculator
  • GGUF Size Calculator
  • Fine-tuning VRAM
  • LoRA Training VRAM
  • Image Generation VRAM
  • Mac VRAM Optimizer

Browse & Compare

  • Models Directory (30 models)
  • Hardware Matrix (79 devices)
  • Image Models
  • Model Licences
  • Comparisons
  • Model Recommender
  • API Pricing & ROI

Guides & Resources

  • Technology
  • Coding
  • Tutorials
  • Freebies
  • Graphics

Copyright © 2026 · Best Free Web Resources · All Rights Reserved.  |  About · Privacy Policy · Contact · AI Tools Hub