models/sam-3/

[giscus[bot]]

models/sam-3/

Discover SAM 3, Meta's next evolution of the Segment Anything Model, introducing Promptable Concept Segmentation with text and image exemplar prompts for detecting all instances of visual concepts across images and videos.

https://docs.ultralytics.com/models/sam-3/

[glenn-jocher]

SAM 3 looks like a big step forward for open-vocabulary segmentation, and we plan to support it natively in ultralytics once Meta releases the official code and weights. Until then, this page is a preview based on Meta's ICLR 2026 submission paper.

We’ll move fast once it’s open-sourced. Feel free to drop questions or comments for the team below!

[UltralyticsAssistant]

👋 Hello, thanks for starting this discussion about SAM 3 in Ultralytics 🚀 This is an automated response to help you get started quickly—an Ultralytics engineer will also assist you here soon.

We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

Feel free to share any specific details or goals for your SAM 3 use case (logs, configs, steps to reproduce, screenshots). This will help us assist you faster 🙌

[koolfox]

Incredible, when you think it’s available for use?

[glenn-jocher]

Later this week I think, YOLO team is working on integration now!

[Abdulhady-WITCo]

Appreciate the effort, and can't wait for your implementation. It's resource heavy already but I'm sure your optimisation will make a difference. Best of luck.

[glenn-jocher]

Thanks for the kind words—most of the credit goes to the Meta SAM team and the Ultralytics community, and once SAM 3 lands in ultralytics we’ll be focusing exactly on those trade-offs so you can balance its open-vocabulary power against lighter options like YOLO11 or SAM 2 on your hardware.

[aliabdollahi-99]

Hi dear ultralytics team, it seems that now fine tuning of sam3 is not available in ultralytics yet. do you have plan to make it possible? if yes, Approximately how long? thanks

[glenn-jocher]

You’re right, SAM 3 fine-tuning isn’t available in Ultralytics yet; the initial integration will be inference-only (similar to our current SAM 2 support in the SAM 2 docs), and while we do plan to explore fine-tuning, there’s no timeline we can share, so please keep an eye on future releases for updates.

[nam787878]

Thanks for the hard work!
Do you have any updated timeline or estimated date for the initial integration(inference-only) ?
I'm really looking forward to using it.

[glenn-jocher]

Thanks for the enthusiasm—SAM 3 integration is actively in progress, but we don’t have an updated ETA to share; please keep an eye on upcoming ultralytics versions and the Ultralytics release notes for when initial inference-only sam3 support is announced.

[Iverson-zq]

Hello, excuse me, I would like to ask if SAM3 can support such an input situation?
For example, given the information of a human detection box, it is prompted to divide a certain part of the human body in the detection box.

[glenn-jocher]

Yes—SAM 3 can do this, but you’ll generally want to constrain the query to the person ROI: run a person detector (e.g., Ultralytics YOLO11) to get the person box, crop that region, then use SAM3SemanticPredictor with a body-part text prompt like "hand"/"face" (or provide an exemplar box around the target part inside the crop for tighter control); the prompt types and requirements (including bpe_path for text) are covered in the SAM 3 docs.

[liorkupfer]

Hi there, am I the only one getting FileNotFoundError: [Errno 2] No such file or directory: 'sam3.pt'?

[glenn-jocher]

That error just means the SAM 3 weights aren’t on disk—sam3.pt is not auto-downloaded, so you need to manually grab it from the official SAM 3 repo and either place it in your working directory or pass the full path (see the “SAM 3 Model Weights Required” note in the SAM 3 docs).

from ultralytics import SAM

model = SAM("/absolute/path/to/sam3.pt")
results = model.predict("image.jpg", points=[900, 370], labels=[1])

[liorkupfer]

gotcha, thank you very much!

[glenn-jocher]

Glad it’s working—if you move to text-based concept prompts (SAM3SemanticPredictor / SAM3VideoSemanticPredictor), make sure you also have the BPE vocab and pass bpe_path as shown in the SAM 3 docs.

[iris01brown]

Does SAM3 now support exporting to ONNX and TensorRT？

[glenn-jocher]

Yes—if you’re on ultralytics>=8.3.237, SAM 3 can be exported via the standard Ultralytics export workflow (see the SAM 3 docs and Export mode docs); TensorRT export requires a CUDA GPU (device=0):

yolo export model=sam3.pt format=onnx imgsz=1024
yolo export model=sam3.pt format=engine imgsz=1024 device=0

If you hit any exporter errors, please paste the full traceback and the output of yolo checks so we can reproduce.

[iris01brown]

Thanks for your reply !
Does this currently only support square images?
Because when I tried yolo export model="D:\Danny\Code\sam3-main\weights\sam3.pt" format=engine imgsz=(780,1920) device=0
I got this error message:

Traceback (most recent call last):
File "", line 198, in run_module_as_main
File "", line 88, in run_code
File "D:\Danny\Tools\anaconda3\envs\sam3\Scripts\yolo.exe_main.py", line 6, in
File "D:\Danny\Tools\anaconda3\envs\sam3\Lib\site-packages\ultralytics\cfg_init.py", line 906, in entrypoint
check_dict_alignment(full_args_dict, {a: ""})
File "D:\Danny\Tools\anaconda3\envs\sam3\Lib\site-packages\ultralytics\cfg_init_.py", line 505, in check_dict_alignment
raise SyntaxError(string + CLI_HELP_MSG) from e
SyntaxError: '1920' is not a valid YOLO argument.

Arguments received: ['yolo', 'export', 'model=D:\\Danny\\Code\\sam3-main\\weights\\sam3.pt', 'format=engine', 'imgsz=', '780', '1920', 'device=0']. Ultralytics 'yolo' commands use the following syntax:

    yolo TASK MODE ARGS

    Where   TASK (optional) is one of ['detect', 'obb', 'segment', 'classify', 'pose']
            MODE (required) is one of ['benchmark', 'predict', 'val', 'train', 'export', 'track']
            ARGS (optional) are any number of custom 'arg=value' pairs like 'imgsz=320' that override defaults.
                See all ARGS at https://docs.ultralytics.com/usage/cfg or with 'yolo cfg'

1. Train a detection model for 10 epochs with an initial learning_rate of 0.01
    yolo train data=coco8.yaml model=yolo11n.pt epochs=10 lr0=0.01

2. Predict a YouTube video using a pretrained segmentation model at image size 320:
    yolo predict model=yolo11n-seg.pt source='https://youtu.be/LNwODJXcvt4' imgsz=320

3. Validate a pretrained detection model at batch-size 1 and image size 640:
    yolo val model=yolo11n.pt data=coco8.yaml batch=1 imgsz=640

4. Export a YOLO11n classification model to ONNX format at image size 224 by 128 (no TASK required)
    yolo export model=yolo11n-cls.pt format=onnx imgsz=224,128

5. Ultralytics solutions usage
    yolo solutions count or any of ['crop', 'blur', 'workout', 'heatmap', 'isegment', 'visioneye', 'speed', 'queue', 'analytics', 'inference', 'trackzone'] source="path/to/video.mp4"

6. Run special commands:
    yolo help
    yolo checks
    yolo version
    yolo settings
    yolo copy-cfg
    yolo cfg
    yolo solutions help

Docs: https://docs.ultralytics.com
Solutions: https://docs.ultralytics.com/solutions/
Community: https://community.ultralytics.com
GitHub: https://github.com/ultralytics/ultralytics

[iris01brown]

yolo checks output:

Ultralytics 8.3.239 Python-3.12.12 torch-2.7.0+cu126 CUDA:0 (NVIDIA GeForce RTX 4050 Laptop GPU, 6140MiB)
Setup complete (24 CPUs, 15.7 GB RAM, 84.4/300.0 GB disk)
OS Windows-11-10.0.26200-SP0
Environment Windows
Python 3.12.12
Install pip
Path D:\Danny\Tools\anaconda3\envs\sam3\Lib\site-packages\ultralytics
RAM 15.71 GB
Disk 84.4/300.0 GB
CPU Intel Core i7-14650HX
CPU count 24
GPU NVIDIA GeForce RTX 4050 Laptop GPU, 6140MiB
GPU count 1
CUDA 12.6

numpy 2.2.6>=1.23.0
matplotlib 3.10.7>=3.3.0
opencv-python 4.12.0.88>=4.6.0
pillow 12.0.0>=7.1.2
pyyaml 6.0.3>=5.3.1
requests 2.32.5>=2.23.0
scipy 1.16.3>=1.4.1
torch 2.7.0+cu126>=1.8.0
torch 2.7.0+cu126!=2.4.0,>=1.8.0; sys_platform == "win32"
torchvision 0.22.0+cu126>=0.9.0
psutil 7.1.3>=5.8.0
polars 1.36.1>=0.20.0
ultralytics-thop 2.0.18>=2.0.18

[glenn-jocher]

That error is just the CLI parsing your PowerShell tuple, not a “square-only” limitation—pass imgsz as a comma-separated pair (and quote it on Windows so it isn’t split), per the Export mode docs and the tested workflow in the SAM 3 docs:

yolo export model="D:/Danny/Code/sam3-main/weights/sam3.pt" format=engine "imgsz=780,1920" device=0

If that gets past argument parsing but the exporter fails later, please paste the full export traceback (your yolo checks looks good).

[iris01brown]

(sam3) PS D:\Danny\Code\sam3-main> yolo export model="D:\Danny\Code\sam3-main\weights\sam3.pt" format=engine "imgsz=780,1920" device=0

Traceback (most recent call last):
File "", line 198, in run_module_as_main
File "", line 88, in run_code
File "D:\Danny\Tools\anaconda3\envs\sam3\Scripts\yolo.exe_main.py", line 6, in
File "D:\Danny\Tools\anaconda3\envs\sam3\Lib\site-packages\ultralytics\cfg_init.py", line 955, in entrypoint
model = YOLO(model, task=task)
^^^^^^^^^^^^^^^^^^^^^^
File "D:\Danny\Tools\anaconda3\envs\sam3\Lib\site-packages\ultralytics\models\yolo\model.py", line 76, in init
super().init(model=model, task=task, verbose=verbose)
File "D:\Danny\Tools\anaconda3\envs\sam3\Lib\site-packages\ultralytics\engine\model.py", line 144, in init
self._load(model, task=task)
File "D:\Danny\Tools\anaconda3\envs\sam3\Lib\site-packages\ultralytics\engine\model.py", line 283, in _load
self.model, self.ckpt = load_checkpoint(weights)
^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\Danny\Tools\anaconda3\envs\sam3\Lib\site-packages\ultralytics\nn\tasks.py", line 1463, in load_checkpoint
model = (ckpt.get("ema") or ckpt["model"]).float() # FP32 model
~~~~^^^^^^^^^
KeyError: 'model'

[Nehcknarf]

Nice work.
How to auto-annotate images using SAM3 at this stage? I wanna produce a YOLO OBB dataset with the power of SAM3 as the 1st step.
It seems the auto_annotate function is not yet support concept segmentation task?

[glenn-jocher]

@iris01brown KeyError: 'model' means the sam3.pt you’re pointing to isn’t an Ultralytics-format checkpoint (it’s likely a raw Meta state_dict), so the loader can’t find the expected model entry—please re-download the official weights exactly as described in the SAM 3 “Model Weights Required” section and then retry export.

@Nehcknarf correct, auto_annotate() doesn’t support SAM 3 concept prompts yet; for now you’ll need a small custom script using from ultralytics.models.sam import SAM3SemanticPredictor (see the SAM3SemanticPredictor reference) to generate masks from text/exemplar prompts and then convert those masks to OBB labels, and if you want first-class support please open a feature request on the Ultralytics issue tracker.

[hn5467]

Where can I find the image "image.jpg" and the video "video.mp4" to test the scripts?

[likebean]

I observed that for the same image and the same confidence (conf) parameter, the masks returned by YOLO + SAM3SemanticPredictor are different from the masks returned by the official SAM3 pipeline. Is this expected behavior?

[glenn-jocher]

Some difference is expected: Ultralytics’ SAM3SemanticPredictor runs SAM 3 through our own predict stack (image resizing/normalization, presence-score–gated confidence, mask thresholding, and postprocessing), so matching conf alone won’t necessarily reproduce the official SAM 3 pipeline outputs; details are in the SAM 3 docs and the SAM3SemanticPredictor reference.

If you want us to sanity-check parity, please paste a minimal repro (exact prompts you pass, and whether you’re feeding YOLO boxes vs text= concepts) plus your environment info from:

yolo checks

[haseebniz01]

It's an amazing plateform to access AI models for the edge devices. I would like to know when Ultralytics team will upload SAM3D weights. Thank you!

[glenn-jocher]

Thanks—credit goes to the wider YOLO community and Ultralytics team; SAM3D weights aren’t currently available/integrated on our side and we can’t give an ETA, so please follow the updates on the SAM 3 docs and watch Ultralytics GitHub releases, and if you share your exact SAM3D use case here we can track it as a feature request.

[jlee-detect]

Does the library allow you to use custom weights? If we've finetuned sam3 model weights from the SAM3 train scripts on the original SAM3 github page, what changes would be needed to import it into this setup?

[glenn-jocher]

Yes—custom SAM 3 weights should work without code changes as long as your finetune kept the same SAM 3 architecture and checkpoint/state-dict layout as the released sam3.pt; just point overrides['model'] (or SAM(...)) at your file path as shown in the SAM 3 usage docs. If your Meta training scripts saved a different checkpoint structure (or you changed the config/backbone), you’ll need to convert/rename keys to match what Ultralytics’ loader expects (see how checkpoints are constructed/loaded in the SAM 3 builder reference); if you paste the exact load traceback plus the top-level keys from torch.load(your_ckpt).keys() here, we can suggest the minimal conversion.

[maitianstudio-arch]

I noticed that the Ultralytics library currently does not support exporting SAM3 models to ONNX.

[glenn-jocher]

You’re correct—SAM 3 in Ultralytics currently supports predict/track but export (including ONNX) isn’t implemented yet; if ONNX/TensorRT deployment is a requirement, the practical path today is to use an Ultralytics YOLO26 segmentation model for your specific classes, and if you want SAM 3 ONNX export considered please open an issue with your target runtime/opset and a minimal example (and feel free to link back to the SAM 3 docs page).

[max-dev-14]

Hi there. I have a video and I want to segment it using SAM3. But there is one issue: the object desired to be segmented is in the middle of the video. I tried to segment my video using SAM3, and I found the model knows nothing about the object and segment it poorly (particularly, SAM3 segmented wrong objects) until the object appears in my video. Is it possible (using Ultralytics framework) to refer to a specific frame from which the model will start segmentation? As far as I know something like this is implemeted in SAM2 original repo.

[glenn-jocher]

We don’t currently expose a simple start_frame= arg in the high-level SAM 3 video API, so the usual approach is to start the input video at the first frame where the object is visible (or, for lower-level control, initialize on that frame via SAM3VideoSemanticPredictor.add_prompt(frame_idx=...); see the SAM3 video add_prompt() reference). For the “start at frame X” workflow, trimming is often simplest:

ffmpeg -ss 00:01:23.500 -i input.mp4 -c copy clipped.mp4

Then run the normal SAM 3 video semantic workflow from the SAM 3 docs; if you share your FPS + the frame index where the object first appears (and whether you’re using text or bboxes prompts), we can suggest the cleanest setup.

[14790897]

amazing!

[glenn-jocher]

Thanks! If you try it and hit any snags, drop your question here in the discussion (ideally with yolo checks output and the full traceback); also note that sam3.pt weights aren’t auto-downloaded, so you’ll need to request/download them first as shown in the SAM 3 docs.

[max-dev-14]

1. Can I parallelize sam3 inference on a video? For instance, 2 GPUs are available and I want to use both of them for the inference. The paramter 'imgsz' must be setted to 1008. The error 'OutOfMemoryError' occurs for single-GPU inference.
2. Is it possible to locate a video in RAM, then sam3 load one frame only and segment on it in GPU?

[glenn-jocher]

For SAM 3 video tracking (SAM3VideoSemanticPredictor/SAM3VideoPredictor) the pipeline is inherently sequential (tracker state carries frame-to-frame), so we don’t currently support “one video, multiple GPUs” parallel inference without splitting the video into independent chunks (which will break/renumber identities at the boundaries); if you only need per-frame concept masks (no temporal tracking), you can parallelize by sharding frames across processes/GPUs using SAM3SemanticPredictor as shown in the SAM 3 docs and the SAM predictor reference. Loading the whole video into RAM won’t fix CUDA OOM (VRAM is the limiter); the recommended approach is already “stream frames and run one frame at a time”, so please paste your full traceback (text) and yolo checks output here (or in-notebook run !yolo checks) along with your exact command/snippet so we can suggest the best memory reduction for your case (e.g., half=True, fewer prompts/objects, or chunking).

[max-dev-14]

I'd compared facebook/sam3 and ultralytics/sam3. I set up both the models for inference as follow:
imgsz=1008, half=True, conf=0.5 and prompts = [
"person",
"car",
"truck",
"tractor",
"building",
"combine harvester",
"power lines",
"clear sky",
"trees",
"crop field",
"harvested field",
"grassy field",
"dark frame edges",
"farm lane",
"windrow"
]
I'd attached scripts I used for inference:
ul-sam3.py - ultralytics/sam3 inference.
fb-sam3.py - facebook/sam3 inference.

I expected that there would be the same results of inference on the same videos and paramteres above. But I'd got different segments, as it can be seen on the screens below (there are original frame to the left, facebook/sam3 segmentation on the middle and ultralytics/sam3 segmentation to the right).

Can someone explain me why it's happened like this?

[glenn-jocher]

Thanks for raising it here — exact parity with facebookresearch/sam3 isn’t expected, because the Ultralytics integration uses its own SAM 3 predictor path for resizing/preprocessing and postprocessing, so matching imgsz, half, conf, and prompts alone won’t guarantee identical masks; see the SAM 3 docs and the predictor reference. If you’d like us to check whether this is a bug rather than an implementation difference, please open a new issue with a minimal reproducible example, confirm both runs use the same sam3.pt, and include yolo checks plus the full text outputs from both scripts.

[tomer196]

Hi, thanks for the great work.
I am trying to use sam3.1. Is it supported?
I tried using sam3.1 weights, and it seems to work, but each frame the objects get new IDs. Am I doing something wrong?
Thanks

[glenn-jocher]

sam3.1 is not documented as a separate Ultralytics variant yet, so if its weights load they’re effectively going through the current SAM 3 integration; for stable identities across frames, make sure you use SAM3VideoPredictor or SAM3VideoSemanticPredictor on the full video rather than running image inference frame-by-frame, since per-frame inference will assign fresh IDs each time—see the SAM predictor reference, and if you’re already doing that please drop a minimal repro plus yolo checks output here so the team can take a closer look.