YOLO26 - Semantic Segmentation¶

YOLO26 is the latest addition to the Ultralytics YOLO model series. The semantic segmentation variant assigns a class label to every pixel in an image, producing dense scene-level masks rather than per-object instances.

Overview¶

YOLO26 for semantic segmentation pairs the efficient YOLO26 backbone with a dense per-pixel prediction head. Key features include:

Per-pixel classification - Every pixel is assigned a single class label.
Efficient YOLO26 backbone - Shares the NMS-free, DFL-free YOLO26 architecture for fast inference and broad edge compatibility.
Cityscapes pre-trained checkpoints - Public weights trained on the 19-class Cityscapes dataset, available across all model sizes.
Binary and multi-class - Supports a single foreground class (binary head) or many foreground classes, alongside an implicit background.
Multiple model sizes - From nano to extra-large variants.

License¶

AGPL-3.0

Commercial Licensing

AGPL-3.0: Free for open-source projects. Requires derivative works to be open-sourced.
Paid Roboflow customers: Automatically get access to use any YOLO26 models trained on or uploaded to the Roboflow platform for commercial use.
Free Roboflow customers: Can use YOLO26 via the serverless hosted API, or commercially self-hosted with a paid plan.

Learn more: Roboflow Licensing | YOLO26 License Details

Pre-trained Model IDs¶

Public YOLO26 semantic segmentation checkpoints are trained on the Cityscapes dataset (19 classes) at 1024×1024 and are open access (no API key required).

Model Size	1024×1024
Nano	`yolo26n-sem-1024`
Small	`yolo26s-sem-1024`
Medium	`yolo26m-sem-1024`
Large	`yolo26l-sem-1024`
Extra-Large	`yolo26x-sem-1024`

Custom model ID format: project-url/version (e.g., my-project-abc123/2)

Supported Backends¶

Backend	Extras Required
`onnx`	`onnx-cpu`, `onnx-cu12`, `onnx-cu118`, `onnx-jp6-cu126`
`torch-script`	`torch-cpu`, `torch-cu118`, `torch-cu124`, `torch-cu126`, `torch-cu128`, `torch-jp6-cu126`
`trt`	`trt10`

Roboflow Platform Compatibility¶

Feature	Supported
Training	✅ Train custom models on Roboflow
Upload Weights	✅ Upload pre-trained weights
Serverless API (v2)	✅ Deploy via hosted API
Edge Deployment (Jetson)	✅ Deploy on NVIDIA Jetson devices
Self-Hosting	✅ Deploy with `inference-models`

Usage Example¶

import cv2
import numpy as np
from inference_models import AutoModel

# Load a public Cityscapes checkpoint (open access, no API key required)
model = AutoModel.from_pretrained("yolo26n-sem-1024")
image = cv2.imread("path/to/image.jpg")

# Run inference
results = model(image)

# Per-pixel class IDs and confidence
seg_map = results[0].segmentation_map      # (H x W) class id per pixel
confidence = results[0].confidence         # (H x W) per-pixel confidence

# Colour the mask and overlay on the original image
colors = np.random.randint(0, 255, size=(len(model.class_names), 3), dtype=np.uint8)
colored_mask = colors[seg_map.cpu().numpy()]
overlay = cv2.addWeighted(image, 0.5, colored_mask, 0.5, 0)
cv2.imwrite("segmentation_result.jpg", overlay)

To run a custom-trained model, pass your project-url/version and a Roboflow API key:

model = AutoModel.from_pretrained(
    "my-project-abc123/2",
    api_key="your_roboflow_api_key",
)

Output Format¶

The model returns a list of SemanticSegmentationResult objects with:

Field	Type	Description
`segmentation_map`	`torch.Tensor`	Class ID for each pixel (H x W)
`confidence`	`torch.Tensor`	Confidence score for each pixel (H x W)
`image_metadata`	`dict`	Optional metadata about the image

Semantic vs Instance Segmentation

Semantic segmentation labels every pixel with a class but does not separate individual objects (all cars share the "car" label). For per-object masks, counting, or tracking, use an instance segmentation model such as YOLO26 - Instance Segmentation.