Source-linked AI summary
YOLO26: A Comprehensive Architecture Overview and Key Improvements
Priyanto Hidayatullah, Refdinal Tubagus
TL;DR
YOLO26 addresses the need for a precise architectural account of a model intended to improve edge-device object detection. The paper investigates its source code and documentation, producing an architecture diagram and analyzing its main refinements. It reports simplified inference, reduced computational cost, and claimed CPU-speed gains of up to 43%.
Problem
Existing technical documentation does not fully expose YOLO26’s architectural mechanisms, creating a need for a precise architectural account for researchers and developers.
Method
The study reconstructs YOLO26 from its source code and official documentation, validates the architecture through tensor tracing, and examines its architectural and training improvements.
Results
YOLO26 refines its architecture to simplify inference, reduce computational cost, and improve accuracy, especially for small-object detection; developers claim up to 43% faster CPU-mode speed.
Takeaways & Limitations
The analyzed refinements position YOLO26 for edge-device use while retaining a single-stage, end-to-end object-detector design.
Abstract
from arXiv · showhide
You Only Look Once (YOLO) has been the prominent model for computer vision in deep learning for a decade. This study explores the novel aspects of YOLO26, the most recent version in the YOLO series. The elimination of Distribution Focal Loss (DFL), implementation of End-to-End NMS-Free Inference, introduction of ProgLoss + Small-Target-Aware Label Assignment (STAL), and use of the MuSGD optimizer are the primary enhancements designed to improve inference speed, which is claimed to achieve a 43% boost in CPU mode. This is designed to allow YOLO26 to attain real-time performance on edge devices or those without GPUs. Additionally, YOLO26 offers improvements in many computer vision tasks, including instance segmentation, pose estimation, and oriented bounding box (OBB) decoding. We aim for this effort to provide more value than just consolidating information already included in the existing technical documentation. Therefore, we performed a rigorous architectural investigation into YOLO26, mostly using the source code available in its GitHub repository and its official documentation. The authentic and detailed operational mechanisms of YOLO26 are inside the source code, which is seldom extracted by others. The YOLO26 architectural diagram is shown as the outcome of the investigation. This study is, to our knowledge, the first one presenting the CNN-based YOLO26 architecture, which is the core of YOLO26. Our objective is to provide a precise architectural comprehension of YOLO26 for researchers and developers aspiring to enhance the YOLO model, ensuring it remains the leading deep learning model in computer vision.
1 Introduction
YOLO26 focuses on improving edge-device efficiency while refining the existing YOLO design. The paper investigates its architecture through source-code analysis and presents an overall architectural diagram and key improvements.
- Motivation: YOLO26 prioritizes faster inference for edge devices while preserving the broader Ultralytics YOLO task coverage.The model supports object detection, instance segmentation, classification, pose estimation, oriented-object detection, and tracking.
- Motivation: YOLO26 extends existing YOLO functionality rather than introducing entirely new computer-vision task categories.The paper states that YOLO functionality was already implemented in YOLOv8, while YOLO26 improves additional capabilities.
- Research gap: The paper identifies a missing architectural diagram as a barrier to understanding and advancing YOLO26.It connects visual architectural documentation with researchers’ and developers’ ability to comprehend and improve the model.
- Research gap: Technical documentation alone does not reveal the placement and operation of architectural modifications such as Distribution Focal Loss.The authors therefore inspect YOLO26 source code and present what they describe as the first overall YOLO26 architecture diagram.
- Contribution: YOLO26 retains a single-stage, end-to-end detector design while adding changes intended to improve efficiency, stabilize training, and remove prediction-stage NMS dependence.The study reports that these refinements seek higher accuracy and better edge-device performance relative to the predecessor.
2 Evolution of YOLO Models
YOLO has evolved through successive architectural changes aimed at improving real-time detection. YOLO26 continues this trajectory with non-NMS detection, dual label assignment, DFL removal, MuSGD, and ProgLoss with STAL.
- Model evolution: YOLO development has progressed rapidly, with multiple versions introducing successive architectural improvements for real-time object detection.The paper reports 11 versions from 2020 to 2026 after YOLOv4.
- Model evolution: Architectural evolution added deeper efficient backbones, refined feature aggregation, and anchor-free heads to improve speed and accuracy.Examples include Darknet-53, CSPDarknet, EfficientRep, PANet, SPP, ELAN, GELAN, and anchor-free head designs.
- YOLO26 evolution: YOLO26 advances recent flexible-design trends through non-NMS detection with dual label assignment.The described changes also include Distribution Focal Loss elimination, MuSGD, ProgLoss, and Small-Target-Aware Label Assignment.
3 YOLO26 Architecture
The YOLO26 architecture uses configurable backbone, neck, and head components derived from source code and validated through tensor tracing. Its three detection heads target small, medium, and large objects.
- Architecture methodology: The architecture diagram is derived from YOLO26 configuration and task source files, then checked through input-output tensor tracing.The source came from Ultralytics release 8.4.14, including yolo26.yaml and tasks.py.
- Architecture configuration: YOLO26 uses depth, width, and maximum-channel parameters to configure its modules and processes three-channel images through backbone, neck, and head.Depth controls bottleneck and PSA repetitions, while width and maximum channels determine block output channels.
- Backbone: The backbone begins with two stride-2, kernel-3 convolutional blocks, followed by C3k2 and additional convolutional and C3k2 stages linked to the neck.Stride 2 halves feature-map spatial resolution at each convolutional block.
- Neck: The neck combines SPPF with a shortcut, C2PSA self-attention, and repeated upsampling and concatenation operations.The shortcut incorporates input into the SPPF output, while C2PSA adds global modeling capabilities.
- Head architecture: An attention block is added to the final C3k2 block before the head.The supplied architecture description identifies this as a new component in that final block.
- Detection heads: YOLO26 has three detection heads specialized for small, medium, and large objects.The heads connect to C3k2 blocks 16, 19, and 22, respectively.
4 YOLO26 Key Improvements Compared to Previous YOLO versions
YOLO26 refines the YOLO11-like architecture through training, inference, and feature-processing changes aimed at efficiency, stability, and accuracy. Key changes include MuSGD, NMS-free prediction, DFL removal, and modified SPPF and attention blocks.
- SPPF modification: YOLO26 adds a shortcut to SPPF, enhancing gradient transmission and stabilizing optimization for high-level semantic representations.
- Attention modification: YOLO26 sets the final C3k2 repeat count to 1 and adds PSABlock attention to improve global context modeling with limited parameter and latency increases.
- Detection head: YOLO26 removes DFL from the detection head and uses explicit box-coordinate regression, streamlining training and inference.
- NMS-free design: During training, YOLO26 uses one-to-many and one-to-one assignments, then retains only the one-to-one head during inference.
- Training mechanisms: ProgLoss shifts training emphasis from the one-to-many head toward the one-to-one head, while STAL ensures very small objects contribute to training.
- Optimizer: MuSGD combines Muon-style and SGD updates for selected parameters while retaining pure SGD for others, supporting smoother and more predictable training.
- NMS-free design: End-to-end inference ranks predictions by global classification confidence with Top-K selection, avoiding IoU computation and NMS.
- Efficiency: 43% is the developers’ claimed maximum CPU-mode speed increase, supporting deployment on edge devices or devices without GPUs.
5 YOLO26 Computer Vision Specific Tasks Improvement
YOLO26 extends task-specific improvements across segmentation, pose estimation, and oriented bounding boxes while supporting its broader computer-vision task suite. These changes target mask quality, keypoint localization, decoding speed, and rotated-object accuracy.
- Supported tasks: The broader Ultralytics YOLO task suite includes object detection, segmentation, classification, pose estimation, oriented-object detection, and tracking.
- Instance segmentation: YOLO26 adds semantic segmentation loss and a multi-scale proto module to improve convergence and mask quality.
- Pose estimation: YOLO26 integrates Residual Log-Likelihood Estimation and optimized decoding for more accurate keypoint localization and faster pose inference.
- Oriented bounding boxes: YOLO26 adds specialized angle loss and optimized OBB decoding to improve square-shaped-object detection and address boundary discontinuity.
6 Performance Comparison
The comparison uses COCO mAP and NVIDIA T4 speed measurements to compare YOLO26 with earlier YOLO versions. YOLO26 has the highest reported accuracy across variants, while only its s and m variants match YOLO11’s speed.
- Accuracy: YOLO26 has the highest accuracy among compared YOLO versions across all variants in the reported comparison.Accuracy is evaluated with mAP on COCO.
- Speed: Only the YOLO26 s and m variants are reported as fastest, with speed equivalent to YOLO11.Speed was measured using an NVIDIA GPU T4.
7 Conclusion
The paper concludes that YOLO26 is an evolutionary refinement that improves architectural and training mechanisms for accuracy, simpler inference, and lower computational cost. Its design is especially suited to edge-device deployment and small-object detection.
- Overall assessment: YOLO26 refines rather than disruptively redesigns YOLO, improving architectural and training mechanisms compared with previous versions.
- Overall assessment: The paper concludes that YOLO26 simplifies inference, reduces computational cost, and improves accuracy, especially for small-object detection.