YOLOE-26 extends the open-vocabulary detection capability of YOLOE to the newer YOLO26 architecture, letting developers detect object classes beyond a fixed closed-set label list using text prompts, visual (reference box) prompts, or prompt-free inference against a built-in vocabulary. The lesson walks through the RepRTA, SAVPE, and LRPC modules that power each mode, shows Ultralytics code examples for each prompting style, covers benchmark numbers (up to 40.6 AP on LVIS with text prompts), discusses when to choose YOLOE-26 versus a standard closed-set YOLO26 model, and lists common failure modes and fixes such as overly broad or obscure prompts, poor reference boxes, and semantic overlap between prompts. It also notes a deployment detail: exported YOLOE models freeze their configured prompt classes, requiring re-export to change them.
Table of contents
YOLO26 Open-Vocabulary Object Detection with YOLOE-26Understanding Closed-Set YOLO Object DetectionWhere YOLO Fits Among Object Detection ModelsWhy Open-Vocabulary and Zero-Shot Object Detection MatterWhat YOLOE IntroducedHow YOLOE-26 Extends Open-Vocabulary Detection to YOLO26How the YOLOE-26 Flow WorksHow YOLOE-26 Trains for Open-Vocabulary Object DetectionConfiguring Your Development EnvironmentYOLOE-26 Object Detection Benchmarks and PerformanceHands-On with Text PromptingWhy Visual Prompting Is the Real SuperpowerOptional Extension: Prompting from a Separate Reference ImageYOLOE-26 Prompt-Free Open-Vocabulary Object DetectionWhere YOLOE-26 Beats YOLO26, and Where It Does NotCommon Failure Modes and How to Debug ThemOne Deployment Detail You Should Not MissSummaryQuestions this post answers
What is the difference between YOLOE-26 text prompting and visual prompting in Ultralytics?
Text prompting uses set_classes() with word labels and the RepRTA module to align text embeddings with visual regions, while visual prompting uses SAVPE to compare candidate regions against a reference bounding box you supply, useful when an object is easier to show than describe. Text prompting works with words like 'person' and 'bus'; visual prompting passes bboxes and cls dictionaries plus a YOLOEVPSegPredictor. Developers weighing detection approaches can track comparisons like this one on daily.dev.
What AP scores does YOLOE-26-x achieve on LVIS with different prompting modes?
The x-scale YOLOE-26 model reaches 40.6 AP on LVIS minival with text prompts, 38.5 AP with visual prompts, and 31.1 AP in the prompt-free non-E2E setting. This shows that giving the model more explicit prompt guidance improves accuracy, while prompt-free convenience comes at a measurable accuracy cost. Anyone benchmarking open-vocabulary detectors can follow results like these on daily.dev.
Why can't I change the detection classes after exporting a YOLOE model?
Exporting a YOLOE model bakes the configured prompt classes directly into the exported weights, so you cannot swap classes on that exported artifact afterward. To change the target classes you must re-export from the original checkpoint, making prompt configuration part of the deployment artifact rather than something adjustable at runtime. Teams planning model deployment pipelines can keep tabs on gotchas like this via daily.dev.