yoloe

YOLOE is a unified object detection and segmentation model that supports text prompts, visual inputs, and prompt-free inference within a single architecture, optimized for real-time performance.

Key capabilities

• Text prompts – Re-parameterizable Region-Text Alignment (RepRTA) refines textual embeddings with zero inference overhead.
• Visual prompts – Semantic-Activated Visual Prompt Encoder (SAVPE) improves visual embedding and accuracy with minimal complexity.
• Prompt-free mode – Lazy Region-Prompt Contrast (LRPC) uses a built-in vocabulary to identify all objects without a language model.

Benchmark highlights

• On LVIS minival, YOLOE-v8-S surpasses YOLO-Worldv2-S by 3.5 AP with 3× less training cost and 1.4× inference speedup.
• Transferring to COCO, YOLOE-v8-L achieves 0.6 AP and 0.4 AP gains over closed-set YOLOv8-L with nearly 4× less training time.
• After re-parameterization, YOLOE becomes a standard YOLO with zero inference or transferring overhead.

Zero-shot detection on LVIS (text / visual prompts)

YOLOE is best suited for applications requiring real-time detection and segmentation across diverse, open-ended object categories without retraining. It is hosted on Gigarouter as a managed, OpenAI-compatible API.