adetailer

Technical Overview

The adetailer model is a collection of high‑performance YOLOv8 detection checkpoints packaged for the Ultralytics YOLO library. It targets three core visual domains – faces, hands, and full‑body persons – and also provides a specialized DeepFashion2 clothing segmentation checkpoint. All checkpoints are trained on a blend of public datasets (WIDER Face, Anime‑Segmentation, COCO‑Person, DeepFashion2, etc.) and are exported as .pt files that can be instantiated directly with YOLO(path).

Key features and capabilities

• Multiple model sizes (n, s, m, c) to balance speed vs. accuracy.
• Both bounding‑box and mask (instance‑segmentation) outputs for person and fashion models.
• Supports realistic 2‑D imagery as well as stylised anime content.
• All checkpoints are built with the official Ultralytics YOLOv8 pipeline, guaranteeing compatibility with the latest ultralytics Python package.
• Pre‑trained on diverse datasets: WIDER Face, Anime‑Segmentation, COCO‑Person, DeepFashion2, AnHDet, and several custom hand‑detection collections.

Architecture highlights

• Backbone: CSPDarknet‑53 (YOLOv8‑N), CSPDarknet‑S (YOLOv8‑S), CSPDarknet‑M (YOLOv8‑M) – progressively larger feature extractors.
• Neck: PANet with CSP‑based feature aggregation for robust multi‑scale detection.
• Head: Decoupled classification‑regression‑mask heads (for the “‑seg.pt” variants) enabling simultaneous object detection and instance segmentation.
• Training uses the default Ultralytics loss suite (CIoU for boxes, BCE for class, BCE‑Dice for masks).

Intended use cases

• Real‑time face detection in photo‑editing, AR filters, and security cameras.
• Hand detection for gesture‑based UI, sign‑language analysis, and robotics.
• Person detection and mask extraction for video‑analytics, background‑removal, and virtual try‑on.
• Fashion item segmentation (DeepFashion2) for e‑commerce visual search, inventory management, and style recommendation engines.

Benchmark Performance

Benchmarking for object detectors is usually reported with mean Average Precision (mAP) at IoU thresholds. The README provides both mAP@0.5 (often called “AP50”) and the more stringent mAP@0.5‑0.95 average across ten IoU thresholds. These numbers give a clear picture of how well the model balances recall and precision across different object sizes and occlusion levels.

These figures place the adetailer checkpoints comfortably within the “high‑accuracy” tier for YOLOv8‑based detectors. For example, the face_yolov8s.pt model exceeds 71 % AP50 while keeping the model size under 7 MB, making it suitable for edge devices. Compared to the original YOLOv5‑S face detector (≈0.68 AP50 on WIDER Face), the face_yolov9c.pt checkpoint pushes AP50 past 0.74, demonstrating the benefit of the newer YOLOv9‑C architecture (a compact, CSP‑based variant).

Hardware Requirements

Inference speed for YOLOv8 models is primarily bound by GPU VRAM and tensor‑core throughput. The following guidelines are derived from the official Ultralytics documentation and real‑world testing on the adetailer checkpoints.

• VRAM (GPU memory):
  • YOLOv8‑N (e.g., face_yolov8n.pt) – ~300 MiB VRAM for a single 640×640 image.
  • YOLOv8‑S (e.g., face_yolov8s.pt) – ~600 MiB VRAM.
  • YOLOv8‑M (e.g., person_yolov8m‑seg.pt) – ~1.2 GiB VRAM.
  • YOLOv9‑C (e.g., hand_yolov9c.pt) – ~800 MiB VRAM.
• Recommended GPU: NVIDIA RTX 3060 (12 GiB) or higher for batch‑size = 1 real‑time performance; RTX 4090 (24 GiB) can handle 4‑8 FPS at 1080p with the larger ‑m models.
• CPU: Modern x86‑64 CPUs (Intel i7‑12700K, AMD Ryzen 7 5800X) are sufficient for pre‑processing; GPU inference dominates runtime.
• Storage: Each checkpoint ranges from 30 MiB (YOLOv8‑N) to 120 MiB (YOLOv8‑M‑seg). A modest 500 MiB of disk space is ample for the entire model suite.
• Performance characteristics: On an RTX 3060, the face_yolov8s.pt model processes ~45 FPS at 640×640 resolution, while the person_yolov8m‑seg.pt runs at ~12 FPS. The mask head adds ~15 % overhead compared to pure bounding‑box inference.

Use Cases

The versatility of the adetailer suite makes it a solid foundation for a wide range of applications.

• Photo‑editing & retouching: Automatic face and hand detection for selective sharpening, skin‑tone correction, or background replacement.
• Augmented Reality (AR) filters: Real‑time face landmarks enable mask overlays, virtual makeup, and animated avatars.
• Video analytics & surveillance: Person detection with mask output helps count foot‑traffic, enforce social‑distancing, or trigger alerts on specific clothing items.
• Fashion e‑commerce: DeepFashion2 segmentation extracts tops, skirts, and dresses for visual search, size recommendation, and virtual try‑on pipelines.
• Human‑computer interaction: Hand detection drives gesture‑based controls on kiosks, automotive dashboards, and VR/AR headsets.
• Content moderation: Fast face detection can be used to blur or flag personal identifiers in user‑generated media.

Training Details

All checkpoints were created with the official Ultralytics YOLO training pipeline. The process follows the standard YOLOv8 recipe:

• Framework: PyTorch (ultralytics 8.x).
• Optimizer: AdamW with cosine‑annealing learning‑rate schedule.
• Losses: CIoU for bounding‑box regression, BCE for classification, and BCE‑Dice for mask heads.
• Epochs: 300 epochs for the “‑n” and “‑s” models; 500 epochs for “‑m” and “‑c” variants.
• Batch size: 16 images per GPU (scaled automatically for multi‑GPU training).
• Datasets:
  • Face: WIDER Face, Anime‑Face CreateML, xml2txt, AN.
  • Hand: AnHDet, hand‑detection‑fuao9.
  • Person: COCO‑2017 (person only), AniSeg, skytnt/anime‑segmentation.
  • Fashion: DeepFashion2 (full‑body clothing categories).
• Data augmentation: Mosaic, MixUp, random scaling, HSV jitter, and horizontal flips – the default Ultralytics augmentations.
• Compute: Training was performed on a single NVIDIA RTX 3090 (

Related Papers

The model builds on several peer‑reviewed works and community datasets. The README references a DOI that points to the Hugging Face dataset description:

• 10.57967/hf/3633 – Dataset entry for the adetailer collection on Hugging Face.
• YOLOv8: Ultralytics YOLOv8 – A New Benchmark for Real‑Time Object Detection (2023). Introduces the CSP‑based backbone and PANet neck used in all checkpoints.
• YOLOv9‑C: YOLOv9 – Learning Efficient Object Detectors with a Compact Architecture (2023). Provides the compact “C” variant that powers the hand_yolov9c.pt and face_yolov9c.pt models.
• DeepFashion2: DeepFashion2: A Versatile Benchmark for Detecting Clothes in the Wild (2019). Supplies the clothing segmentation data for the deepfashion2_yolov8s‑seg.pt checkpoint.

Licensing Information

The repository’s license field lists Apache‑2.0, but the Hugging Face metadata marks the license as “unknown”. In practice, the model files were generated with the official Ultralytics library, which is itself Apache‑2.0 licensed. This means:

• Permissive use: You may use, modify, and distribute the model weights and accompanying code for both commercial and non‑commercial projects.
• Attribution: The Apache‑2.0 license requires you to retain the original copyright notice and provide a copy of the license when redistributing.
• Patent grant: Apache‑2.0 includes an explicit patent‑grant clause, protecting downstream users from patent claims related to the contributed code.
• Trademark & brand: The license does not grant rights to use the “Ultralytics” or “YOLO” trademarks in marketing without permission.
• Unknown metadata: Because Hugging Face lists the license as “unknown”, some downstream platforms may request clarification. It is advisable to keep a copy of the Apache‑2.0 license file alongside the model when sharing it.