General
The DIV2K super-resolution network real_esrgan_x2 now calibrates on a DIV2K train-split set; the previous calibration set was cut from the validation split.
New Models:
STARK-Lightning - First single-object tracker in the Model Zoo (stark_lightning_repvgg_a0).
MiVOLO - Age and gender estimation (mivolo_d1), 6-channel face+person input with face-only and body-only modes.
YamNet - First audio-classification network in the Model Zoo (yamnet), a 521-class AudioSet event classifier.
YOLO26-seg-semantic - Cityscapes semantic segmentation (1024x2048 input, 19 classes).
CABiNet - Aerial (UAV) semantic segmentation with a MobileNetV3 context branch and non-local attention (cabinet_mnv3l_uavid, UAVid, 8 classes).
PaddleOCR-v6 - The complete PP-OCRv6 text detection and recognition pipeline.
all-MiniLM-L6-v2 - First sentence-embedding model in the Model Zoo, 128-token input and 384-dim embeddings.
DINOv2 - Self-supervised Vision Transformer backbone with a linear ImageNet classifier head (dinov2_vits14_lc), ViT-S/14 at 224x224.
PIDNet - Real-time Cityscapes semantic segmentation with a three-branch (proportional / integral / derivative) pure-conv design.
SPAN - Swift Parameter-free Attention Network, the NTIRE 2024 Efficient Super-Resolution winner.
SPANF - The lighter NTIRE 2025 successor to SPAN.
PP-LiteSeg - Real-time Cityscapes semantic segmentation from PaddleSeg with an STDC encoder - pp_liteseg_t and pp_liteseg_b.
BiSeNetV2 - Real-time Cityscapes semantic segmentation with a bilateral detail / semantic design (bisenetv2, 3.36M parameters).
DDRNet - Real-time Cityscapes semantic segmentation with a deep dual-resolution design and a DAPPM context head.
ByteTrack - Multi-object pedestrian tracking by associating every detection box, low-score ones included.
ZipDepth - Lightweight zero-shot monocular depth estimation (zipdepth_256x832, zipdepth_384x384).
New variants of models already in the Model Zoo:
YOLO26-seg-semantic aerial - YOLO26 semantic-segmentation models fine-tuned on aerial imagery.
YOLOv11-seg - Instance-segmentation variants of the YOLOv11 family.
YOLO26 - yolo26s_768_448 and yolo26s_768_576, yolo26x_seg.
YOLO-World - yolo_world_v2m, the medium YOLO-World v2 open-vocabulary detector.
SCRFD - scrfd_2.5g_384_640 and scrfd_10g_384_640, 384x640 variants of the SCRFD face detectors.
Compatibility
Updated for Dataflow Compiler v5.5.0
Updated for HailoRT v5.5.0