NanoVerse YOLO: Compact Edge Detectors

• NanoVerse YOLO is a research space that refines YOLO’s one-stage detection for compact, deployment-oriented models with strict size and compute constraints.
• It leverages heterogeneous design elements and machine-driven architecture search to balance high-resolution scaling, small-object sensitivity, and embedded efficiency.
• The approach demonstrates competitive mAP and reduced operational cost, validated by models like YOLO Nano, Octave-YOLO, and microYOLO across diverse deployment scenarios.

NanoVerse YOLO is not identified in the supplied literature as a formal model name. In the cited work, the term is most plausibly interpreted as denoting the compact and deployment-oriented branch of the YOLO family: architectures that preserve the one-shot, multi-scale detection paradigm while pushing model size, operation count, latency, or tiny-object sensitivity toward edge, mobile, microcontroller, or high-resolution real-time regimes. Within that reading, YOLO Nano is the clearest foundational reference point, and later systems such as LeYOLO, Octave-YOLO, microYOLO, SOD-YOLO, Hyper-YOLO, VajraV1-Nano, and YOLO11-4K can be understood as adjacent developments along the same design frontier (Wong et al., 2019, Hollard et al., 2024, Shin et al., 2024).

1. Scope and representative models

The supplied literature suggests that “NanoVerse YOLO” is best treated as an umbrella over several related, but technically distinct, goals. One line of work minimizes model size and arithmetic cost for embedded deployment; another retains YOLO-like throughput while improving scaling to 1080p or 4K inputs; a third adds finer-scale heads or richer fusion blocks to recover small-object performance that standard pyramids tend to lose. This suggests that compact YOLO research is not a single architecture family so much as a recurring set of constraints on backbone design, neck fusion, head resolution, and deployment operators (Hollard et al., 2024, Shin et al., 2024, Wang et al., 17 Jul 2025, Deutel et al., 2024, Feng et al., 2024, Hafeez et al., 18 Dec 2025, Makkar, 15 Dec 2025).

Representative model	Primary design target	Representative reported result
YOLO Nano	Embedded compact detection	4.0 MB, 4.57B operations, 69.1% mAP on VOC 2007
Octave-YOLO-N	High-resolution real-time detection	1.8M params, 5.3G FLOPs, 37.5 AP; 1.56 times faster than YOLOv8 at 1080
LeYOLO-Small@640	Low-FLOP YOLO-style detection	38.2 mAP, 4.5 GFLOP, 1.9M params
microYOLO	MCU-class single-shot detection	about 3.5 FPS, less than 800 KB Flash, less than 350 KB RAM
SOD-YOLO	Small-object UAV detection	0.351 mAP$_{50:95}$, 0.526 mAP$_{50}$, 22.6M params, 94.9G FLOPs
Hyper-YOLO-N	High-order cross-level representation	41.8 AP, 4.0M params, 11.4G FLOPs
VajraV1-Nano	High-accuracy nano-scale YOLO	44.3% COCO val AP, 3.78M params, 13.7B FLOPs, 1.1 ms/img
YOLO11-4K	4K panoramic small-object detection	1,377,444 params, 2.4 GFLOPs, 0.950 mAP@50, 28.3 ms

The common denominator across these systems is still recognizably YOLO: dense one-stage prediction, multi-scale features, and deployment-oriented engineering. The differences lie in what is treated as the dominant bottleneck: storage, FLOPs, memory traffic, post-processing, high-resolution redundancy, or small-object recall.

2. YOLO Nano as the foundational compact detector

YOLO Nano was introduced as a highly compact single-shot object detector for embedded, edge, and mobile deployment, explicitly motivated by the mismatch between modern detector accuracy and the storage, compute, and power budgets of practical low-power systems. The paper contrasts this setting with large YOLO-family and proposal-based baselines, noting that YOLOv3 is roughly 240 MB and requires more than 65 billion operations, while proposal-based detectors such as Faster R-CNN can fall to low single-digit frame rates on embedded processors. Its stated objective is unusually strict: produce a detector dramatically smaller and cheaper than Tiny YOLO baselines while maintaining, and in fact improving, detection accuracy (Wong et al., 2019).

The design method is framed as a human-machine collaborative strategy. The human-specified prototype, denoted $\varphi$, fixes the detector family: YOLO-style single-shot inference, shortcut-connected feature extraction, and three-scale prediction in a feature-pyramid-like arrangement. Machine-driven design exploration then determines the actual macroarchitecture and microarchitecture through generative synthesis, formalized as

$$\mathcal{G} = \max_{\mathcal{G}} \ \mathcal{U}(\mathcal{G}(s)) \quad \text{subject to} \quad 1_r(\mathcal{G}(s))=1,\ \forall s \in S.$$

The explicit constraints are central to the model’s identity: mAP $\ge 65\%$ on VOC 2007, computational cost $\le 5$ billion operations, and 8-bit weight precision. The resulting network is therefore not a manually shrunken YOLO baseline, but a searched detector under hard embedded constraints.

Architecturally, YOLO Nano is heterogeneous at both macroarchitecture and microarchitecture levels. The discovered network mixes residual projection-expansion-projection (PEP) modules, expansion-projection (EP) modules, lightweight fully-connected attention (FCA) modules, and standalone $3\times 3$ and $1\times 1$ convolutions. The most distinctive block is the residual PEP module, whose sequence is: $1\times1$ projection, $1\times1$ expansion, depthwise convolution, and a final $_{50}$0 projection, wrapped in a shortcut path. The efficiency rationale is the usual depthwise-separable reduction:

$_{50}$1

The FCA module adds selective channel reweighting through two fully connected layers that learn nonlinear interdependencies between channels and modulate them by channel-wise multiplication. At the whole-network level, YOLO Nano retains three-scale YOLO-style prediction of bounding boxes, objectness, and class outputs in one forward pass.

3. Reported efficiency, embedded behavior, and limitations

On VOC 2007 test with $_{50}$2 input, YOLO Nano is reported at 4.0 MB model size, 4.57B operations per inference, and 69.1% mAP. The comparison against compact YOLO baselines is the paper’s main empirical claim: Tiny YOLOv2 is reported at 60.5 MB, 57.1% mAP, and 6.97B operations; Tiny YOLOv3 at 33.4 MB, 58.4% mAP, and 5.52B operations. Relative to these baselines, YOLO Nano is more than $_{50}$3 and $_{50}$4 smaller, improves mAP by 12.0 and 10.7 percentage points, and reduces operations by about 34.4% and 17.2%, respectively (Wong et al., 2019).

The embedded measurements reinforce that the model was evaluated as a deployment system rather than as a purely offline benchmark architecture. On the NVIDIA Jetson AGX Xavier, YOLO Nano is reported at about 26.9 FPS and about 1.97 images/sec/W under a 15 W power budget, and about 48.2 FPS with about 1.61 images/sec/W under 30 W. These results expose a typical edge tradeoff: higher power budget materially improves throughput, but reduces throughput per watt.

The same paper also shows the main limitations of early compact-YOLO reporting. Evaluation is confined to PASCAL VOC rather than COCO; the supplied text does not include a full stage-by-stage architecture listing, exact head parameterization, or ablations isolating PEP, FCA, heterogeneity, or the search procedure. Embedded timing is reported only for YOLO Nano itself on Xavier, not as a normalized side-by-side Jetson comparison with Tiny YOLO baselines. The result is still historically important: it established that a searched, heterogeneous YOLO-like detector could be simultaneously smaller, cheaper, and more accurate than manually simplified Tiny YOLO baselines.

4. Post-YOLO Nano compactness: low-FLOP and high-resolution efficiency

Later work diversified the compact-YOLO problem. Octave-YOLO addresses the cost of processing high-resolution images in real time by using octave-style frequency decomposition. Its Cross Frequency Partial Network (CFPNet) splits features into high-frequency high-resolution and low-frequency low-resolution branches, routes expensive bottlenecks and self-attention only through the low-frequency branch, and then fuses the representation back. In the standard octave-convolution notation used by the paper,

$_{50}$5

The paper reports that Octave-YOLO-N reduces parameters from 3.2M to 1.8M and FLOPs from 8.7G to 5.3G relative to YOLOv8-N while slightly improving AP from 37.3 to 37.5, and at $_{50}$6 resolution it is 1.56 times faster than YOLOv8. At the same time, its latency results are measured on an RTX 3090 with ONNX Runtime rather than on an embedded SoC, so the “embedded” claim is supported indirectly through efficiency and scaling trends rather than by Jetson-class benchmarking (Shin et al., 2024).

LeYOLO takes a different position. It is explicitly FLOP-centric and aims to bridge YOLO-style detectors and SSDLite-like efficiency. Its backbone uses modified inverted bottlenecks with optional first pointwise expansion, the neck is FPANet or LeNeck, and the head is a decoupled Network-in-Network design called DNiN. Reported COCO results include LeYOLO-Small@640 at 38.2 mAP with 4.5 GFLOP and 1.9M parameters, and LeYOLO-Medium@640 at 39.3 mAP with 5.8 GFLOP. On Jetson TX2 4GB using TensorRT, LeYOLO-Small@640 is reported at 24.0 QPS and LeYOLO-Medium@640 at 19.89 QPS. The paper is unusually explicit that the model is not always the fastest because of “speed imperfections due to the (deliberate) lack of parallelizable architecture,” which makes it a good example of the distinction between FLOP efficiency and accelerator-optimal latency (Hollard et al., 2024).

A later nano-scale comparison point is VajraV1-Nano. It does not minimize compute as aggressively as YOLO Nano, LeYOLO, or microYOLO; instead it spends more width in its main block and recovers efficiency through ADown, parameter-efficient deep blocks, re-parameterized RepVGG structure, and low-resolution partial attention. The reported result is 44.3% COCO val AP at 3.78M parameters, 13.7B FLOPs, and 1.1 ms/img, outperforming YOLOv12-N and YOLOv13-N in AP while remaining within the same high-end real-time GPU regime. This suggests a later reinterpretation of “nano” as a model-scale label rather than a strict ultra-low-resource constraint (Makkar, 15 Dec 2025).

5. Small-object, high-resolution, and feature-fusion branches

Another major branch of the NanoVerse landscape prioritizes tiny-object sensitivity rather than minimum arithmetic cost. SOD-YOLO is built on YOLOv8-m for UAV imagery and modifies three places in the pipeline: Adaptive Scale Fusion in the neck, a new high-resolution P2 detection layer, and Soft-NMS in post-processing. On VisDrone2019-DET-val, the paper reports improvement from YOLOv8-m’s $_{50}$7 and $_{50}$8 to SOD-YOLO’s 0.351 and 0.526, described as a 36.1% increase in $_{50}$9 and a 20.6% increase in $\varphi$0. The tradeoff is clear: parameters fall from 25.8M to 22.6M, but FLOPs rise from 78.7G to 94.9G, and the paper provides no FPS or latency numbers. SOD-YOLO is therefore a small-object-optimized YOLO, not a nano detector in the strict embedded sense (Wang et al., 17 Jul 2025).

YOLO11-4K addresses a related problem in 4K panoramic imagery. Its best reported model combines a hybrid GhostConv/Conv backbone, a P2 detection branch, and kernel–stride tuning. The paper reports 1,377,444 parameters, 2.4 GFLOPs, $\varphi$1 mAP@50, and 28.3 ms inference per frame, compared with YOLO11 at $\varphi$2 mAP@50 and 112.3 ms. It also reports detection of objects as small as $\varphi$3 px. The architecture is thus compact in parameter count, but the deployment evidence is GPU-based rather than edge-device-based, and the paper contains minor inconsistencies between 28.3 ms and 21.4 ms latency statements and between $\varphi$4 and $\varphi$5 descriptions of the 4K setting (Hafeez et al., 18 Dec 2025).

FA-YOLO and Hyper-YOLO illustrate more accuracy-oriented extensions that remain relevant to the same conceptual space. FA-YOLO inserts a Fine-grained Multi-scale Dynamic Selection module and an Adaptive Gated Multi-branch Focus Fusion module into YOLOv9’s RepNCSPELAN4 blocks, improving VOC 2007 mAP from 65.1 to 66.1 and raising $\varphi$6, $\varphi$7, and $\varphi$8, but at 30.7M parameters and 101.9 GFLOPs it is not nano-scale (Huo et al., 2024). Hyper-YOLO replaces conventional neck fusion with hypergraph computation through HyperC2Net and strengthens the backbone with MANet; Hyper-YOLO-N is reported at 41.8 AP, 4.0M parameters, and 11.4G FLOPs. Its limitation is deployment overhead: the paper notes that TensorRT does not fully optimize the torch.cdist() step used in hypergraph construction, so the gain in AP comes with extra engineering complexity and some latency cost (Feng et al., 2024).

6. Extreme deployment, applied pipelines, and search-driven evolution

At the extreme low-resource end, microYOLO tests whether a YOLO-style single-shot detector can fit on a Cortex-M7 microcontroller. The reported deployment on the OpenMV H7 R2 achieves about 3.5 FPS on $\varphi$9 RGB images while using less than 800 KB Flash and less than 350 KB RAM. The model is much closer to a YOLOv1-style grid predictor than to modern multi-scale YOLOs, and its reported $$\mathcal{G} = \max_{\mathcal{G}} \ \mathcal{U}(\mathcal{G}(s)) \quad \text{subject to} \quad 1_r(\mathcal{G}(s))=1,\ \forall s \in S.$$0 varies sharply by task: 56.4% on the fridge dataset, 27.7% on humans, and 12.3% on vehicles. This makes microYOLO an important lower bound for the NanoVerse idea: single-shot detection remains possible under MCU-class budgets, but only with aggressive simplification, tiny inputs, and strong sensitivity to scene complexity (Deutel et al., 2024).

Application-oriented papers show a different sense in which compact YOLO work evolves. A surveillance study based on standard YOLOv3 and transfer learning from MS COCO uses $$\mathcal{G} = \max_{\mathcal{G}} \ \mathcal{U}(\mathcal{G}(s)) \quad \text{subject to} \quad 1_r(\mathcal{G}(s))=1,\ \forall s \in S.$$1 input, three-scale prediction, and confidence and NMS thresholds around 0.6, but its strongest numerical claims are incompletely supported: the paper reports an “improved mAP of 61.5” and a “99.71%” accuracy figure while also listing lower video-level accuracies between 0.701 and 0.892, so it is better read as a deployment recipe than as a rigorously benchmarked detector contribution (2209.12447). A face-mask detection study likewise uses YOLOv3 as the main detector, but adds a weighted loss,

$$\mathcal{G} = \max_{\mathcal{G}} \ \mathcal{U}(\mathcal{G}(s)) \quad \text{subject to} \quad 1_r(\mathcal{G}(s))=1,\ \forall s \in S.$$2

to compensate for class imbalance; its weighted-loss YOLOv3 reaches precision 0.919, F1 0.731, and 95.78 ms per 100 images, versus Faster R-CNN at precision 0.932, F1 0.721, and 175.2 ms per 100 images. These papers do not define NanoVerse architectures, but they show how lightweight YOLO research also includes training and deployment adjustments under data and hardware constraints (Liu et al., 2021).

The most recent direction in the supplied literature is search and edge-first simplification. YOLO-NAS-Bench is not itself a detector but a surrogate benchmark for YOLO architecture search; it builds a 1,500-architecture pool on COCO-mini and reports that self-evolving surrogate training improves ensemble predictor performance from $$\mathcal{G} = \max_{\mathcal{G}} \ \mathcal{U}(\mathcal{G}(s)) \quad \text{subject to} \quad 1_r(\mathcal{G}(s))=1,\ \forall s \in S.$$3 and sKT $$\mathcal{G} = \max_{\mathcal{G}} \ \mathcal{U}(\mathcal{G}(s)) \quad \text{subject to} \quad 1_r(\mathcal{G}(s))=1,\ \forall s \in S.$$4 to $$\mathcal{G} = \max_{\mathcal{G}} \ \mathcal{U}(\mathcal{G}(s)) \quad \text{subject to} \quad 1_r(\mathcal{G}(s))=1,\ \forall s \in S.$$5 and sKT $$\mathcal{G} = \max_{\mathcal{G}} \ \mathcal{U}(\mathcal{G}(s)) \quad \text{subject to} \quad 1_r(\mathcal{G}(s))=1,\ \forall s \in S.$$6, with discovered architectures surpassing official YOLOv8–YOLO12 baselines at comparable latency on COCO-mini. This suggests that the compact-YOLO frontier is increasingly search-driven rather than purely hand-designed (Li et al., 10 Mar 2026). YOLO26, by contrast, is framed as an edge-first detector that removes DFL and NMS while adding ProgLoss, STAL, and MuSGD, and claims up to 43% faster CPU inference for the nano model; however, the supplied text does not provide direct Jetson Nano or Orin benchmark tables despite presenting Jetson deployment as a central theme. That gap is significant because it shows how “nano” rhetoric can outpace reproducible embedded evidence (Sapkota et al., 29 Sep 2025).

Taken together, the literature supports a precise interpretation of NanoVerse YOLO: not a single standardized model, but a research space within the YOLO lineage defined by tension among compactness, FLOPs, memory movement, high-resolution scaling, and small-object retention. YOLO Nano remains the canonical origin of that space in the supplied sources, because it made the embedded objective explicit and paired it with a heterogeneous machine-designed architecture. Later models broaden the same agenda in different directions—sub-gigaflop efficiency, microcontroller feasibility, 1080p and 4K scaling, dynamic feature fusion, high-order reasoning, or high-accuracy nano-scale competition—without removing the central question that YOLO Nano posed: how far a one-stage detector can be compressed or specialized before it ceases to remain both accurate and practically deployable.