FluentVision

A fluent PHP 8.3+ API for YOLO object detection powered by Ultralytics YOLO26 and NanoDet-Plus.

Detect, segment, classify, and annotate images with an elegant chainable interface — same PHP result types regardless of which backend runs inference.

Quick Start

composer require b7s/fluentvision

use B7s\FluentVision\FluentVision;
use B7s\FluentVision\Enums\Provider;
use B7s\FluentVision\Enums\YoloModel;

$result = FluentVision::make()
    ->useUltralytics()
    ->model(YoloModel::YOLO26s)
    ->useCpu()
    ->confidence(0.5)
    ->media('photo.jpg')
    ->detect();

echo $result->getDetectionCount() . " objects found\n";

foreach ($result->detections as $detection) {
    echo sprintf("- %s (%.1f%%)\n", $detection->class, $detection->confidence * 100);
}

Installation

First, install Python dependencies and download models:

# Install the PHP package
composer require b7s/fluentvision

# Set up Python venv + packages + models
php ./vendor/bin/fluentvision install

# Or install only one provider
php ./vendor/bin/fluentvision install --provider=ultralytics
php ./vendor/bin/fluentvision install --provider=nanodet

# Download a specific model
php ./vendor/bin/fluentvision install --model=yolo26s.pt
php ./vendor/bin/fluentvision install --model=yoloe-26s-seg.pt
php ./vendor/bin/fluentvision install --model=nanodet-plus-m-416

The local user running your system must be the same user who will be using FluentVision.

Try something like this:

sudo -u www php ./vendor/bin/fluentvision install

This will create the environment in the user's "www" folder: /home/www/.fluentvision.

Check your environment:

php ./vendor/bin/fluentvision doctor

Providers

Provider	Backend	Best For
Ultralytics	YOLO26 (n/s/m/l/x), YOLOE-26 (s/m/l + PF)	Full-featured, multi-task, open-vocabulary detection
NanoDet	NanoDet-Plus (M/T/G)	Ultra-lightweight, edge devices, real-time

Both providers return identical PHP result types — switch backends without changing your code.

YOLOE-26 Open-Vocabulary Detection

YOLOE models support text prompts to detect anything you can describe — not just the 80 COCO classes:

use B7s\FluentVision\Enums\YoloModel;

$result = FluentVision::make()
    ->useUltralytics()
    ->model(YoloModel::YOLOE26s)
    ->useCpu()
    ->confidence(0.25) // default: 0.4
    ->prompts(['person', 'yellow hard hat'])
    ->media('factory.jpg')
    ->detect();

Variant	Suffix	Prompts	Best For
Text-prompted	yoloe-26*-seg.pt	->prompts([...]) required	Targeted attribute/concept detection
Prompt-free	yoloe-26*-seg-pf.pt	Not supported	Auto-detect without specifying prompts

Ultralytics Solutions

Run 12 built-in Ultralytics solutions (counting, heatmaps, speed estimation, etc.) with a single fluent call:

use B7s\FluentVision\Enums\UltralyticsSolution;

$result = FluentVision::make()
    ->media('highway.mp4')
    ->solution(UltralyticsSolution::Count, [
        'region' => '[[20,400],[1080,400],[1080,800],[20,800]]',
    ])
    ->withAnnotation(true)
    ->process();

echo "In: {$result->inCount}, Out: {$result->outCount}\n";

See docs/solutions.md for all 12 solutions and their parameters.

Detection Examples

Modern Workspace

$result = FluentVision::make()
    ->useUltralytics()
    ->model(YoloModel::YOLO26s)
    ->media('modern-workspace-with-laptop-coffee-plants.jpg')
    ->detect();

Person + Cup

$result = FluentVision::make()
    ->useUltralytics()
    ->model(YoloModel::YOLO26s)
    ->confidence(0.6)
    ->media('woman-cup-coffe.jpg')
    ->detect();

Street Scene with Segment

$result = FluentVision::make()
    ->useUltralytics()
    ->model(YoloModel::YOLOE26mPF) // Segment with Prompt free
    ->confidence(0.8)
    ->media('woman-bike-cars-trees-road-day.jpg')
    ->detect();
// 9 detections: person (90.6%), bicycle (91.2%), 7x car

Video Detection — Street Traffic

$result = FluentVision::make()
    ->useUltralytics()
    ->model(YoloModel::YOLO26s)
    ->confidence(0.4)
    ->everyNframes(10)
    ->media('/path-to-video/people-crossing-cars.mp4')
    ->detect();

echo $result->getFrameCount() . " frames\n";
echo $result->getTotalDetections() . " total detections\n";
// 48 frames, 342 detections — person, car, bus, traffic light

Detection Result Array

The detect() method returns an InferenceResult object. Call toArray() to get a plain array:

$result->toArray();

// [
//     'image_path' => '/path/to/photo.jpg',
//     'provider' => 'ultralytics',
//     'model' => 'yolo26s.pt',
//     'inference_time' => 0.1367,
//     'detection_count' => 2,
//     'detections' => [
//         [
//             'class' => 'person',
//             'confidence' => 0.910,
//             'box' => ['x1' => 198.0, 'y1' => 242.0, 'x2' => 675.0, 'y2' => 836.0],
//         ],
//         [
//             'class' => 'cup',
//             'confidence' => 0.646,
//             'box' => ['x1' => 638.0, 'y1' => 459.0, 'x2' => 844.0, 'y2' => 630.0],
//         ],
//     ],
// ]

Fluent API

use B7s\FluentVision\FluentVision;
use B7s\FluentVision\Enums\Device;
use B7s\FluentVision\Enums\Provider;
use B7s\FluentVision\Enums\YoloModel;
use B7s\FluentVision\Enums\NanodetModel;
use B7s\FluentVision\Enums\YoloTask;

FluentVision::make()
    ->provider(Provider::Ultralytics) // or ->useUltralytics() / ->useNanodet()
    ->model(YoloModel::YOLO26s) // or ->model('yolo26s.pt') or ->model('/path/to/custom.pt')
    ->media('/path/to/image.jpg') // media type auto-detected from extension
    // ->media('/path/to/clip.mp4') // video — auto-detected from extension
    ->useCpu() // or ->useGpu()
    ->confidence(0.5) // confidence threshold
    ->iou(0.45) // IoU threshold (NMS)
    ->imgsz(640) // inference image size
    ->maxDet(100) // max detections per image
    ->classes(['person', 'car']) // filter to specific classes
    ->prompts(['person wearing red', 'hard hat']) // YOLOE text prompts
    ->augment() // test-time augmentation
    ->half() // FP16 inference (GPU required)
    ->withDetections() // include detection data (default: true)
    ->withAnnotation() // include annotated image (default: false)
    ->process(); // returns ProcessResult with both
    // ->media('rtsp://camera/live')
    // ->streamConfig(function ($frame, $num, $result) { ... }, null, 100) // real-time stream (Ultralytics only)
    // ->process(); // returns StreamResult

Image Detection

$result = FluentVision::make()
    ->media('photo.jpg')
    ->detect();

Video Detection

$result = FluentVision::make()
    ->media('clip.mp4')  // .mp4 auto-detected as video
    ->everyNframes(10) // process every 10th frame - default: 5
    //->vidStride(10) // Alias for ->everyNframes(10)
    ->detect();

echo $result->getFrameCount() . " frames processed\n";
echo $result->getTotalDetections() . " total detections\n";

For realtime stream, check: Real-Time Streaming

Image Annotation

$result = FluentVision::make()
    ->media('photo.jpg')
    ->annotate();

echo "Annotated image saved to: " . $result->annotatedPath . "\n";

Detect + Annotate in One Call

Use process() to get both detections and an annotated image in a single inference run:

use B7s\FluentVision\Results\ProcessResult;

$result = FluentVision::make()
    ->media('photo.jpg')
    ->withDetections()       // include detection data (default: true)
    ->withAnnotation()       // include annotated image (default: false)
    ->process();             // returns ProcessResult

echo "Detections: " . $result->getDetectionCount() . "\n";
echo "Annotated: " . $result->getAnnotatedPath() . "\n";

// Access the individual results
$detections = $result->detections;  // InferenceResult|VideoInferenceResult
$annotation = $result->annotation;  // AnnotatedResult

process() runs inference once — more efficient than calling detect() and annotate() separately.

Working with Results

$result = FluentVision::make()->media('photo.jpg')->detect();

// Counts
$result->getDetectionCount();
$result->isEmpty();

// Filter detections
$persons = $result->filterByClass('person');
$highConf = $result->filterByMinConfidence(0.8);

// Unique classes
$classes = $result->getClasses();

// Individual detections
foreach ($result->detections as $d) {
    echo $d->class;           // "person"
    echo $d->confidence;      // 0.92
    echo $d->box->x1;         // 100.0
    echo $d->box->width();    // 150.5
    echo $d->box->area();     // 22650.25
}

// Serialize
$data = $result->toArray();

NanoDet Example

use B7s\FluentVision\Enums\NanodetModel;

$result = FluentVision::make()
    ->useNanodet()
    ->model(NanodetModel::PlusM416)
    ->media('photo.jpg')
    ->detect();

Custom Trained Models

Pass a path to your own trained model — provider is auto-inferred from the file extension:

// Ultralytics (.pt, .onnx, .engine, etc.) — auto-detected
$result = FluentVision::make()
    ->model('/path/to/my-trained-model.pt')
    ->media('photo.jpg')
    ->detect();

// NanoDet — use nanodetCustom() for config + checkpoint
$result = FluentVision::make()
    ->nanodetCustom('/path/config.yml', '/path/model.ckpt')
    ->media('photo.jpg')
    ->detect();

See Custom Models for full details on supported formats, model resolution, and provider auto-inference.

Configuration

Create fluentvision-config.php in your project root:

<?php

declare(strict_types=1);

return [
    'default_provider'       => 'ultralytics',
    'ultralytics_default_model' => 'yolo26s.pt',
    'nanodet_default_model'  => 'nanodet-plus-m-416',
    'default_device'         => 'cpu',
    'default_conf'           => 0.25,
    'default_iou'            => 0.7,
    'default_imgsz'          => 640,
    'python_path'            => null,           // auto-detect
    'python_venv_path'       => null,           // default: ~/.fluentvision/venv
    'model_dir'              => null,           // default: ~/.fluentvision/models
    'nanodet_repo_path'      => null,           // default: ~/.fluentvision/nanodet
    'timeout'                => 0,              // 0 = no timeout
    'verbose'                => false,
];

Or load from a custom path:

$vision = FluentVision::make('/path/to/my-config.php');

CLI Commands

# Environment check
vendor/bin/fluentvision doctor

# Install all dependencies
vendor/bin/fluentvision install

# Install specific provider
vendor/bin/fluentvision install --provider=ultralytics
vendor/bin/fluentvision install --provider=nanodet

# Download a model
vendor/bin/fluentvision install --model=yolo26m.pt
vendor/bin/fluentvision install --model=nanodet-plus-m-416

# Use custom config
vendor/bin/fluentvision doctor --config=/path/to/config.php
vendor/bin/fluentvision install --config=/path/to/config.php

Requirements

• PHP 8.3+
• Python 3.8+ with pip
• Ultralytics or NanoDet Python packages (installed via fluentvision install)

Documentation

• Installation Guide — step-by-step setup
• Configuration Reference — all config options
• Usage Guide — complete fluent API reference
• Providers — Ultralytics vs NanoDet details
• Custom Models — using your own trained models
• Result Objects — InferenceResult, ProcessResult, StreamResult, SolutionResult, DetectionResult, BoundingBox API
• Ultralytics Solutions — object counting, heatmaps, speed estimation, and 9 more built-in solutions
• Real-Time Streaming — RTSP, RTMP, webcam frame-by-frame detection
• CLI Commands — install, doctor, and options

License

MIT

---

Image by freepik