Find the model.
Build the idea.

Turn an aligned face crop into a 512-dim unit vector.

Higher-fidelity background removal with sharper edges than U-2-Net.

Motion-only tracker that holds IDs through short occlusions.

Zero-shot scoring between an image and a list of text prompts.

Add plausible color to a grayscale photo in one pass.

A modern CNN that closes most of the gap to ViT accuracy on CPU.

A 23 MB segmentation model for the 21 Pascal VOC classes.

ByteTrack motion plus an OSNet appearance gate for crowded scenes.

Estimate per-pixel depth in meters from a single photo.

Intel DPT-Large for relative depth from a single image, quantized to uint8.

The CPU-friendly default classifier, around 21 MB on disk.

Class-agnostic instance masks for every salient region, in one pass.

The heavy FastSAM backbone for higher recall on small or cluttered objects.

Dense 106-point face landmarks from an aligned crop.

68 face landmarks with a depth axis, from an aligned crop.

Estimate gender and a coarse age band from an aligned face crop.

The smallest credible ImageNet-1k classifier, around 10 MB on disk.

Click a point, get a mask. Distilled SAM for live interactive use.

Recover a sharp image from a single motion-blurred frame.

Score whether an uploaded image is safe or not safe for work.

Embed a pedestrian crop into a 512-dim vector for tracking across frames.

Three-stage OCR for live feeds: detect, classify orientation, recognize.

PP-OCRv5 with the 88 MB server detector for harder pages.

Parse a document page into 25 region types: titles, paragraphs, tables, figures.

Upscale a real-world photo by 4x with restored detail.

The well-understood ImageNet baseline, 102 MB on disk.

Meta's original SAM ViT-B for higher-quality masks when latency is fine.

InsightFace SCRFD with 5-point keypoints, the accuracy-pinned variant.

The live-feed SCRFD variant, 2.4 MB and 21 fps on CPU.

Tag every pixel with one of 150 ADE20K scene classes.

Google SigLIP2 with sigmoid scoring for independent yes-or-no prompts.

Cut the salient subject from a photo with a clean alpha mask.

The 2024 Ultralytics nano detector, the recommended live-feed default.

The Ultralytics flagship covering five tasks in one family.

One model, three outputs for driving: vehicles, drivable area, lane lines.

The 2023 Ultralytics nano detector, kept as a stable baseline.