Best Zero-Shot Object Detection Models
Compare the best 8 zero-shot object detection models and try 3 of them on your own image, free in the Roboflow Playground. 8 are open-weight, so you can self-host them for free under their licenses.
8 models · 8 open-weight · 3 free to try · prices synced Jul 30, 2026
Open-Source Zero-Shot Detection Models
8 models with downloadable weights you can self-host under their licenses (Custom, MIT, and AGPL 3.0). 3 run live in the Playground through hosted APIs, so self-hosting is optional.
SAM 3Meta | — | Custom | Nov 2025 | |
Florence-2Microsoft | 230M | MIT | Jun 2025 | |
YOLOETHU-MIG | 10M-50M | AGPL 3.0 | Mar 2025 | |
YOLO WorldTencent AI Lab | 13M | GPL v3 | Feb 2024 | |
 Grounded SAMIDEA Research | — | Apache 2.0 | Jan 2024 | |
SAM-CLIPApple | — | Custom | Oct 2023 | |
Grounding DINOIDEA Research | 172M-341M | Apache 2.0 | Mar 2023 | |
OWL-ViTGoogle | — | Apache 2.0 | May 2022 |
Getting Good Results from Zero-Shot Detection
Zero-shot detectors are only as good as their prompts and thresholds, and both behave differently than people expect. A few habits separate a frustrating demo from a reliable labeling pipeline.
These models ground words in appearance, so concrete visual phrases beat category jargon: "orange striped shirt" outperforms "non-compliant uniform", and "person wearing a hard hat" beats "PPE compliance". Attribute plus noun is the sweet spot; long relational descriptions ("the shirt next to the woman holding a coffee") degrade into phrase-grounding territory where precision drops. When a concept keeps failing, try synonyms: detectors often know "high-visibility vest" but not your industry term for it.
Score distributions vary wildly between prompts on the same model: 0.3 can be conservative for "person" and hopelessly noisy for "cracked weld". Treat the threshold as per-prompt configuration, not a global constant, and set it by inspecting a handful of your own images rather than trusting defaults.
Zero-shot output looks convincing, so measure it: label a small ground-truth set and check both misses and false positives before wiring the detector into anything automated. It shines for prototyping, ad hoc analysis, and pre-labeling datasets; for a fixed class list at production volume, use it to bootstrap labels and train a specialized detector, which will beat it on accuracy, speed, and cost. The open-vocabulary object detection task page covers choosing between the detectors themselves.
The bottom line: Concrete visual phrases, per-prompt thresholds, and a small ground-truth check before trusting the output. Prototype and label zero-shot; ship a trained detector once the classes settle.
Frequently Asked Questions About Zero-Shot Object Detection Models
Zero-shot detection models locate objects described in free text without being trained on those categories. Instead of a fixed class list, they align language and image regions (Grounding DINO, OWL-ViT), run prompt-conditioned one-stage detection at video speed (YOLO-World, YOLOE), or extend to segmentation everything they detect (Grounded SAM, SAM 3). You change what they find by changing the prompt, which makes them the fastest way to prototype a detector, pre-label a dataset, or handle categories that keep changing. On any fixed class list, a detector fine-tuned on your own data still beats them on accuracy, speed, and cost. This page lists 8 zero-shot object detection models, including 8 open-weight options you can self-host; 3 of them run live in the Playground so you can test them on your own images.
They describe the same capability from two angles: zero-shot emphasizes that no training examples of the target class were seen, open-vocabulary emphasizes that the class list is unbounded text. In practice the terms are used interchangeably for models like Grounding DINO and YOLO-World. We also maintain an open-vocabulary object detection task page listing models with that output wired up.
It depends on your task and constraints. For fixed categories in production, a model fine-tuned on your own data typically beats any general-purpose model. Compare the zero-shot object detection models on this page and try them on your own image to see which fits.
Yes. 8 of the 8 models here are open-weight (for example SAM 3, Florence-2, and YOLOE), free to self-host under their licenses (Custom, MIT, and AGPL 3.0).
Yes. You can run 3 of them in the Roboflow Playground for free. Upload an image and compare the models' output side by side, no setup required.
This page lists all 8 zero-shot object detection models in the Roboflow Playground catalog, all of them open-weight and free to self-host; 3 of them run live in the Roboflow Playground on your own images. Compare licenses, parameters, and release dates side by side, or open any model page for full details.