Runs as a custom pluginDetectionVerified 2026-11-16
Grounding DINO Tiny
IDEA-Research/grounding-dino-tiny
Grounding DINO Tiny is an open-vocabulary object detection model (Apache-2.0). Can run on Mixpeek as a custom plugin.
Mixpeek does not run it today. It can run as a custom plugin: the weights load without remote code.
No demo for Grounding DINO Tiny yet
A video, image, slideshow or playground goes here once we record one on our own data.
What it is good and bad at
- Best at
- Open-vocabulary boxes from a text prompt, with a commercial license; cropping a query photo to its logos lifted open-model logo search from R@1 .833 to .956 on the ip-check benchmark (SigLIP base + CLIP L/14, whole photo plus three crops).
- Worst at
- Speed (a Swin-T backbone and a text encoder per image, slower than YOLO-class detectors); a prompt like "a logo." also boxes text and packaging, so crops are scored beside the whole image rather than replacing it.
- Task
- Detection
- Input
- Image, Text
- Output
- Boxes, Labels, Scores
- Use cases
- Brand detection
Templates that use it
No template uses this model yet. Browse templates.
Input and output
Input
Image
- Resolution
- shortest side 800, longest 1333 (processor default)
Prompts
lower-case phrases ending in a period, such as "a logo. a brand logo."
Output
Boxes
xyxy pixels
Labels
the prompt phrase each box matched
Scores
0..1, box and text thresholds
How Mixpeek wires it
- Extractors
- None today
- Feature URIs
- None today
- Vector indexes
- None
- Retriever stages
- None
- Combines with
- query photo -> Grounding DINO "a logo." boxes -> padded crops -> SigLIP / CLIP vectors -> best score per drawing beside the whole-photo score
How it runs
- Status
- Runs as a custom plugin No Serve app or built-in extractor loads it. It runs as a custom plugin through transformers AutoModelForZeroShotObjectDetection with safetensors weights.
- Serve apps
- None
- Runtime
- transformers AutoModelForZeroShotObjectDetection + AutoProcessor, post_process_grounded_object_detection
- Weights
- safetensors, no remote code
- Parameters
- 172 M
- Size on disk
- 690 MB
- Preprocessing
- Processor resize and normalize; the text prompt tokenized with the BERT tokenizer.
- Postprocessing
- Box threshold 0.2 and text threshold 0.2 (as measured), top three boxes, each padded 15% before cropping.
- Serving
- both
- Hardware
- Runs on CPU (slower without a GPU)
Scale to zero
- Minimum replicas
- Not recorded
- Maximum replicas
- Not recorded
- Idle timeout
- Not recorded
- Memory
- Not recorded
- CPUs
- Not recorded
- GPUs
- Not recorded
- Cold start, node warm
- Not recorded
- Cold start, new node
- Not recorded
Measured cost
Unmeasured
No committed throughput or cost measurement.
Our evaluations
| Dataset and metric | Score | Baseline | Date |
|---|---|---|---|
ip-check logo benchmark, 90 photos with a known owner against 5,278 USPTO drawings and 3,003 Wikidata logos recall@1, SigLIP base + CLIP L/14 z-summed, whole photo plus up to three Grounding DINO crops (k=1), held out | 0.956 | the same two models on the whole photo only (SigLIP base alone .833): 0.844 | 2026-11-16 |
Questions
- What is Grounding DINO Tiny best at?
- Open-vocabulary boxes from a text prompt with a commercial license; cropping a query photo to its logos lifted open-model logo search from R@1 .833 to .956 on the ip-check benchmark.
- What are the limits of Grounding DINO Tiny?
- It is slower than YOLO-class detectors, and a prompt like "a logo." also boxes text and packaging, so crops are scored beside the whole image.
- Can Grounding DINO Tiny be used commercially?
- Yes. Apache-2.0 weights and code; no use restrictions.
- How does Grounding DINO Tiny run on Mixpeek?
- As a custom plugin through transformers; no Serve app or built-in extractor loads it yet.
- What does Grounding DINO Tiny output?
- Boxes in pixels, the prompt phrase each box matched, and a score from 0 to 1.