The New Standard in Open‑World Segmentation
Moondream Segmentation turns text prompts, points, or boxes into pixel-accurate SVG. State-of-the-art segmentation, available today in Moondream Cloud.
Beyond fixed classes
Traditional segmentation only recognizes what it was trained on. Need “hairline cracks” or “overripe fruit”? If it's not in the training set, you're stuck — curating data, labeling examples, retraining models.
Moondream understands language, not just labels. Describe any object, any attribute, any spatial relationship. Get pixel-accurate boundaries instantly. Your vocabulary is the only limit.
Pixel Perfect Prompting
Same API, endless applications. See Segmentation across different domains.
Use Case
Defect Tracking
Use Case
Agriculture
Use Case
Robotics
Use Case
Media
Top performance, single model
Moondream pairs top-tier grounding with precise segmentation in a single model, removing the need for any multi-model setup.
Moondream | SAM3 | Gemini Flash | SAM3 + Gemini 2.5 Pro | |
|---|---|---|---|---|
| RefCOCO-M | 86.9% | 42.3% | 73.9% | 86.3% |
| RefCOCO | 81.8% | 39.8% | 66.6% | 74.9% |
| RefCOCO+ | 74.7% | 27.0% | 60.9% | 66.9% |
| RefCOCOg | 76.4% | 34.8% | 68.1% | 73.3% |
| LVIS | 62.6% | 62.6% | -- | -- |
| Avg Time | 5.3s | 0.4s | 2.6s | 10.5s |
| Cost / 1K images | $0.40 | $2.20 | $9.00 | $40.00 |
How Segmentation works
Prompt, points, and boxes can be used alone or together to guide pixel‑accurate SVG polygons.
Inputs
Tell it what to find
- Describe it: “the person in blue”, “cracked tiles”, “ripe fruit” — plain English, no class labels
- Point to it: Click inside one or more objects to guide the model
- Box it in: Draw a rectangle to constrain the search area
Outputs
Get pixel-perfect vectors
- SVG polygons that trace exact object boundaries
- Render in-browser, export to design tools, or feed downstream pipelines
- Compact, editable, resolution-independent
FAQ
Common questions about Segmentation, pricing, and integration.