Run Vision Models
Anastig Vision Lab
Private visual AI workbench for images, video, live camera, and trained checkpoints.
Image
Run a model on one image and inspect overlays.
Video
Sample frames from a video, run a model, and inspect detections on a timeline.
Live Camera
Run a model continuously on a camera feed with FPS, latency, and overlays.
Batch
Run a model on many images and review the result grid.
Vision Lab runs in your browser or on your own device — no account needed. Checking your session for cloud projects…
Checking for a connected worker…
What you can run
Object detection
Find objects and draw boxes.
Open-vocabulary detection
Type what to find using a text prompt.
Segment anything
Create masks/outlines. No class names by default.
Classification
Label the whole image — pretrained classifiers or a model trained on your classes.
Embedding similarity
Compare image/crop similarity against an example.
Run my trained checkpoint
Use a model trained on my project classes.
Works across your domain
Open-vocabulary prompts find what you describe — pick your industry, or type your own.
Manufacturing / QC
Agriculture
Traffic / safety
Retail / logistics
Research / microscopy-like
What Vision Lab produces
Real capabilities — object detection, open-vocabulary prompts, segmentation, live camera, and sampled video. Diagrams are illustrative; run it on your own media for real results.
Object detection (closed vocabulary)
- Input
- Image
- Model
- RF-DETR / D-FINE / LibreYOLO
- Output
- Labeled boxes in the detector’s own vocabulary, one box per object (class-aware NMS).
Open-vocabulary prompt
- Input
- Image
- Model
- GroundingDINO + SAM / OWLv2
- Output
- Type what to find ("black car", "scratch") → labeled boxes/masks for exactly that.
Segment anything
- Input
- Image (box/point prompt)
- Model
- SAM 2.1 / MobileSAM
- Output
- Draw a box or click a point → a precise mask. No class until you save it.
Live camera
- Input
- Webcam stream
- Model
- Any detection / open-vocab model
- Output
- Real-time overlays with adaptive FPS, latency, and dropped-frame metrics. Frames never stored.
Video (sampled frames)
- Input
- Uploaded MP4
- Model
- Any detection model
- Output
- Sampled-frame analysis on a timeline; per-frame detection counts; codec-aware ingestion.
One pipeline: Studio and Vision Lab
Anastig Studio builds the dataset; Vision Lab runs the models. Annotate and export in Anastig Studio, then run detection, segmentation, and prompt-based models here — and turn results back into reviewable annotations.
1 · Studio
Upload images, annotate with AI assist, and export a clean, versioned dataset.
2 · Vision Lab
Run detectors, segmenters, and open-vocabulary prompts on your images, video, or webcam.
3 · Review
Inspect overlays, export an MP4 or labels, and feed accepted results back into annotation.
Private by design — what is live, what is next
Live today
- Image, video, and live-camera inference on a worker your org controls
- Detection, promptable segmentation, and open-vocabulary (text) prompts
- Overlay review and MP4 export — media transits the anastig.ai proxy to YOUR worker, never stored on Anastig servers
On the roadmap
- One-click training of a custom detector on your own worker
- Managed (Anastig-hosted) inference and deployment
- Turnkey private-worker onboarding for every organization
See the roadmap for live capability statuses.
Privacy: Vision Lab runs on a worker your organization controls. Uploaded media transits the anastig.ai proxy to that worker and is never stored or computed on Anastig servers; on-device webcam Fast Mode never uploads frames at all. Nothing is stored unless you explicitly save it. See our security page.