# Are foundation models becoming the default starting point for CV applications? — retained record

> Retained reddit record from reddit:computervision.

Canonical: https://highsignal.app/data/records/33af99df1b671481

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:computervision)  published 2026-11-03 16:03 ingested 2026-11-04 02:37

[original source ↗](https://www.reddit.com/r/computervision/comments/1xmwwi0/are_foundation_models_becoming_the_default/)

## retained text



```
A few years ago, building a CV application often meant collecting data, annotating it, and training or fine-tuning a task-specific model.

Now, for many tasks, it seems possible to take a different approach: use off-the-shelf foundation models (VLMs, open-vocabulary detectors, etc.) with some preprocessing and application-specific logic, without training anything.

For example, checking whether tables in a food court need cleaning, inspecting assembled products for missing components, or monitoring workplace safety compliance. With fixed cameras, you could define ROIs, process cropped images with pretrained models, and add some temporal logic where needed.

Of course, there are trade-offs in accuracy, latency, inference cost, and reliability. You'd still need labeled data for evaluation. But skipping the training pipeline could save a lot of development time, especially for smaller applications.

I'm curious how people approach this in practice. Are you deploying foundation models directly, only moving to task-specific training when needed? Or are you mostly using them to speed up annotation before training or fine-tuning a smaller model?
```



document fetched 2026-11-04 00:23 · canonical https://reddit.com/r/computervision/comments/1xmwwi0/are_foundation_models_becoming_the_default
