# One photo, one question, three models: one returned a bounding box, one wrote prose for 2.5 seconds, and one sent back HTTP 400 — retained record

> Retained reddit record from reddit:SaaS.

Canonical: https://highsignal.app/data/records/389e441eced91ab1

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:SaaS)  published 2026-10-29 15:28 ingested 2026-10-30 02:38

[original source ↗](https://www.reddit.com/r/SaaS/comments/1xin3nr/one_photo_one_question_three_models_one_returned/)

## retained text



```
The question was plain enough to compare cleanly, a photo of a crushed front end and a single instruction, find the damage on the car. Bivia answered with coordinates, box \[187, 50, 1000, 919\] on a 0 to 1000 scale, so the result arrives ready to hand straight to whatever step runs next.

The general model spent 2.5 seconds composing a description of the photograph, which leaves prose to parse for a location that four integers already carried.

Jev returned HTTP 400, invalid image dimensions, since it caps input at 1568 px per side and leaves resizing to the caller, so the original file bounced ahead of any inference.

The clip runs all three in parallel across six photos, a white sedan in Japan, van corner damage in Oban, a wreck on a bridge, a smashed front in Esbjerg and a rear end collision in Moscow, so the same question lands on each one in turn.
```



document fetched 2026-10-30 00:29 · canonical https://reddit.com/r/SaaS/comments/1xin3nr/one_photo_one_question_three_models_one_returned
