# Which OpenRouter vision model would you use for building defect detection? (cost vs “good enough”) — retained record

> Retained reddit record from reddit:computervision.

Canonical: https://highsignal.app/data/records/a48b4521ac4ddb1f

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:computervision)  published 2026-11-11 05:41 ingested 2026-11-12 02:38

[original source ↗](https://www.reddit.com/r/computervision/comments/1xt9rld/which_openrouter_vision_model_would_you_use_for/)

## retained text



```
I’m working on flagging defects in residential building inspection photos taken on a phone: cracking, water staining and damp, mould, rust, timber rot, and roof damage. A human reviews every output. I’m aiming for about 80–90% recall on obvious defects with rough bounding boxes, not pixel-level segmentation. Cost per image matters because one job can mean 50–200 photos.

**What I’ve tried so far** (via OpenRouter, same prompt, 100 full-res photos from one real inspection report, scored against the engineer’s notes: 81 documented defects, all photo-level labels, no boxes yet):

* **Gemini 3.8 Flash:** 65/81 (80%) with the fewest false alarms of anything I tried (17 “not a defect” after engineer review). \~$0.34 per 100 images.
* **Qwen3-VL:** 64/81 (79%), but about twice the false alarms (35). Very cheap at \~$0.09 per 100 images.
* **Claude Sonnet 5.5:** 72/81 (89%), the best single model, but also the noisiest (56 false alarms, 52 “can’t tell from photo”). \~$1.44 per 100 images.
* **GPT-6 Luna:** 59/81 at \~$0.40 per 100. Not worth it here.
* **Ensemble:** Gemini Flash and Qwen miss *different* things (7 vs 6 unique catches), so taking the union gets 71/81 (88%) for \~$0.43 per 100. That’s close to Sonnet at under a third of the cost. Flags both models agree on are much cleaner than single-model flags, so I rank by agreement for the reviewer.
* **Prompting:** I compared a neutral prompt (“list clearly visible defects”) with a recall-focused one (“list anything plausibly a defect; misses are costly”). The recall prompt added +7 to +13 catches per model, at the cost of more noise. Common false positives were stains read as gaps or holes, textured render read as chipping, and patch repairs read as damage. Cracks were nearly solved (39–40/40). Small chips in floor tiles were the big gap (7–10/18).

**Specific questions:**

1. **Localisation.** So far I’ve only scored “does this photo contain defect type X”. Has anyone gotten reliably usable boxes from a general VLM, or is “VLM describes, small detector localises” the better pattern?
2. **Tiling.** I tried splitting 21 hard photos into crops. It recovered 10–11 of the misses per model and lost none, but extra flags roughly doubled or tripled. Cost went up \~3–8× (Gemini Flash $0.34 → $2.70 per 100). Has anyone found a good way to tile selectively, e.g. only when a first pass is uncertain or only on floor and close-up shots, rather than tiling everything?
3. **Datasets.** Are there public building or concrete defect datasets you’d recommend for a quick YOLO baseline to compare against? So far I’ve found \[SDNET2018, CODEBRIM\]. I’m especially interested in anything covering mould, damp, rust or timber rot, since my test set is mostly cracks, chips and tile damage.

Caveat: this is a small test (100 photos, one site, one engineer’s notes), so treat the numbers as directional. I’ll post benchmark results back here once I’ve compared more, including a second site.
```



document fetched 2026-11-12 00:24 · canonical https://reddit.com/r/computervision/comments/1xt9rld/which_openrouter_vision_model_would_you_use_for
