AI Vision Assist¶
AI Vision Assist connects FindRomCover to a vision-capable language model that actually looks at candidate images and decides which one is the correct cover. Instead of comparing filenames, the model compares artwork — a much stronger signal when filenames are ambiguous, abbreviated, or wrong.
AI Assist is optional. Everything in FindRomCover works without it, and no AI provider is contacted unless you enable it.
What AI Assist can do¶
| Capability | Where | Description |
|---|---|---|
| AI Pick Best | Local Files and Google API tabs | Ranks the candidate images and highlights the best cover |
| Auto-run | Local Files tab | Starts the AI pick automatically after each local search |
| Auto-save | Local Files and Google API tabs | Saves the AI pick automatically when confidence exceeds the threshold |
| Verify before saving | Image folder watcher | Confirms that a newly saved image really matches the selected game before renaming it |
| Batch Fill | Missing covers list | Processes many missing covers in one run, with optional Google API fallback |
How a pick works¶
- The candidate images are downscaled to a maximum dimension (default 512 px) to keep token usage low.
- The top-ranked candidates (default 6) are sent to the model together with the game name.
- The model returns a JSON verdict: the chosen image, a confidence score, and a short explanation.
- The winner is highlighted in the results; depending on your settings it may be saved immediately.
- Verdicts are cached for 30 days, so repeating the same pick does not cost anything.
The filename matcher pre-filters local candidates first: only images whose similarity is at or above the Candidate similarity value (default 70%) are sent to the model.
Supported providers¶
| Provider | Type | Default model | API key |
|---|---|---|---|
| OpenRouter | Cloud | qwen/qwen3.7-flash |
Required |
| OpenAI | Cloud | gpt-4o-mini |
Required |
| Anthropic | Cloud | claude-sonnet-4-5 |
Required |
| Gemini | Cloud | gemini-2.5-flash |
Required |
| GLM | Cloud | glm-4.5v |
Required |
| Local | Local server | qwen2.5vl:7b |
Not required |
| Custom (OpenAI-compatible) | Any | User-defined | Optional |
| Custom (Anthropic-compatible) | Any | User-defined | Optional |
See Providers & Models for setup details and Recommended Models for cost comparisons.
Cost control¶
AI Assist is designed to keep API usage predictable:
- Downscaling — images are resized before upload.
- Candidate limit — only the top N candidates are sent.
- Candidate similarity — weak filename matches never reach the model.
- Verdict cache — identical picks are answered from a 30-day cache.
- Query history — covers already queried are not asked again (180-day memory).
- Confidence threshold — auto-save only happens above your threshold.
With the default OpenRouter model, a typical pick costs a fraction of a cent. See Recommended Models for estimates.
Safety¶
- AI settings and API keys are stored encrypted in the settings database and are never logged.
- AI request failures are logged as warnings and never trigger the automatic bug-report pipeline.
- The AI never deletes anything. It can only choose an image; saving and renaming follow your settings and confirmation rules.
- Verify before saving protects against wrong images being renamed automatically.
Getting started¶
- Open
Settings > AI Settings.... - Pick a provider, enter your API key, and click Test / Load Models.
- Select a model from the loaded list (vision-capable models are shown by default).
- Enable AI-assisted image selection.
- Optionally enable Run AI Pick automatically and Automatically save the AI pick.
- Use AI Pick Best or AI Fill Missing Covers... in the main window.