How naming field values are auto-detected (including AI vision)

Last updated June 10, 2026

If your ad naming convention uses field tokens — things like Dimensions, Asset Type, Style, or Hook — uplads can fill most of them in for you instead of making you pick a value on every ad. Some fields are detected for free with no AI involved; others use AI vision when you have a provider connected.

You'll find this in the Review step of the launch wizard, in the Ad naming card. Each ad unit lists its field tokens with a dropdown or text box, plus a wand button to detect values.

Mechanical detection: dimensions, asset type and video length (free, no AI key)

Three kinds of fields are resolved deterministically, with no model call and no cost — they work even if you've never set up an AI provider:

Dimensions / aspect ratio — parsed from the filename (patterns like _9x16_ or -1x1.) or from the ratio uplads already probed when you uploaded the creative.
Asset type — derived from the file's MIME type: videos map to your "Video" option, images to "Static" (or "Image" if that's what your option is called).
Video length — bucketed from the probed duration into whatever ranges your field defines, such as Under10s, 10-20s, or 1m30s+.

uplads recognizes these fields by their key and label, so a token named "Aspect Ratio", "Format", "Asset Type", or "Length" is picked up automatically. The detected value is snapped to one of your exact options (so 9:16 becomes your 9x16 option).

If you click Auto-fill fields and have no AI provider connected, this is everything that runs — you still get Dimensions, Length, and Asset Type filled in for free.

AI vision detection for custom fields like Style, Hook, Creator Age

Fields that can't be read off the file — Style, Hook, Creator Age, and similar subjective tokens — go to your connected AI provider's vision model. The button reads Detect all with AI once a provider is configured.

For each ad, the model sees:

The cleaned-up filename
Any facts already resolved (dimensions, duration, asset type)
The exact list of allowed options for each remaining field
The image itself (for static creatives; videos fall back to filename-based inference)

To enable this, connect a provider under AI provider and API keys. When both Anthropic and OpenAI are configured, uplads prefers Claude for vision. If a field is genuinely ambiguous, the model is instructed to leave it blank rather than guess.

You can detect the whole batch at once, or click the wand on a single ad to re-run just that one without spending tokens on the rest.

How vision output snaps to your exact field options

The model never injects free-form text into your ad names. For any field that has a defined options list, uplads matches the model's answer back to your exact option — so ugc or "user generated content" both resolve to your UGC option. If the model returns something that matches no option, that field is simply left empty for you to set. For free-text fields (no options), the value is trimmed to a short PascalCase label so it stays naming-safe.

Seeing which fields were auto-detected vs. manually filled

Detected values land directly in the same dropdowns and text boxes you'd edit by hand, so a filled field looks the same however it got there. After a detection run, a toast tells you how many ads were filled and how many were skipped (for example, "Detected 12 ads with Claude" or "Detected 9 — 3 skipped"). Anything still showing a blank value (—) wasn't confidently detected and is yours to pick. You can always override any value before launch — manual choices win over detected ones. Names update live as you change values, and the on-screen ad name is exactly what ships to Meta (see review and edit ad names).

Where vision detection cost appears in AI usage stats

Mechanical detection is always free. When a vision run actually calls the model, the token usage and estimated cost are logged so they show up in your AI usage and cost view, tagged separately from ad-copy generation. Each detection is roughly one vision call per ad, and runs are capped at 100 ads per batch with limited concurrency to stay under provider rate limits. If you've set a monthly AI spend cap, vision detection counts toward it — mechanical detection does not.