Will your ad actually perform?

Meet Treviso

Treviso is a frozen-feature model: every creative runs through a pretrained CLIP ViT-B/32 vision encoder whose weights are never touched, and the only thing we train is a small gradient boosted head (LightGBM) that maps the image embedding to an engagement rate. No fine-tuning, no training from scratch, nothing guessed from your brand or your budget. The pixels do the talking.

On held-out data it reaches a Spearman correlation of 0.71 with real engagement, and 0.78 pairwise accuracy when ranking sibling ads inside the same campaign. That pairwise number is the one we care about: same account, same budget, same audience, only the creative differs. Random guessing scores 0.5.

The results

Treviso was validated the hard way: held out by account with 5-fold cross validation, so every number comes from ads and advertisers the model had never seen. Cold start, no shortcuts, and it still ranks winning creatives correctly 78% of the time inside the same campaign.

The headline finding: budget and targeting context alone predicted nothing (0.53 pairwise, basically random), while the creative lifted ranking accuracy to 0.78. The creative, not the budget, drives engagement. That is exactly what Treviso reads, and exactly what this page scores.

What we learned

We also trained a sibling model, Bardolino, to predict conversion cost from the same creatives. It barely beat random (0.59 pairwise), which is a finding in itself: conversion depends on your offer, landing page and audience, not on the image. That is why this page scores engagement, the thing the creative actually controls, and why we publish the numbers instead of pretending to predict your sales.

• Scores your creative against the full corpus in seconds
• Explains what is working and what to fix, from the pixels alone
• Ranks your variants so you launch the strongest one
• Your image is used only for scoring and improving the model