builditwithgk/adresyze-ad-layouts

AdResyze Ad Layout Annotations (v2)

Layout annotations for 302 Indian brand advertisements: which region of an ad is the logo, the headline, the call to action, the product, the price, the background — and where each one sits.

Built for AdResyze, which uses these layouts to reflow a single creative into every platform aspect ratio without squashing the logo or cropping the call to action.

Try it: builditwithgk--adresyze-ui.modal.run  ·  Code: github.com/builditwithgk/adresyze

Annotations only. The advertisements themselves are brand-owned and are not redistributed.

Format

One JSON file per ad.

{
  "image_file": "121101721_337678557303047_605074869721584267_n.jpg",
  "source_width": 600,
  "source_height": 400,
  "aspect_ratio": "other",
  "aspect_ratio_value": 1.5,
  "platform_guess": "other",
  "dominant_colors": ["#ffffff", "#000000"],
  "elements": [
    {
      "type": "product",
      "bbox": [0.25, 0.3878, 0.6599, 0.7398],
      "priority": 1,
      "must_preserve": true
    }
  ],
  "generator": {"model": "Qwen/Qwen2.5-VL-7B-Instruct", "adapter": null}
}

field	meaning
bbox	[x1, y1, x2, y2], normalized 0..1 against the source image, origin top-left
type	logo, headline, cta, product, price, background, other
priority	1 critical, 2 important, 3 optional
must_preserve	whether the element should survive a crop
aspect_ratio	nearest standard placement, or other when the ad is not within 3% of one
aspect_ratio_value	the true width / height — this is the authoritative value

Because boxes are normalized against source_width × source_height, they can be placed on the image at any resolution.

Contents

302 ads, 1493 elements.

type	count		ratio	count
headline	358		1:1	198
background	306		other	33
other	210		4:5	30
product	202		9:16	27
cta	201		1.91:1	11
logo	176		16:9	3
price	54

How it was made

Ads collected from the Meta Ad Library (India) on discount and offer search terms. Annotated with stock Qwen/Qwen2.5-VL-7B-Instruct, greedy decoding, using a prompt that states the element vocabulary and value formats explicitly. Every record was then validated against a schema; 302 of 302 passed.

Note the generator block on every record: v2 was not produced by the adresyze-lora adapter. Measured over 25 ads with an identical prompt, the adapter lost or tied the stock model on every conformance metric — 96% vs 100% parse rate, 3.9 vs 4.5 elements found per ad — so it is not in the pipeline. The adapter remains published as provenance.

Changes from v1

• Boxes are grounded. v1 stored coordinates in the model's internal patch grid and recorded no source dimensions, so they could not be mapped onto the image. v2 stores normalized coordinates plus source_width / source_height.
• priority and must_preserve vary. In v1 they were effectively constant (1121 of 1126 elements had priority: 1) because nothing defined them. Defining them in the prompt made them informative.
• aspect_ratio is computed from the image, not guessed by the model.
• 302 records instead of 289, and 1493 elements instead of 1126.

Limitations

• Machine-generated, not human-verified. These are model outputs validated for schema conformance, not ground truth. No hand-labelled accuracy figure is published.
• ~3% of raw boxes extended past the model's own coordinate grid and were clamped.
• other is 14% of elements — regions the model declined to classify.
• Indian market, discount and offer creatives. Do not assume it generalizes.

Licence

CC BY 4.0. Attribution requested.