Layer
retrieval

Can image models separate the parts of a finished design?

Flattened Kiln and Hand cover with a studio interior, vase illustration and lettering — Flattened cover

Complete background layer saved before flattening — 01Background

Complete illustration layer saved before flattening — 02Illustration

Complete typography layer saved before flattening — 03Typography

Dataset · privateExplore on Harbor Hub Source · privateView task code on GitHub

00 / Task definition

From pixels to a layer

01 / 03 · Build

Complete illustration layer — Illustration

1 · Build

Save the design parts

We make a finished design from separate artwork, background and text layers. Titles and details can be separate requests.

The single flattened Kiln and Hand cover given to each model — Input

Nano Banana 2

Gemini 3 Pro

Sunburst

2 · Extract

Ask for one part

Three image models see the same flattened design and a short request for one named layer.

Saved complete vase illustration layer — Saved layer

Recovered vase with green overlap and red missing or extra shapes — Recovered layer

● Matching shape● Missing or extra

3 · Score

Compare with the saved part

We line up each output with the original layer and measure its pixels and shape.

01 / Model results

Model results

Separate useful parts of finished designs: artwork, backgrounds, titles and details.

Nano Banana 2

25.2%

Mean match within saved layers

Full canvas: 72.3%

Gemini 3 Pro Image

28.2%

Mean match within saved layers

Full canvas: 73.6%

GPT Image 2.5 Sunburst

11.3%

Mean match within saved layers

Full canvas: 67.5%

50 distinct designs across five categories, with ten designs in each and 90 scored layer requests in total. Each model saw a flattened 1024 × 1024 design and a one-sentence request for one part. We average the layers of each design first, then give each design equal weight. These means count matches within the saved layer, so empty transparency does not raise them. This is a selected set of practical tasks, not a random performance estimate.

02 / Scoring

How we score

We compare each extraction with the layer saved when the design was made. Every image is 1024 × 1024 pixels. A pixel matches when its alpha differs by at most 4 and its weighted RGB distance is at most 10.

Full canvas

M=100\,\frac{\sum_{p\in P}\mathbf{1}[\operatorname{match}(p)]}{|P|}

Count close matches over every pixel. Shown for continuity, but empty transparency can raise it.

Saved layer pixels

T=100\,\frac{\sum_{p\in A}\mathbf{1}[\operatorname{match}(p)]}{|A|}

Count matches only on nontransparent pixels in the saved layer. This is the displayed mean.

Isolation

I=\begin{cases}0,&|B|=|P|\text{ and }|A|<|P|\\100,&|A\cup B|=0\\100\,\frac{|A\cap B|}{|A\cup B|},&\text{otherwise}\end{cases}

Compare the visible areas of both layers. A full-canvas output gets 0 when the saved layer has transparency. This score does not check color.

In these formulas: P is every pixel. A contains pixels with saved-layer alpha of at least 128, including parts covered in the finished design. B contains output pixels with alpha of at least 128. If A and B are both empty, isolation is 100%. Exact background scores can penalize plausible fills behind a subject.

A design row averages its scored layers. A category bar averages ten design means, and the headline averages all 50 design means. Each design has equal weight. These averages use saved layer pixels (T), not full-canvas matches (M).

Exact pixel match rule

For each pixel, subtract the two 8-bit alpha values. For each RGB channel, first multiply its 8-bit value by alpha divided by 255, then subtract the two results.

|\Delta_\alpha|\le 4\quad\land\quad \sqrt{3\Delta_R^2+4\Delta_G^2+2\Delta_B^2}\le 10

A plausible image can still get a low exact-match score when its texture, placement, or lettering differs.

03 / Designs and layers

Explore the designs

Photo edits

Photos, titles, details and individually selectable design objects. 10 designs · 17 layers.

Nano Banana 228.7%

Gemini 3 Pro Image22.5%

GPT Image 2.5 Sunburst10.3%

Swipe horizontally to see all model scores →

Nano Banana 2	Gemini 3 Pro Image	GPT Image 2.5 Sunburst
0.0%	0.0%	0.0%
0.0%	0.1%	0.0%
2.2%	13.3%	2.1%
2.2%	28.3%	0.8%
34.7%	0.0%	0.0%
16.7%	18.4%	2.7%
39.6%	2.0%	0.7%
53.7%	13.6%	11.7%
60.4%	62.4%	2.3%
77.7%	87.4%	83.2%

04 / Cost

Cost to repeat

List-rate estimate from the recorded token usage for one successful run of the 90 displayed cases.

Model	90 image edits
Nano Banana 2	≈ $6.17
Gemini 3 Pro Image	≈ $12.44
GPT Image 2.5 Sunburst	≈ $5.50
All three models	≈ $24.11

Standard global rates as of October 9, 2026. Includes recorded input, image output, and reasoning tokens where applicable. Excludes retries, failed requests, taxes, and account discounts; this is not an invoice. Google pricing · OpenAI pricing