computer-vision on S Anand

Gemini copies images almost perfectly

Thu, 22 Jan 2026 09:33:02 +0530

Summary: Nano Banana Pro is much better than recent models at copying images without errors. That lets us do a few useful things, like:

Pre-process images for OCR, improving text recognition by cleaning up artifacts while preserving text shapes exactly.
Convert textbook raster diagrams into clean vector-like images that vectorizers can process easily.
Create in-betweens for cartoon animations
Copy torn, stained 1950s survey maps into pristine, high-contrast replicas with boundary lines preserved pixel-perfectly.
Redraw sewage map blueprints or refinery blueprints into clean schematics, separating the “pipes” from the “background noise”.
… and more!

GPT Image 1.5 has a good reputation for drawing exactly what you tell it to.

But right now, nothing beats Gemini 3 Pro Image Preview (Nano Banana Pro) on copying images exactly.

To check this, Nitin asked a bunch of models to:

Recreate the provided image EXACTLY:

- Same resolution and aspect ratio
- Same colors, lighting, composition
- Same shapes and positions
- No stylization, no creativity, no cleanup
- Do not add/remove text, objects, or artifacts
- Output should be the closest possible pixel-level match

… and passed it different kinds of images.

Here’s their structural similarity index (SSIM):

Image Type	gemini-3-pro-image-preview	gemini-2.5-flash-image	gpt-image-1.5	chatgpt-image-latest	gpt-image-1	gpt-image-1-mini
Document	89.4%	33.6%	14.0%	10.0%	7.9%	8.3%
Diagram	97.9%	32.2%	32.4%	24.4%	26.9%	31.6%
Stereogram	98.5%	37.8%	41.1%	34.1%	36.1%	39.6%
Scatterplot	89.1%	85.9%	46.3%	41.9%	36.0%	38.5%
Surreal painting	99.7%	66.7%	77.2%	76.6%	72.2%	51.4%
Tuvalu map	96.2%	99.9%	72.4%	61.8%	64.5%	65.4%
Cartoon	99.8%	81.3%	72.3%	71.9%	72.4%	72.3%
Human photo	99.3%	76.1%	82.9%	82.2%	74.0%	71.8%
Human portrait	99.9%	84.7%	93.4%	89.8%	89.5%	77.0%
Landscape	99.8%	95.7%	93.1%	92.4%	90.3%	92.0%
Overall Average	97.0%	69.4%	62.5%	58.5%	57.0%	54.8%

For example, take this portrait:

Gemini 3 Pro is nearly pixel perfect. (Red areas show differences. You can see that most areas are barely different.) Whereas GPT Image 1.5 makes more changes:

The nose is smaller.
The mouth is wider.
The lips are thinner.
The eyes are smaller.
The moustache is smaller.
The goatee is neater.
The face is narrower.
The skin is yellower.
… and so on.

But even the best models aren’t perfect when copying diagrams and documents.

The text is slightly displaced. Background writing is removed. Gemini 3 Pro actually sharpens the text. But even here, interestingly, almost every letter is copied correctly - the differences are small.

What does this mean?

Clearly, Gemini 3 Pro Image Preview outperforms the others by a huge margin. Winner in 10 categories, with a 97% average. The runner up (Gemini 2.5 Flash Image) wins barely nudges it out on the Tuvalu map, and has a 69% average.

Also, the models are clearly better at natural images (human photo, portrait, landscape) than with symbolic images (cartoon, diagram, scatterplot, document).

Among the OpenAI models, GPT Image 1.5 is far better than the others, but lags far behind Gemini 3 Pro. The gap is larger for symbolic images - so definitely prefer the Gemini models to edit diagrams, documents, charts etc.

Also, Gemini 3 Pro Image Preview seems quite different from Gemini 2.5 Flash (20% correlation)- almost a different model, whereas GPT Image 1.5 seems fairly similar to GPT Image 1 and GPT Image 1 Mini (97% correlation) - like incremental improvements to the same model.

You can try out your own images at https://nitin399-maker.github.io/imagediffer/ or run the code yourself.

LLMs still do not locate bounding boxes well

Fri, 01 Nov 2024 16:20:03 +0000

I sent an image to over a dozen LLMs that support vision, asking them:

Detect objects in this 1280x720 px image and return their color and bounding boxes in pixels. Respond as a JSON object: {[label]: [color, x1, y1, x2, y2], …}

None of the models did a good-enough job. It looks like we have some time to go before LLMs become good at bounding boxes.

I've given them a subjective rating on a 1-5 scale below.

Model	Positions	Sizes
gemini-1.5-flash-001	🟢🟢🟢🔴🔴	🟢🟢🟢🟢🔴
gemini-1.5-flash-8b	🟢🟢🟢🔴🔴	🟢🟢🟢🔴🔴
gemini-1.5-flash-002	🟢🟢🔴🔴🔴	🟢🟢🟢🔴🔴
gemini-1.5-pro-002	🟢🟢🟢🔴🔴	🟢🟢🟢🟢🔴
gpt-4o-mini	🟢🔴🔴🔴🔴	🟢🟢🔴🔴🔴
gpt-4o	🟢🟢🟢🟢🔴	🟢🟢🟢🟢🔴
chatgpt-4o-latest	🟢🟢🟢🟢🔴	🟢🟢🟢🟢🔴
claude-3-haiku-20240307	🟢🔴🔴🔴🔴	🟢🟢🔴🔴🔴
claude-3=5-sonnet-20241022	🟢🟢🟢🔴🔴	🟢🟢🟢🔴🔴
llama-3.2-11b-vision-preview	🔴🔴🔴🔴🔴	🔴🔴🔴🔴🔴
llama-3.2-90b-vision-preview	🟢🟢🟢🔴🔴	🟢🟢🟢🔴🔴
qwen-2-vl-72b-instruct	🟢🟢🟢🔴🔴	🟢🟢🔴🔴🔴
pixtral-12b	🟢🟢🔴🔴🔴	🟢🟢🟢🔴🔴

I used an app I built for this.

Here is the original image along with the individual results.

Update

Adding gridlines with labeled axes helps the LLMs. (Thanks @Bijan Mishra.) Here are a few examples:

Google may acquire Riya

Thu, 17 Nov 2005 12:00:00 +0000

Google may acquire Riya. Riya can recognise faces and lettering in your pictures.

Comments

Sai 8 Dec 2005 11:50 pm: Hi, do post the rest of the tips for excel when you get the time. Good Stuff.

Identify location by photographs

Wed, 14 Apr 2004 12:00:00 +0000

There’s software that can locate where you are using a photograph of buildings.

Cats

Tue, 26 Mar 2002 12:00:00 +0000

Amazing, what some people try and do with cats. And image recognition.