Coco Labs Vyb: A Design Model That Returns Editable Layers, and What Is Only Vendor Benchmark

Most AI graphics tools hand you a picture. Coco Labs says its first model, Vyb, hands you a design file. Announced around October 20, 2026, Vyb builds ads, posters and similar marketing graphics element by element from your real logo, fonts and photos, and every result comes back in layers that you can edit individually. The company was founded by former Google Brain and DeepMind researchers, and describes itself on cocolabs.ai as having built "the model," not "an AI design app."

The pitch is concrete and the evidence is thin: access is a waitlist, there is no public API or price, and the benchmark behind the headline numbers is one Coco Labs wrote itself. This guide separates what the company says from what has been independently shown.

Claimed versus verified

What Vyb is supposed to do

The company's homepage makes three statements. "All AI images look the same, so we built the first design model." "Vyb builds designs, not images. Each one is made element by element from your real logo, fonts and photos." And "Designers have worked in layers for 30 years. AI forgot. Every Vyb design comes back in layers, and every layer is yours to edit. Edit just the headline, like any design file."

The site's gallery shows prompts such as a skincare launch ad, an editorial lip-product ad citing survey percentages, a pet-accessories ad, a snack-bar ad with fiber and protein claims, a denim collection poster, a dog tracker ad with app-style notifications, a dental ad and a dumpling poster. Those are curated examples, so they show intent, not typical output.

According to the launch post by David Zats and Anusha Ramesh, Vyb composes designs in an internal structured format called Cocolang, in which text, vector shapes and images keep native formats (text stays text, graphics stay vectors, images stay pixels) rather than being flattened to pixels. Because a color can be written as a value instead of painted, exact color reproduction is a natural advantage of the format.

The benchmark: read it as a vendor demo

In its launch post, Introducing Vyb, Coco Labs reports a comparison covering 145 random Google fonts and 200 chat edits (three attempts each) against ChatGPT Image, Muse and Recraft, with these company-reported results:

The launch post reports a measurement noise ceiling of 94.4% for typeface fidelity, so 90.3% is near the practical maximum for that test.

How to read these:

• The benchmark's author is the winner. That does not make the numbers false, but nobody outside Coco Labs has reproduced them.
• Edit locality is a property of the format. A model editing a document changes little by design, while a model that regenerates an image changes almost everything. The comparison is therefore between two input contracts, not two models given identical tasks.
• Fonts are public Google Fonts only. Brands often use custom typefaces, which the comparison did not measure; the authors expect a larger gap there.
• ChatGPT Image completes more edits. The judge scored it 100% against Vyb's 94.7%, which the authors attribute to Vyb sometimes declining edits that would hurt design quality, such as making text unreadable. That is an interpretation, not a finding we can check.

The layers argument

Coco Labs argues that converting a finished flat image back into layers cannot recover real text or vector layers, since the text in a flat image is only pixels and hidden regions must be guessed, so a genuine layer stack has to be generated. It is a defensible position and also the position of the company selling the alternative. Teams with large libraries of approved flat assets would still want a way to make them editable, which Vyb, as described, does not address. For context on neighboring tools, see our coverage of Claude Design from Anthropic Labs, the OpenAI product design plugin that goes from prototype to Figma, and ChatGPT Images 2, plus Krea 2's open-weight image model.

Why layers matter to working designers

Anyone who has fixed a typo in an AI-generated ad knows the problem: you re-run the prompt, get a slightly different design, and lose the layout you had approved. A flat output makes the model the only editor. A layered output moves the control back to the designer, who can change copy, swap a photo, adjust spacing or localize a headline without touching the rest of the file. That is a workflow argument rather than a quality argument, and it is why the edit-locality row of the benchmark is the one that deserves the most attention. If the numbers hold, the useful unit of comparison is not how good the first image looks but how many minutes a designer spends getting from an approved layout to ten localized variants.

It also changes who can use the tool. Marketing teams that already work from brand kits, with fixed logos, fonts and color values, are the natural audience, because Vyb's pitch is fidelity to those inputs rather than inventing a style. Teams that want exploratory, painterly visuals are better served by conventional image generators, and the company's own framing, that images "all look the same," is aimed at the template-like marketing graphics where brand correctness matters more than novelty.

Where the company stands

Its team background, launching early image models at Google and building models for YouTube Ads, is a credible pedigree for the problem, but it is not evidence that this product works. Vyb is its first model release, and the company has not published parameter counts, an architecture description or training data details in what we found. The waitlist model also means the reach of the product is unknown: we do not know how many designs have been generated, by whom, or with what success rate.

What is still unproven

1. Price. No rate card exists. The $10 credit says nothing about the cost of a finished campaign.
2. Independent testing. Every number comes from the vendor.
3. File quality. "Lossless export to Figma" has not been confirmed by designers opening the files. Whether the layer tree is tidy, named and usable matters more than whether it exists.
4. Coverage of formats. The examples show posters, ads and announcements. Multi-page documents, packaging with dielines and print color profiles are not addressed in the material we saw.
5. Brand and rights. Extracting logos, fonts and photos from a website raises licensing questions; check the terms before using assets you do not own.

How to evaluate it when you get access

• Bring three real brand kits, including one with a custom font, and ask for the same ten designs from Vyb and your current tool.
• Open each file in Figma and have a designer rate editing effort, layer naming and fidelity to brand rules.
• Time the full path from brief to approved asset, including revisions, because the claimed advantage is in the edit loop, not the first draft.
• Check what happens when you ask for a change that conflicts with legibility, to see how often it refuses.

Bottom line

Vyb is an interesting bet that the next step in AI graphics is structure rather than prettier pixels. The idea that text should stay text, vectors should stay vectors and every element should sit on its own layer matches how designers already work, and it is easy to see why a team localizing campaigns would care. But today the evidence consists of a company website, a curated gallery and a benchmark written by the winner. Treat the 90.3% typeface figure and the roughly 12% edit footprint as hypotheses to test on your own brand kits once access opens, not as settled results. Until independent designers have opened real exported files in Figma, the honest status is promising, unverified and waitlist-only.

Related reading

• Claude Design from Anthropic Labs
• OpenAI product design plugin: prototype to Figma
• ChatGPT Images 2 (gpt-image-2)
• Krea 2 open-weight image model
• Figma Config 2026 recap
• Official: Coco Labs and Introducing Vyb

Details are accurate as of October 22, 2026 and come from the Coco Labs website and launch post. Benchmarks are the vendor's own.