Coco Labs

More and more, image models get ranked by human vote. A vote can miss that your font changed. Today @DavidZats and I are releasing VybBench, an open benchmark that scores design fidelity with numbers instead of votes. Ask a model to change one headline and it can nudge your font, shift your blue and redraw a corner of your logo too. Each edit changes a little more, until the design isn't yours. VybBench scores each piece: 145 fonts, matched pixel by pixel against the real font 100 hex codes, scored on how far each color drifts 200 small edits, scored on how many pixels changed, with a judge checking the edit was made We are not including model scores in the post, but you can see how each model scores on the Hugging Face. Every prompt, output, judge verdict and the scorer is on Hugging Face, link in the reply. Full Disclosure: At @cocoailabs we also build Vyb - our first design model built to preserve design fidelity and taste.

I launched the first image and video generation models at Google, and kept asking why all AI designs look the same. AI has brought us the WORST era of design. It made content creation effortless, but it also made 90% of it look painfully generic. My co-founder and I spent 2 years building in stealth to change that. Meet Vyb, the first AI design model that understands your taste like a designer would. RT + comment your website URL and we’ll make your first design.

The only way to edit an AI-generated design is to generate another one. The text in it is not really text. It is part of the picture, so a one-word fix comes back as a whole new image, different everywhere. People try to get around that by pulling the layers back out. The model never drew the background behind the logo, so it has to invent what was underneath. It reads the text off the picture and picks a font that looks close. You end up with a copy of the design you already had.

Design generation is not the same as image generation. An image model makes one flat layer of pixels, the same way it would make a photograph. It looks like a design. But the logo in it cannot be moved and the headline cannot be rewritten, because they are not separate things. They are pixels. So the only thing left to do is ask for another one, which is not the right process for designers. Design generation means building the design out of a company's own logo, fonts and photos, and placing every part of it on purpose. A better image model does not get you there.

Design is judged on novelty, so companies throw out ads that still look good and pay for new ones. A model gives you back the patterns it has seen most, which means you can pay for a design and get one that could belong to your competitor. Nobody has worked out how to get new design work into a model. I would not say it cannot be done. I think it is the open question underneath all of this.

A model built for one domain already understands the context you are working in, so you are not describing it again on every request. That is most of what makes a domain-specific model easier to prompt than one built to answer anything.

Open weights decide who can realistically build a model. Training one from nothing takes the resources of a frontier lab. By allowing companies to avoid the prohibitive costs of starting from scratch, companies can create state of the art models for specific domains. @cursor_ai post-trained @Kimi_Moonshot K2.5 into Composer 2.5. @perplexity_ai built its Sonar search model on Llama 3.3 (@AIatMeta) @harvey post-trained Kimi K3 into Tenet, its legal model. All of them are doing work a frontier lab was never going to do for them.

At @Google, I had the opportunity to be part of the team that took the company's first image and video generation models from research to launch. Two years ago I left to start a lab, and it comes out of stealth at the end of August. Send me a message if you’d like to test it.