Comment by jjcm

A reasonable conclusion, but you don't have all the data here in this particular instance. These are all part of a set of benchmark prompts I have that are designed to try and push the images to their extremes. The banking one is dense because the prompt itself specifies each of these things - the goal with this test is to see how well it can handle many different UI elements. Here's an excerpt from the json blob that is used as the prompt:

  "subjects": [
    {
      "description": "A transparent header spans the top of the teal hero. The white ANNUI wordmark sits at the upper left, followed by Product, For Agents, Resources, and Company links. For Agents is underlined as the active page. Sign in appears near the upper right beside a large white rounded Get started button.",
      "position": "top edge, full width, approximately 8% of viewport height",
      "subject": "Navigation Header"
    },
    {
      "description": "Left-aligned hero content includes the spaced eyebrow 'FOR AGENTS', a large cream serif headline reading 'Your clients’ brighter tomorrow.', supporting copy about modern financial software, real support, and helping independent agents grow their annuity business, plus a white pill-shaped 'Build with Annui' button and a circular play control labeled 'Watch for Agents (1:30)'.",
      "position": "left half, upper-middle through lower-middle, approximately 46% width and 58% viewport height",
      "subject": "Hero Copy and Actions"
    },
    {
      "description": "A monumental classical female statue dominates the center-right of the hero. The figure has a pale blue-green face, dark teal hair, a radiant gold helmet and laurel details, and layered gold and blue garments. Architectural columns and cloudy sky fill the background, creating a sculptural financial-services editorial scene.",
      "position": "center-right, extending from the top edge to the bottom of the hero, approximately 55% width and 85% viewport height",
      "subject": "Classical Statue Hero Artwork"
    },
    {
      "description": "A narrow text panel overlays the far-right architectural columns with vertically spaced uppercase text reading 'INDEPENDENT ADVICE A BRIGHTER TOMORROW', a short horizontal divider, and 'BUILT TOGETHER'.",
      "position": "right edge of hero, upper-middle through lower-middle, approximately 15% width and 42% viewport height",
      "subject": "Vertical Message Panel"
    },
    {
      "description": "A full-width warm off-white strip beneath the hero contains four evenly spaced benefit columns separated by thin vertical rules. Each column has a pale beige circular icon and dark text: 'Streamline Sales' with quote and compare copy, 'Serve Clients Better' with carrier and education copy, 'Grow Your Practice' with tools and marketing support copy, and 'Stay Independent' with client, advice, technology, and support copy.",
      "position": "below the hero, full width, approximately 18% viewport height",
      "subject": "Agent Benefits Strip"
    }
  ],

That's not to say that GPT-Image-2.5 doesn't have a certain smell to it though. It definitely has its isms.

I very firmly disagree. IMO image models are extremely under-used in this area.

Just for context, I ran design systems at Figma for 5 years / led their first prompt->design stream before quitting to make my own startup in this space.

Image models are trained on visual creativity, color usage, whitespace, etc - all the ingredients you need for good design principles. Where they struggle is often with information density and text rendering, which LLMs excel at. You can use the best parts of both of them though - use an image model for the initial design and an LLM to convert it to code.

One thing that's very telling is with openai's sites feature, in its chain of thought it actually renders out an image first as a design reference, then builds towards that.

I actually was just talking about this with a peer. He was saying that people inside labs and data brokers use image models for RL training of LLMs to get better at these sort of tasks, which would make sense as a way to try to embed the knowledge into the LLM - similar to how Astra was clearly trained with more CAD, 3D, and Robotics in mind.

Or the gpt-image "fried" texture. Interestingly the best approach I've found so far is to use gpt-image for the initial design, but nano banana for the asset export once I have the design. IE in the case of the "Annui" design, I'd take that design and its backing json structure, hand it off to Opus to build it in html, then use nano banana to build out the images of the statues/background images.

You just gotta use the best parts of all of them :)

Thank you haha

And benchmarks do not account for it sadly

Also, sometimes you want to do “ugly” things on purpose or add imperfections like film grain and such

> do you find that a lot of people use diffusion models for UI mockups? i’ve always been surprised that people wouldn’t use LLMs exclusively for that instead

Not yet, but it's what I quit my role at Figma to make a bet on for my startup. In my experience diffusion is a far, far better creative designer than LLMs. LLMs are trained for consistent output, which makes sense for a billing function, but leads to some very obvious smell when doing frontend work (take claude design as an example - it's often very obvious what comes out of it). Image models have a bit more creativity baked in, and are trained on color usage, whitespace, and visual balance a lot more. I've found the image->html approach leads to much better results, basically using an image model as an architect.

Would love to be connected with the team for when you all are ready to finetune it. My email is j@diffui.ai

When VL models started becoming popular a while back, a relatively common workflow was to use generative models (like the original Nano Banana) to generate initial UI mockups and then throw the images at a VL model to convert them into CSS and HTML.

However feels like LLMs such as Opus 5.5 in particular are surprisingly adept at generating UI mockups right off the bat these days. And since they’re already code, they’re much easier to modify and tweak.