AI models struggle with generating high-quality, on-brand UI. We developed a model fine-tuned for UI code generation and worked with OpenAI to build it.
Generating front-end UI code that has taste and is aligned with your brand design is a big unlock for anyone building or experimenting with websites & UIs in a production setting. This model is a major step towards this.
We accomplished this with a large, curated set of websites with sections masked out, prompting the model to generate the masked-out section. The resulting model produced 26% more visually-aligned code than the GPT-4o base model, as measured by our LLM-as-a-judge eval.
We will be making this model available through our product soon. We're still working on a few things before releasing it, including making the generated code responsive, more reliable, and compatible with our product and content data model.
We'll be sharing more details and opening up a benchmark to the community in the near future. For now, you can read more about this and other Coframe features in our blog: https://lnkd.in/gZ6MmkSQ
We are excited to see the productivity, engagement, and conversion lifts that these new code generation capabilities will drive for our users. Read more about our work in OpenAI's Vision Fine-Tuning announcement: https://lnkd.in/g43WTXx6
If you want to drive more conversions on your website, please book a call with us: https://lnkd.in/gVSWjrje
We're hiring for product engineering, AI engineering, infra, and more! DM me or apply within: https://lnkd.in/gCfqWS_n