SFT made it slower: OpenUI's self-training loop lifted generative-UI validity from 13% to 71.7%

1glasspaani · reddit · 2026-09-09

The OpenUI team details fine-tuning an open-weight model for Generative UI on consumer GPUs. Starting from DiffusionGemma, their first SFT attempt—a LoRA on 700 synthetic examples—lowered training loss but worsened benchmarks: broken component props and references. Constrained to a single component library, structural validity rose from 13% to 28.8%, but generation slowed from 1.6s to 4.3s with roughly double the denoising steps.

They then switched to a self-training loop:

Result: 57.1% structural validity at 1.9s generation; the median repair changes one statement, and each pass takes 1–2 hours on a single A100. Repeating across 27 component libraries, the final model OUI-1 reaches 71.7% (structural validity only—not a guarantee of good looks or task completion). Blog, weights, and benchmark are public.

Original post →

More from Research

Research channel →