Using "underdrawings" for accurate text and numbers (samcollins.blog)

A new technique discovered for creating accurate text and numbers in AI-generated images, termed "underdrawings," has shown promising results, outperforming current state-of-the-art models like Gemini 3.0 Pro and ChatGPT-Images-2. The method involves constructing a precise layout of text and numbers in the first layer, termed the "underdrawing," followed by a generative image creation step. This two-layer approach allows users to leverage the strengths of AI in both analytical precision and artistic creativity, capturing complex structures, such as a game board with numbered stepping stones arranged in a spiral. This innovation is significant for the AI/ML community as it fills a notable gap in current generative models, which often struggle with accurate representations of text and numbers within complex compositions. By employing this underdrawing technique, developers and artists can produce images that are not only visually appealing but also semantically accurate. The ease with which existing models like Claude can assist in this process highlights the potential for further advancements in AI-driven creative tasks, indicating that with refinement, this approach could become a standard practice in the industry.