news.volyx.in

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

379 points by samcollins · 121 days ago · 131 comments on HN

Article summary

The article discusses a technique for generating accurate text and numbers in AI-generated images by combining deterministic image rendering with SVG and rich image transformation with Gemini 3.0 Pro. This method involves creating an outline image with precise text and numbers using SVG, and then passing it to an image model to transform it into a richer visual while preserving the key details. The author found this technique to be reliable and useful for creating images with accurate text and numbers. The method is conceptually similar to traditional underdrawings used in art.

Main themes

  • AI-generated images
  • Text and number accuracy
  • Image rendering techniques
  • SVG and Gemini 3.0 Pro
  • Multimodal LLMs
  • Image generation limitations

What commenters say

  • The technique described in the article is a useful trick for achieving accurate text and numbers in AI-generated images, but it may not be a novel approach.
  • Some commenters argue that the limitations of LLMs are not fundamental, but rather a result of current technology and training data.
  • Others believe that LLMs have inherent limitations, such as hallucinations, that cannot be fully overcome, even with improvements in technology and training data.
  • The definition of hallucinations in LLMs is disputed, with some arguing that it is too broad and others believing it is a fundamental limitation of the technology.
  • The use of auxiliary tools, such as SVG, can help improve the accuracy of LLMs, but some argue that this is not a long-term solution.
  • The article's technique is seen as a step towards understanding the strengths and weaknesses of LLMs and how to use them effectively.
  • Some commenters believe that the burden of proof lies with those who claim that LLMs can overcome their limitations, rather than those who argue that they are fundamental.
  • The discussion highlights the need for a deeper understanding of LLMs and their limitations, as well as the development of a taxonomy of work and studies on their capabilities and limitations.