The video demonstrates an iterative SVG builder powered by GPT-5 that generates and progressively refines SVG images through multiple iterations, showcasing improvements in detail and realism, especially with GPT-5 Mini at high reasoning levels. While the system is straightforward and customizable via prompts, the creator notes some limitations in refinement consistency and plans to share the code on Patreon for further experimentation and development.
In this video, the creator demonstrates an iterative SVG builder powered by GPT-5, which generates and then refines SVG images through multiple iterations. The process begins by inputting an SVG description, such as “Eiffel Tower,” and specifying the number of refinement iterations. Users can optionally provide instructions for each iteration to guide improvements, though in the demonstration, this was left empty. The tool generates SVG files saved in a designated folder, progressively enhancing the image with each iteration.
The video showcases examples of SVGs created by different GPT-5 models, including GPT-5 Mini and GPT-5 High, at varying reasoning levels. The creator highlights how GPT-5 Mini, particularly at high reasoning, tends to add more detailed and realistic features with each iteration, such as improving eyelashes and adding depth to the eyes. However, some aspects like the background remain less developed. Comparisons between iterations reveal incremental improvements, though sometimes changes are subtle or unexpected, like the addition of extra eyebrows.
The creator also explores the performance of GPT-5 at minimal reasoning effort, noting that if the initial SVG is poor, the model struggles to recover fully. Despite this, the model still manages to add interesting details over iterations, such as a small blue gem. The iterative process takes some time, and the creator suggests that refining the iteration prompt could lead to better results, as some images showed limited improvement or even loss of detail in later stages.
Throughout the video, the creator emphasizes the straightforward nature of the system, which relies on a system prompt defining the expert SVG artist role and an iteration prompt aimed at enhancing realism and color harmony. These prompts can be adjusted to potentially improve the output quality. The creator plans to share the code for this project on Patreon, making it accessible to paying subscribers, while also providing links for free viewers to access the post and code.
In conclusion, the video serves as an experimental showcase of using GPT-5 for iterative SVG generation and refinement. While the results are promising, especially with GPT-5 Mini at high reasoning, there is room for improvement in prompt design and model tuning. The creator invites viewers to explore the project further through Patreon, where it is available alongside hundreds of other projects, encouraging experimentation and development in AI-assisted SVG creation.