Show HN: Pelican-bicycle alternatives
summary
This project evaluates the ability of various large language models to generate Scalable Vector Graphics (SVG) based on complex, surreal prompts. By comparing model outputs from 2025 and 2026, the site tracks progress in AI's spatial reasoning and visual composition. The benchmark uses thirty specific, imaginative scenarios to ensure consistency across different AI versions. This approach reveals how different models handle the intersection of code generation and visual creativity.
key points
- The benchmark consists of thirty distinct prompts featuring animals performing human tasks, such as an octopus playing a pipe organ, to test the limits of AI visual logic.
- The site compares results from multiple top-tier models including GPT, Claude, Gemini, and DeepSeek across two different timeframes.
- Detailed metadata for each generation is provided, including the time taken to produce the SVG and the monetary cost of the API call.
who should read it
AI researchers, prompt engineers, and developers interested in the evolution of LLM capabilities in code-based image generation.
discussion positive
consensus
Commenters generally agree that while AI models have improved significantly in SVG generation, they still struggle with spatial coherence and complex entity intertwining. There is a recurring debate regarding whether these benchmarks are 'Goodharted' due to training data leakage or if the improvements reflect genuine spatial reasoning. Many users observe a striking similarity in output designs across different models, suggesting shared biases or overlapping training sets.
top insight
SimonW argues that Gemini's superior SVG performance isn't just due to general web scraping, but likely stems from targeted effort, such as RLHF or RLVF, specifically aimed at improving SVG output. This distinguishes the result from simple data contamination, suggesting that specific architectural or training optimizations can yield visible gains in a niche visual task.
Generated by an LLM from the article and thread. Check the sources before relying on it.