Show HN: Extending LLM SVG generation beyond pelicans and bicycles (gally.net)

A Show HN benchmark tested six leading large models on 30 creative natural-language → SVG generation prompts to push evaluation beyond toy examples like “pelicans and bicycles.” Rather than image outputs, the exercise focuses on models’ ability to produce valid, compact, and stylistically faithful SVG code from freeform prompts, exposing how LLMs handle vector primitives, path commands, transforms, fills/gradients and other programmatic drawing constructs. The comparison covers Anthropic’s Claude Sonnet 4.5, xAI’s Grok Code Fast 1 (314B MoE), Google’s Gemini 2.5 Pro, DeepSeek V3.2‑Exp (685B/37B MoE), Zhipu AI’s GLM‑4.6 (355B/32B MoE), and Alibaba’s Qwen3‑VL‑235B‑A22B‑Thinking (235B/22B MoE). Technically this matters because text-to-SVG is a different challenge than pixel generation: correctness and renderability of generated code, succinct use of vector primitives, and predictable composability are essential for downstream design, prototyping, and programmatic graphics pipelines. The benchmark therefore highlights practical concerns for the AI/ML community—MoE vs dense scaling trade-offs, instruction tuning for code-structured outputs, hallucination and syntax errors, and the need for integrated render-check feedback or programmatic verification. Results from such comparisons can guide model selection for tools that require reliable vector outputs and motivate research into tighter grounding, differentiable raster feedback, and compactness-optimized generation.