Search papers, labs, and topics across Lattice.
This paper benchmarks a two-stage pipeline for generating safety-relevant automotive instructional imagery, where LLMs extend manual descriptions into image-generation prompts for ADAS scenarios. A factorial evaluation of prompt-extension LLMs, thinking modes, CoT sources, prompt styles, scenarios, and image generators revealed that the image generator is the primary quality determinant, with specific Google models significantly outperforming others. LLM and CoT effects were found to be generator- and scenario-dependent, highlighting the importance of evaluating the entire coupled system for domain-constrained instructional image generation.
Most generative image models fail to produce usable automotive instructional imagery, with only two Google-based generators achieving >14% usable outputs, underscoring the critical role of the image generator over prompt engineering.
Generative image models can support instructional visual content, but safety-relevant automotive imagery requires more than photorealism: generated scenes must preserve cockpit geometry, human-machine-interface (HMI) iconography, and physically plausible vehicle behavior. We evaluate a two-stage pipeline in which large language models (LLMs) extend technical owner-manual descriptions and reference-image constraints into downstream image-generation prompts, across four Cadillac LYRIQ advanced driver assistance system (ADAS) scenarios. Using a factorial design over prompt-extension LLM, thinking mode, injected Chain-of-Thought (CoT) source, prompt style, scenario, and image generator, we collect N = 18, 551 manually rated images (Excellent/Good/Fair/Poor). The downstream generator was the dominant quality determinant: Nano Banana Pro and Nano Banana 2 produced 16.9% and 14.3% usable images, whereas GPT-Image-1.5 produced 0.5% and collapsed to near-uniform failure. Within the viable Google generators, LLM and CoT effects were generator- and scenario-dependent; because both viable generators are Google, the CoT-source effect reflects trace鈥揼enerator alignment (Google-aligned CoT helped, OpenAI-sourced CoT hurt) rather than a general CoT benefit. A robustness analysis further reveals a trade-off between high-performing specialist and lower-variance generalist configurations. Domain-constrained instructional image generation should therefore be evaluated as a coupled system spanning generator, prompt-extension model, reasoning trace, prompt template, and scenario.
