Tech Graphics Benchmark
Tech Graphics Benchmark: test prompts Write-up: Can image models spell the labels? A first look at my Tech Graphics Benchmark Test prompts for Tech-Graphics-Benchmark: can an image model make a characterful banner or a restyled diagram for a tech repo or blog post without getting the text or the arrows wrong? prompts.jsonl, one record per prompt: Field Meaning id tgb-t2i-NN (text-to-image), tgb-i2i-NN (image-to-image) or tgb-he-NN (Hebrew track) modality… See the full description on the dataset page: https://huggingface.co/datasets/danielrosehill/Tech-Graphics-Benchmark.
About This Project
Tech-Graphics-Benchmark provides a structured collection of prompts designed to stress-test text-to-image and image-to-image models. The project specifically evaluates whether generative AI can produce high-quality technical assets, such as banners or diagrams, without introducing errors in text rendering or spatial alignment of graphical elements like arrows.
The benchmark is organized into a JSONL format, categorizing prompts by modality and track, including support for Hebrew language integration. It serves as a tool for researchers and developers to measure the proficiency of AI models in handling technical communication tasks, providing a standardized set of test cases for assessing visual accuracy in technical contexts.