Should I Render or Should AI Generate?
Crafting Synthetic Semantic Segmentation Datasets with Controlled Generation
New scientific milestone from the DDS ecosystem!
In collaboration with CITIC – Centro de Investigación TIC, Universidade da Coruña, our team has published a peer-reviewed paper that powers the core of our generative platform for conscious design.
"Should I Render or Should AI Generate? Crafting Synthetic Semantic Segmentation Datasets with Controlled Generation"
Read the full paper
A breakthrough in dataset generation
The study introduces a novel approach using Diffusion Models + ControlNet to generate high-quality, annotated synthetic datasets — faster, cheaper, and with better results than traditional rendering techniques.
This research proves a bold point:
AI-generated datasets — when crafted wisely — can be comparable and sometimes even better than traditional physically-based rendered synthetic datasets.
Moreover, it makes the process much easier and cheaper to author, providing wider variability and reflecting many uncommon conditions in real-world datasets.
Authors and collaborators
Huge kudos to the authors and our research partners at CiTIC:
- Luis Omar Álvarez Mures
- Manuel Silva Díaz
- Manuel Lijó Sánchez
- Emilio José Padrón González
- José A. Iglesias
And deep thanks to the institutions and supporters backing this research.
Final thoughts
This is how deep science becomes transformative technology.
Stay tuned — the future of design is being generated.
In the car parts domain, segmentation models trained on rendered, generative, and real images were evaluated. Generative datasets produced clearer, less noisy outputs with stronger separation between semantic components (grille, headlamps, logos, etc.).