A new AI framework that delivers high-quality image generation without the extensive computational power typically associated with modern artificial intelligence has earned researchers at the University of Surrey one of the highest honours for computational efficiency and sustainability in AI.
The team from the Surrey Institute for People-Centred AI has received the inaugural Compute Gold Star at the 2026 Computer Vision and Pattern Recognition Conference (CVPR) – one of the world’s leading AI and computer vision events. Out of more than 16,000 submissions to the conference, the Surrey team was one of only 18 globally to receive the award.
The award recognises their new AI framework, CaricHarmony. While many modern AI systems often require vast computing resources and expensive, time-consuming fine-tuning to learn new concepts, CaricHarmony operates without additional pre-training data. The system can run on a single consumer-grade graphics card (RTX 4090) and generate complex images in under 16 seconds, significantly reducing the computational demands associated with many comparable models.
To demonstrate the efficiency of their architecture, the research team applied it to one of the most challenging tasks in computer vision – generating high-quality caricatures from a rough sketch.
AI models have historically suffered from what researchers call "signal contamination" when trying to process a person's identity and an exaggerated shape at the same time, often resulting in either a bland portrait or an unrecognisable distortion. CaricHarmony separates these instructions into parallel pathways, allowing the system to balance both requirements without the need for computationally intensive training.
Dongyu Wang, lead author of the study and postgraduate researcher at the University of Surrey, said:
“Current AI models often require massive computing power and expensive retraining just to learn a single new concept. We wanted to prove that you can achieve highly complex, creative results using a fraction of the energy. By solving the fundamental architectural conflict, we removed the need for heavy computation entirely.”
Dar-Yen Chen, co-author of the study and postgraduate researcher at the University of Surrey, said:
"The challenge was not simply generating better caricatures, but doing so efficiently. Separating the conflicting signals into parallel pathways meant we could bypass the heavy optimisation loops that make other systems so slow and environmentally costly."
Professor Yi-Zhe Song, co-author of the study and Co-Director of the Surrey Institute for People-Centred AI at the University of Surrey, said:
"As AI research increasingly relies on ever-larger computing resources, competing for top conference honours as a university research lab can become increasingly challenging. This award is a significant validation of the impact that academic research can still have.
"The CVPR community's decision to recognise this reality and reward compute efficiency is particularly encouraging. It shows there is still a clear path for pure algorithmic innovation. Demonstrating that a single consumer-grade GPU and no additional training data can deliver a world-class AI breakthrough highlights what is possible for democratised and sustainable AI."
[ENDS]