
Research Model Playground
Papers
Fine-Tuning Next-Scale Visual Autoregressive Models with Group Relative Policy Optimization
Authors: Matteo Gallici, Haitz Sáez de Ocáriz BordeFine-tuning next-scale visual autoregressive models with reinforcement learning enhances image quality, style control, and generalization beyond pre-trained distributions.
Our Research Philosophy
At Supermodel, we are dedicated to redefining what's possible in generative AI, guided by five core principles:
Science, Art & the Human Experience
There is a unique beauty in bringing together seemingly unrelated fields of human knowledge.
We believe that the most significant advancements occur where cutting-edge scientific research, artistic expression, and genuine human connection intersect.
Fast First–Principles Innovation
We combine rapid experimentation with profound theoretical understanding.
This allows us to push the boundaries of foundational theory while also developing real-world AI applications and products for the public to experience and interact with.
Empowering Creative Control
Generative models to empower human creativity.
By building fine-grained control mechanisms into our generative models, we empower creators to safely and effectively harness our technology.
Collaboration
We believe in the power of collective human intelligence.
We foster an open, cross-disciplinary culture that leverages collective expertise to solve deep learning's most challenging problems. We believe in sharing knowledge, and we are open to collaborations with universities and other research institutions.
These pillars collectively drive us to create next-generation generative products, from interactive avatars and ultrafast video synthesis to sophisticated image generators. Our research covers the entire generative model creation pipeline, including critical but often-overlooked areas like data curation, compression algorithms, and alignment strategies.