Redirecting to original paper in 30 seconds...
Click below to go immediately or wait for automatic redirect
📄 Abstract
Abstract: Current text-to-image generative models are trained on large uncurated
datasets to enable diverse generation capabilities. However, this does not
align well with user preferences. Recently, reward models have been
specifically designed to perform post-hoc selection of generated images and
align them to a reward, typically user preference. This discarding of
informative data together with the optimizing for a single reward tend to harm
diversity, semantic fidelity and efficiency. Instead of this post-processing,
we propose to condition the model on multiple reward models during training to
let the model learn user preferences directly. We show that this not only
dramatically improves the visual quality of the generated images but it also
significantly speeds up the training. Our proposed method, called MIRO,
achieves state-of-the-art performances on the GenEval compositional benchmark
and user-preference scores (PickAScore, ImageReward, HPSv2).
Authors (5)
Nicolas Dufour
Lucas Degeorge
Arijit Ghosh
Vicky Kalogeiton
David Picard
Submitted
October 29, 2025
Key Contributions
Proposes MIRO, a method that conditions text-to-image models on multiple reward models during training, rather than using post-hoc selection. This approach significantly improves visual quality and training efficiency, aligning generated images better with user preferences and achieving state-of-the-art performance on compositional benchmarks.
Business Value
Enables the creation of more aesthetically pleasing and contextually relevant images, accelerating creative workflows and improving user engagement in applications relying on image generation.