The field of generative models is rapidly advancing with a focus on efficiency, innovation, and improved performance. Recent research has explored alternative methods to traditional function estimation, such as estimation-free generative methods and proximal diffusion models, which offer theoretical and practical benefits. Another area of focus is the acceleration of generative models, including the use of region-adaptive latent sampling, warm starts, and rectified flow matching. These innovations have the potential to significantly improve the performance and efficiency of generative models. Noteworthy papers include 'Data Generation without Function Estimation', which proposes an estimation-free generative method, and 'RODS: Robust Optimization Inspired Diffusion Sampling', which introduces a novel method for detecting and reducing hallucinations in generative models. 'Upsample What Matters' is also noteworthy for its region-adaptive latent sampling approach, which achieves significant speed-ups without compromising image quality.
Efficient and Innovative Generative Models
Sources
Data Generation without Function Estimation
Upsample What Matters: Region-Adaptive Latent Sampling for Accelerated Diffusion Transformers
Beyond Scores: Proximal Diffusion Models
Theory-Informed Improvements to Classifier-Free Guidance for Discrete Diffusion Models
Learning Diffusion Models with Flexible Representation Guidance
Shortening the Trajectories: Identity-Aware Gaussian Approximation for Efficient 3D Molecular Generation
Can Contrastive Learning Improve Class-Imbalanced Diffusion Model?
Warm Starts Accelerate Generative Modelling
RectifiedHR: High-Resolution Diffusion via Energy Profiling and Adaptive Guidance Scheduling
Inversion-DPO: Precise and Efficient Post-Training for Diffusion Models
Schr\"odinger Bridge Consistency Trajectory Models for Speech Enhancement
RODS: Robust Optimization Inspired Diffusion Sampling for Detecting and Reducing Hallucination in Generative Models
Hierarchical Rectified Flow Matching with Mini-Batch Couplings