Advances in Multi-Subject Generation and Flow Matching

The field of multi-subject generation and flow matching is rapidly advancing, with a focus on improving the fidelity and diversity of generated images and videos. Recent research has introduced new frameworks and techniques, such as spatially disentangled attention and identity-aware reinforcement learning, to address the challenges of attribute leakage and subject entanglement. Additionally, there is a growing interest in flow matching methods, including diffusion bridge and flow matching, which have been shown to be effective in various applications, including image and video generation. Noteworthy papers in this area include MultiCrafter, which proposes a novel framework for high-fidelity multi-subject generation, and Diffusion Bridge or Flow Matching?, which provides a unified theoretical and experimental validation of diffusion bridge and flow matching methods. Other notable papers include Optimal Control Meets Flow Matching, which introduces a principled route to multi-subject fidelity, and DisCo, which achieves state-of-the-art multi-subject fidelity across models.

Sources

MultiCrafter: High-Fidelity Multi-Subject Generation via Spatially Disentangled Attention and Identity-Aware Reinforcement Learning

High-Order Progressive Trajectory Matching for Medical Image Dataset Distillation

Diffusion Bridge or Flow Matching? A Unifying Framework and Comparative Analysis

Spectral Flow Learning Theory: Finite-Sample Guarantees for Vector-Field Identification

Energy Guided Geometric Flow Matching

Exact Solutions to the Quantum Schr\"odinger Bridge Problem

Data-to-Energy Stochastic Dynamics

Learning Passive Continuous-Time Dynamics with Multistep Port-Hamiltonian Gaussian Processes

Multi-Marginal Flow Matching with Adversarially Learnt Interpolants

DisCo: Reinforcement with Diversity Constraints for Multi-Human Generation

Multi-marginal temporal Schr\"odinger Bridge Matching for video generation from unpaired data

Optimal Control Meets Flow Matching: A Principled Route to Multi-Subject Fidelity

Built with on top of