Topic
diffusion-models
Divide-and-Denoise: Game-Theoretic Method Ensures Fair Composition of Diffusion Models
Researchers propose Divide-and-Denoise, a game-theoretic method for composing multiple pre-trained diffusion models fairly. At each timestep, an allocation divides the noisy sample into regions, maximizing utility under fairness constraints. The method outperforms baselines on the GenEval benchmark, resolving common failures like missing objects and mismatched attributes.
Trust-Region Diffusion Policies Enable Expressive AI for Complex Control Tasks
Researchers introduce Trust-Region Diffusion Policies (TruDi), a method that enables diffusion models to be used in massively parallel on-policy reinforcement learning. By enforcing a KL-divergence constraint over the entire diffusion trajectory, TruDi achieves stable training and outperforms strong baselines across 73 diverse tasks, showing particular gains on challenging humanoid control problems.
Who Should Lead Decoding Now? Tracking Reliable Trajectories for Ensembling Masked Diffusion Language Models
Masked Diffusion Language Models (MDLMs) have emerged as a distinct paradigm for sequence generation, but combining their knowledge is an underexplored problem. Researchers introduce TIE (Trajectory-based Iterative Ensembling), a framework that tracks confidence dynamics over answer-relevant positions to relay decoding trajectories between models, achieving strong performance on diverse reasoning tasks.
First Wasserstein-2 Convergence Proof for Decentralized Diffusion Models with ODE Samplers
A team of researchers has proven the first convergence guarantee in Wasserstein-2 distance for ODE-based samplers in decentralized diffusion models. The work addresses the missing theoretical foundation for decentralized generative architectures that replace a single global velocity field with multiple local experts and a routing mechanism. The result shows distribution converges at rate O(N^{-1/2}+ε), paving the way for privacy-scalable AI deployments.
DifFRACT Brings Circuit Tracing to Diffusion Transformers for Better AI Interpretability
Researchers introduce DifFRACT, a method for mechanistic interpretability of multimodal diffusion transformers. By training timestep-conditioned transcoders on FLUX.1[schnell], they achieve exact feature-to-feature attribution and recover compact circuits, outperforming sparse autoencoders in precision.