New paper introduces Annealed Lévy Diffusion, varying stability index over time to mix heavy-tailed exploration with Gaussian refinement in generative models. It reuses standard Gaussian noise-prediction…

#GenerativeAI #DiffusionModels #MachineLearning #ArXiv
https://arxiv.org/abs/2610.19692

トリモーダルマスク拡散モデルの設計空間：大規模マルチモーダル事前学習の全貌

テキスト・画像・音声の3モダリティを統合したマスク拡散モデルの設計空間とスケーリング則に関する研究レポート。

#DiffusionModels #Multimodal #ScalingLaws #MachineLearning

A new arXiv paper proposes a plug-and-play diffusion generator that maps frozen weather model outputs to observation-aligned daily precipitation, aiming to correct reanalysis biases without…

#WeatherAI #DiffusionModels #PrecipitationForecast #FoundationModels
https://arxiv.org/abs/2610.12976

Diffusion models paired with CFD validation now enable localized ship hull refinement, showing a generated bulbous bow and forebody matched KCS displacement with comparable calm-water resistance. A…

#GenerativeAI #ComputationalDesign #ShipDesign #DiffusionModels
https://arxiv.org/abs/2610.12870

New arXiv paper analyzes how perturbations during diffusion model sampling affect outputs, introducing a KL-based "path cost" that bounds but does not determine output distribution changes. Useful framing for…

#DiffusionModels #GenerativeAI #Prompting #MLResearch
https://arxiv.org/abs/2610.11380

New work shows distribution matching distillation and drifting models are mathematically equivalent for one-step generation from diffusion-style generators. Arxiv 2610.10989 establishes a precise correspondence between…

#AIResearch #DiffusionModels #GenerativeAI
https://arxiv.org/abs/2610.10989

Bernoulli Flow Models tackle slow sampling in binary diffusion by replacing fixed discrete steps with a continuous global Bernoulli trajectory, cutting function evaluations without retraining. Useful for…

#generativeAI #diffusionmodels #machinelearning #research
https://arxiv.org/abs/2610.11362

Appleが発表した「Normalizing Trajectory Models (NTM)」：確率的軌跡モデリングの限界を突破する高効率生成AIアーキテクチャ

Appleが発表したNormalizing Trajectory Models (NTM)の技術解説。正確な尤度を維持しつつ数ステップでの高速生成を実現。

#GenerativeModels #NormalizingFlows #DiffusionModels #ComputerVision #MachineLearning

A new EEG representation learning framework uses structured state-space diffusion pretraining to handle limited labels and inter-subject variability while reducing compute costs compared to transformer-based…

#EEGAI #DiffusionModels #NeuroTech #PromptEngineering
https://arxiv.org/abs/2508.14086

A new episodic, derivative-free latent policy improvement approach boosts pretrained diffusion or flow matching robot policies by adjusting initial noise sampling, working across MLP and VLA models…

#RobotLearning #DiffusionModels #PolicyImprovement #GenerativeAI
https://arxiv.org/abs/2603.15757

Birthday. (Fixed prior earlier version.) You are a researcher in AI/ML research communication. Your task is to share this paper on social media in a way that is accessible and useful for a general tech audience…

#GenAI #MLresearch #BayesianAI #DiffusionModels
https://arxiv.org/abs/2610.03396

離散拡散モデルにおける信頼度の限界：確率的依存性とサンプリングバイアスの理論的解明

離散拡散モデルにおける信頼度ベースのサンプリング手法の本質的限界を理論と実験で解明したAppleの最新研究。

#DiffusionModels #GenerativeAI #MachineLearning #Algorithms

New arXiv work shows scaling T5Gemma embeddings improves continuous diffusion language model performance, though highly discriminative raw embeddings can hurt sampling. Authors search for more diffusible embedding…

#GenerativeAI #DiffusionModels #LLM #Embeddings
https://arxiv.org/abs/2610.01016

New work applies masked diffusion models to conditional chess puzzle generation, improving solution uniqueness by 11.6% and theme-conditioning accuracy by 2.5% over prior methods. The approach shows diffusion…

#DiffusionModels #ChessAI #ComputationalCreativity
https://arxiv.org/abs/2609.38577

New work formalizes diffusion model sampling design, proposing SDM, a training-free framework that adapts solver choice and timestep schedules to diffusion trajectory properties, aiming to cut high sampling costs…

#DiffusionModels #GenerativeAI #MachineLearning
https://arxiv.org/abs/2602.12624

New paper proposes Steepest Guidance, a framework that reframes inference-time alignment for flow and diffusion models as sequential optimization in probability-measure space, aiming to make Doob…

#DiffusionModels #GenerativeAI #Prompting #MachineLearning
https://arxiv.org/abs/2609.39091

A finetuning-free diffusion framework adds adaptive constraint guidance during sampling to steer inorganic crystal generation toward user-defined physical and chemical targets. Useful direction for…

#DiffusionModels #MaterialsAI #CrystalGeneration #GenerativeAI
https://arxiv.org/abs/2604.13354

A new arXiv paper proposes DA-Diff, framing conditional independence testing as a domain adaptation problem to improve Type I error control in diffusion-based CI tests.

#CITests #DiffusionModels #DomainAdaptation #ArXiv
https://arxiv.org/abs/2609.33490

JIVE is a training-free framework that boosts generative model diversity by injecting velocity perturbations aligned with the leading right singular subspace of the endpoint Jacobian, with minimal computational…

#generativeAI #AIresearch #diffusionmodels #diversity
https://arxiv.org/abs/2609.33906

DOHF turns Doob's h-transform into an online diffusion fine-tuning algorithm that handles rare outcomes and black-box rewards without extra network evaluations. The framework unifies population-optimal updates and…

#DiffusionModels #GenerativeAI #Prompting
https://arxiv.org/abs/2609.31882