generative-model

Duke researchers present StormGen, a generative ML model that completes short observed rainfall records into full, physically plausible storm lifecycles, reproducing rare extremes like Hurricane Harvey and outperforming stochastic storm transposition for stress-testing flood defenses.

BigCell是首个能在原生千兆像素分辨率下生成连贯全切片病理图像的方法，采用流匹配Transformer在2048×2048像素上训练超过12万张公开WSI，可生成高达32,768×32,768像素的图像，并通过合成样本将少样本切片级分类性能提升最多18%。

JoHan：一个从视频序列直接联合生成2D和3D手部姿态序列的统一生成框架，无需依赖逐帧姿态预测，在遮挡和不完整观测下显著提升手部运动重建的精度、速度和平滑度。

This paper presents REALM, a coarse-to-fine generative framework for audio-driven reactive listening in embodied AI, addressing challenges in timing and facial motion with evaluations and robot deployment showing improvements.

GeoCR introduces a generalist cloud removal model that unifies RGB-only and multispectral cloud removal (with optional SAR guidance) in a single flow transformer, trained on ten datasets with 883,331 cloud-free images, achieving state-of-the-art FID/DISTS without dataset-specific fine-tuning.

The paper proposes CorrFlow, a correlation-guided flow matching framework for predicting spatial transcriptomics from histology images, explicitly modeling gene-gene dependencies to improve biological coherence in generated profiles.

The paper introduces SolarFlowRefiner, a refinement-aware flow-matching framework for downscaling surface solar radiation from coarse ERA5 data to high-resolution SolarCube fields, demonstrating consistent improvements over standalone generation and post-hoc refinement methods.

OneBid is a unified auto-bidding foundation model that addresses challenges in diverse oCPX advertising scenarios by using Mixture-of-Experts architecture and CROP optimization, with deployment at Kuaishou showing significant improvements like +13.1% in ROAS.

BoundInk introduces a writer-conditioned framework for online handwriting generation that explicitly models inter-character boundaries to improve cursive continuity and spatial alignment, outperforming existing methods with significant reductions in error and high human preference.

Mira-Scene introduces a compositional 3D scene reconstruction framework using pixel-aligned canonical coordinate maps for accurate object layouts, achieving significant improvements in layout accuracy over existing methods.

This paper proposes an Affect-Prototype-Conditioned Fusion (APCF) framework for open-vocabulary multimodal emotion recognition with incomplete modalities, using an affect-prototype library to guide feature fusion and an LLM decoder for generating natural language emotion labels.

This paper evaluates MiniMax-H3, an omni-modal generative model, by introducing a comprehensive framework to assess its reasoning about the physical world through multimodal inputs. The evaluation reveals that video-based decision reasoning performs best, while audio-based disambiguation reasoning is the weakest.

Fraglingo introduces an attachment-aware autoregressive model for molecular design that generates molecules by jointly predicting fragment identity and attachment in a continuous latent space, enhancing property control and flexibility.

KaiNinja extends a native 3D generator to produce part-level outputs using a dual-volume representation, improving both part and whole-object fidelity without requiring segmentation.

This paper proposes a conditional generative latent framework to encode connectome graphs, enabling reconstruction, generation, and functional analysis in reservoir computing, with insights into task-specific structural mechanisms.

Introduces Generative Late-Interaction Embeddings (GLIE) for compressing visual document retrieval vectors, improving accuracy under storage constraints by regenerating full embeddings on demand.

SPAR3S is a sparse auto-regressive generative model that efficiently generates complete 3D scenes from sparse multi-view images without requiring 3D ground truth data, using a voxel-aligned latent space and differentiable 3D Gaussian Splatting.

WorldSculpt is a framework that reconstructs cluttered scenes from grounded videos into compositional per-object meshes, enabling interactive 3D world generation and editing.

SimpleDesign introduces a joint model for protein sequence and structure codesign, trained end-to-end in data space using a single-stage objective, achieving competitive performance on co-design and generation benchmarks.

OutageDiT is a generative foundation model for power outage forecasting that uses a Diffusion Transformer architecture to generate seven-day outage trajectories, improving forecast accuracy and enabling zero-shot transfer to new regions.