Tag
Vathography is a new creative approach that turns a photograph's inferred depth map into an artwork, using a diffusion transformer to estimate depth and letting authors treat focus as a mapping of depth to tone and color. Inspired by Marigold V2's detailed depth decomposition, it reframes a normally hidden technical pass as a form of artistic expression.
HairDeck 提出使用集合扩散 transformer 从单张图像生成风格化发卡与弹簧骨动画，将 CHARM 的 Chamfer 距离降低 44%，速度提升 33 倍，可直接导出为可用的下游资产。
OrderFlowDiT提出了一种带非平衡最优传输正则的条件扩散Transformer，用于预测未来五分钟限价订单簿的成交量-价格分布，相比MSE训练将全字段传输成本降低45.8%。
MOSS-Image is an open 5B diffusion transformer combining high-fidelity image synthesis with intent-aware prompt enhancement for bilingual long-form text rendering and layout-coherent generation, achieving state-of-the-art open-weight results on the new LongLongTextBench and RealLayoutArena benchmarks.
AcrossVAM2.0 提出一种冲突感知的点对应监督方法，用于参数高效的视频扩散模型适配，通过置信度加权跟踪头、零初始化残差反馈和非对称 PCGrad，在保持预训练外观先验的同时提升对应一致性，但并不均匀改善视频质量。
OmniXRef 提出一个统一的多模态多参考视频编辑框架，支持结合文本、图像和视频参考来控制源视频的多种属性，并配套发布了模型、训练数据集和评测基准 OmniXRef-Bench。
Dream4ACT introduces a shared visual action interface ("action views") rendered via URDF-based forward kinematics to enable joint video-action modeling across multiple robot embodiments, combining observations and actions in a single diffusion world model with masked flow matching. It achieves 88.98% success on RoboTwin 2.0 and 65.66 on TriWorldBench using a training-free URDF-constrained recovery mechanism for executable actions.
FlowTool is a framework that models tool-based image editing as a flow matching problem, achieving superior performance and efficiency compared to autoregressive multimodal models.
Meta AI introduces Muse Realtime Avatar, a state-of-the-art embodiment technology that turns voice into synchronized, expressive avatars in real time using a streaming Diffusion Transformer.
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.
Marigold V2 transforms an image-editing Diffusion Transformer into a single-step model for sharp dense prediction, achieving best zero-shot results in depth estimation and related tasks with Apache 2.0 license.
LynnReal-Omni is a unified multimodal video diffusion framework that integrates agentic visual controls with high-fidelity generation and real-time acceleration for stable, controllable video creation.
A detailed write-up on training a 210M text-to-image diffusion transformer from scratch on a single GPU, sharing key measurements on attention sinks, loss as a health signal, and timestep shifting benefits.
StepAudio 3 Music introduces a large-scale, long-form music generation model with explicit musical planning via ABC-CoT, achieving high scores in audio quality and similarity metrics compared to other systems.
UniMate is a unified diffusion transformer model that generates articulated motion for diverse skeletons from text and rigged 3D assets without per-skeleton retraining, using topology-aware attention and a large curated dataset.
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.
LLaDA-Image presents a unified framework that combines a 6B diffusion transformer with a frozen vision-language module for generating photorealistic images with precise editing, achieving state-of-the-art results among open-source models through efficient training and fast inference.
World Labs introduces Atlas, a multimodal world model that can generate, reconstruct, and simulate 3D environments with camera control and spatial consistency.
Wan-Animate-2 is a new end-to-end character animation framework that consumes driving videos directly in a redesigned Diffusion Transformer, achieving high-fidelity motion generation and identity preservation. It also introduces a lightweight variant for real-time streaming animation, with open-source weights released.
AtlasVLA introduces a dual-memory architecture with persistent world-ego state modeling to overcome perception and task-progress forgetting in vision-language-action models, achieving state-of-the-art long-horizon manipulation from a single wrist camera.