本文贡献:SlotDiT 让文本条件 DiT 直接在 slot 潜空间里工作:先把场景分解成一组对象级的 slot,再依据指令与已观测上下文自回归去噪未来的 slot 轨迹。论文的落点是系统性比较潜空间设计——在统一 DiT 框架下把 slot 表示与 VAE 式、语义对齐式潜空间放在一起对照。
% Overview of SlotDiT. % (a) Given a reference image $\ImageT{1}$ and text instruction $\Caption$, SlotDiT parses the scene into an object-centric slot representation $\SlotsT{1}$. % Conditioned on the encoded instruction and context slots, an object-centric diffusion transformer (DiT) autoregressively denoises future slot trajectories $\PredSlotsT{2}, \ldots, \PredSlotsT{T+1}$, % optionally decoded into future video frames $\PredImageT{2},\, \ldots,\, \PredImageT{T+1}$. % % (b) Comparison of SlotDiT's slot-based latent space against established representation spaces used in DiTs. % % (c) Across four robotic datasets, SlotDiT achieves higher task-completion rates while remaining substantially more efficient than the baselines.
% Overview of SlotDiT. % (a) Given a reference image $\ImageT{1}$ and text instruction $\Caption$, SlotDiT parses the scene into an object-centric slot representation $\SlotsT{1}$. % Conditioned on the encoded instruction and context slots, an object-centric diffusion transformer (DiT) autoregressively denoises future slot trajectories $\PredSlotsT{2}, \ldots, \PredSlotsT{T+1}$, % optionally decoded into future video frames $\PredImageT{2},\, \ldots,\, \PredImageT{T+1}$. % % (b) Comparison of SlotDiT's slot-based latent space against established representation spaces used in DiTs. % % (c) Across four robotic datasets, SlotDiT achieves higher task-completion rates while remaining substantially more efficient than the baselines.
AlayaVista: Streaming World Modeling from Panoramic States to Perspective Video | Alaya Lab;Beijing Institute of Technology;The University of Tokyo | arXiv:2609.14462
Overview of MDN-Control. Mask-guided localization localizes the target, depth-aware occlusion control guides editing near occlusions, and noise latent prompting initializes appearance generation. Denoising uses mask-depth guidance at early steps and mask-only guidance thereafter.
Efficient Text-to-Image Generation: An Adaptive Step Schedule Controller for Diffusion Models | SAP;National University of Singapore;A*STAR I2R | arXiv:2609.16572
Effect of text prompt and seed combinations on diffusion steps for generating visually sufficient images. Step counts shown are examples; optimal steps may vary.
The denoising process begins with the longest schedule (e.g. 40 steps in the example) and transitions to shorter schedules (20 followed by 10) upon detecting convergence.
The overall architecture of the proposed DiTAR+ framework. (a) The standard two-stage DiTAR baseline. (b) Hierarchical Acoustic Masking (HAM), which masks historical acoustic context in shallow layers to decouple semantic alignment and acoustic rendering. (c) Dilated Context Sampling (DCS), which expands the macro-receptive field via dilated sampling while preserving temporal continuity.
Chunk-level speaker similarity over continuous generation steps (3.0,s per chunk). The standard baseline experiences severe speaker drift in the long tail, whereas DCS effectively stabilizes the temporal coherence.
VOR-Bench: A Human Perception-Driven Benchmark for Video Object Removal | Beijing University of Posts and Telecommunications;China Telecom AI Technology;Xi'an Jiaotong University | arXiv:2609.16878
评论 (0)