Image generation models trained on large datasets can synthesize high-quality images but often produce spatially inconsistent and distorted images due to limited information about the underlying structures and spatial layouts. In this work, we leverage intrinsic scene properties (e.g., depth, segmentation maps) that provide rich information about the underlying scene, unlike prior approaches that solely rely on image-text pairs or use intrinsics as conditional inputs. Our approach aims to co-generate both images and their corresponding intrinsics, enabling the model to implicitly capture the underlying scene structure and generate more spatially consistent and realistic images. Specifically, we first extract rich intrinsic scene properties from a large image dataset with pre-trained estimators, eliminating the need for additional scene information or explicit 3D representations. We then aggregate various intrinsic scene properties into a single latent variable using an autoencoder. Building upon pre-trained large-scale Latent Diffusion Models (LDMs), our method simultaneously denoises the image and intrinsic domains by carefully sharing mutual information so that the image and intrinsic reflect each other without degrading image quality. Experimental results demonstrate that our method corrects spatial inconsistencies and produces a more natural layout of scenes while maintaining the fidelity and textual alignment of the base model (e.g., Stable Diffusion).

Recent advances in neural networks have enabled agents to perform high-level reasoning in complex environments, where robust environmental understanding must often be obtained from accumulated observations rather than from fixed scene labels. Existing spatial abstractions commonly rely on scene-level representations, object-presence cues, scene graphs, or language-grounded category prompts; depending on the task, these abstractions can either discard distributional object evidence or introduce additional structure that is not directly grounded in the agent’s own experience. To address this information abstraction imbalance, we design an object-centric spatial abstraction (OSA) that models environments as empirical distributions over object embeddings. Building on this abstraction, we present PEU-OSA, a Probabilistic Environment Understanding framework via Object-centric Spatial Abstraction. PEU-OSA comprises three conceptual modules: an object perception module, the OSA module, and a memory component. Scene observations are transformed into object representations, abstracted into empirical distributions through OSA, and accumulated in memory for comparison with new observations. The OSA module captures three fundamental relationships in the latent space via kernel density estimation: object-object, object-environment, and environment-environment relationships. To characterize when these measurements are reliable, we propose the ( )-Statistically Separable (EDS) function, which quantifies the separability and concentration of object representations and provides theoretical conditions and empirical evidence for OSA estimates. We evaluate PEU-OSA through EDS-based analysis on ImageNet, scene classification on Places, chained inference in Replica, and applicability studies for embodied agents in Replica and Minecraft. In a Replica chained-inference setting designed to test generalization to unseen environments, PEU-OSA improves top-1 retrieval accuracy from 0.36 with CLIP-based vision-language retrieval to 0.78, suggesting that object-distribution memory can provide a useful grounding source when scene-level language retrieval is insufficient. These findings support PEU-OSA as an annotation-free, experience-grounded framework for probabilistic environment understanding across recognition, reasoning, and embodied memory scenarios.

Causal dynamics learning has recently emerged as a promising approach to enhancing robustness in reinforcement learning (RL). Typically, the goal is to build a dynamics model that makes predictions based on the causal relationships among the entities. Despite the fact that causal connections often manifest only under certain contexts, existing approaches overlook such fine-grained relationships and lack a detailed understanding of the dynamics. In this work, we propose a novel dynamics model that infers fine-grained causal structures and employs them for prediction, leading to improved robustness in RL. The key idea is to jointly learn the dynamics model with a discrete latent variable that quantizes the state-action space into subgroups. This leads to recognizing meaningful context that displays sparse dependencies, where causal structures are learned for each subgroup throughout the training. Experimental results demonstrate the robustness of our method to unseen states and locally spurious correlations in downstream tasks where fine-grained causal reasoning is crucial. We further illustrate the effectiveness of our subgroup-based approach with quantization in discovering fine-grained causal relationships compared to prior methods.