absorb.md

jasmine sun

Chronological feed of everything captured from jasmine sun.

paper / jasminewsun / Jun 9

Bi-S Network Governs Stability and Optoelectronics in Disordered AgBiS2

This work reveals the critical role of the three-dimensional Bi-S network in stabilizing the cation-disordered AgBiS2 phase and dictating its optoelectronic properties. Analysis combining machine learning potentials and deep learning Hamiltonians shows that this network preserves dispersive band edges and a direct band gap, essential for semiconductor applications. The findings clarify structural ambiguities and link disorder to material performance in lead-free alloys.

materials-sciencecomputational-physicsoptopyroelectricslead-free-materialssemiconductor-alloysmachine-learning
“The Bi-S network is the primary structural element that stabilizes the cation-disordered phase of AgBiS2.”
paper / jasminewsun / Jun 9

SemDINO: A DINOv3-Enhanced Network for Robust Cross-Temporal Semantic Change Detection

SemDINO is a novel end-to-end network designed for semantic change detection, addressing limitations in cross-temporal alignment, multi-scale representation, and robustness to pseudo-changes. It features a dual-branch encoder fusing CNN and frozen DINOv3 features, a multi-scale temporal transformer for alignment, and specialized modules for enhancing genuine changes while suppressing pseudo-variations caused by illumination or seasonal shifts. Experimental results show SemDINO outperforms existing methods on public remote sensing datasets, demonstrating strong generalization.

change-detectionremote-sensingsemantic-segmentationtransformer-modelsmulti-modal-fusioncomputer-visiondino
“SemDINO integrates a dual-branch encoder, multi-scale temporal interaction, semantic purification, change enhancement, and decoupled multi-task prediction into a unified framework to address challenges in semantic change detection.”
paper / jasminewsun / May 31

Future-Aware KV Cache Compression for AR Video Generation via Pre-RoPE Query Stationarity

Autoregressive video generation models suffer from KV cache bloat at long horizons, and existing compression methods fail because they score token importance using only short-horizon, historically-grounded signals — discarding tokens that become critical for future frames. This paper identifies that canonical pre-RoPE query distributions remain approximately stationary across autoregressive steps, meaning future query behavior can be estimated from past statistics without any retraining. Leveraging this, the authors propose Future Forcing: a training-free policy that constructs a future query proxy, scores cached tokens against it, and merges redundant token pairs within the induced affine subspace. On VBench-Long 60-second generation benchmarks, the method achieves up to 1.49× improvement in subject consistency over existing AR video KV cache policies.

video-generationkv-cacheautoregressive-modelsinference-optimizationcomputer-visiontransformer-efficiencylong-horizon-generation
“Pre-RoPE query distributions in trained AR video models remain approximately stationary across autoregressive generation steps, despite RoPE-modulated queries evolving at each step.”
paper / jasminewsun / May 31

OmniCD: Multimodal Semantic Guidance Enables Generalizable Remote Sensing Change Detection

OmniCD is a unified framework for remote sensing change detection (CD) that incorporates multimodal inputs — image prompts, text descriptions, semantic maps, and geospatial metadata — into a single architecture capable of handling tasks from binary CD to zero-shot semantic change understanding. It addresses the core limitation of traditional CD methods: poor generalization across diverse scenarios and domains. The framework pairs a hierarchical scene retrieval module with a style disentanglement mechanism for cross-domain robustness, and is supported by RSITCD, a newly introduced large-scale dataset of 300K+ annotated image-text pairs. Experiments report state-of-the-art performance across benchmarks.

remote-sensingchange-detectionmultimodal-learningcomputer-visionzero-shot-learninggeospatial-aiimage-text-pairs
“OmniCD achieves state-of-the-art performance across remote sensing change detection benchmarks.”
paper / jasminewsun / May 31

minWM: Open-Source Pipeline for Converting Video Diffusion Models into Real-Time Interactive World Models

minWM is a full-stack, open-source framework that transforms existing bidirectional text-to-video (T2V) and text-image-to-video (TI2V) diffusion foundation models into camera-controllable, few-step autoregressive world models capable of real-time interactive rollout. The pipeline covers the full lifecycle — data construction, controllable fine-tuning, AR diffusion training, causal ODE/consistency distillation, and streaming inference — addressing the key bottlenecks of latency and causal control. The framework is architecture-agnostic, demonstrated on both cross-attention-based (Wan2.1-T2V-1.3B) and MMDiT-style (HY1.5-TI2V-8B) backbones, and supports adaptation of existing world models like HY-WorldPlay to new distributions and latency targets. Code, checkpoints, and documentation are fully released.

video-generationworld-modelsdiffusion-modelsopen-sourcereal-time-inferencecomputer-visionautoregressive-models
“Existing video diffusion foundation models can be converted into real-time interactive world models via a structured distillation pipeline (Causal Forcing / Causal Forcing++) without training from scratch.”