Donghyeon Cho, Joowon Kim, Seungho Shin, Eunho Yang
Diffusion vision-language models (DVLMs) process their entire input at every denoising step. Since visual tokens make up 89-97% of this input on the benchmarks we study, most of the decoding computation is spent repeatedly processing the same image. Visual token pruning methods partially relieve this burden by removing part of the visual tokens, but they lose the removed information and still process the remaining image at every step. To measure how often the image is actually needed, we compare LLaDA-V's predictions with and without the image at every decoding step and find that removing the image changes a token being unmasked at only 22.8% of steps on average and at as few as 7.9%. We therefore propose LOWN, a decoding method that decides at every step whether to process the image and how many tokens to unmask. LOWN predicts the masked tokens without the image and processes the image only when this prediction disagrees with the last image-conditioned prediction. Agreement certifies the earlier prediction for reuse, and the number of certified tokens sets how many are unmasked in that step. We further observe that the need for the image differs across benchmarks, so LOWN's adaptive decision is needed rather than a fixed schedule. Without training or a tuned threshold, LOWN retains 99.6% of LLaDA-V's score on ten benchmarks at 35.3% of the decoding FLOPs on average. LOWN also generalizes across LLaDA-based DVLMs, composes orthogonally with visual token pruning, and runs under key-value caching for further efficiency gains. Code and project page are available at https://hyeon9698.github.io/LOWN/
Joonhyung Park Korea Advanced Institute of Science & Technology, Jaeyun Song KRAFTON Inc., Sihwan Park Korea Advanced Institute of Science & Technology, Eunho Yang Korea Advanced Institute of Science & Technology
Recent proprietary video generation models have demonstrated remarkable proficiency in synthesizing highly realistic videos from textual instructions. Most open-source text-to-video models, however, still struggle to accurately simulate real-world physics and dynamic entity interactions. Existing approaches rely on scaling laws and large-scale, high-quality video datasets to implicitly learn physical dynamics, yet this paradigm is constrained by prohibitive costs and the burdensome demands of data curation. Motivated by this, we propose a novel framework that integrates graph-structured temporal knowledge into video latent diffusion models to enhance compositional generation and interaction fidelity. Our framework constructs video scene graphs specifically designed to capture entity relationships, temporal dynamics, and global scene context. These graph-structured representations guide the generation process through cross-attention mechanisms. Additionally, we introduce Graph-Aligned Denoising Loss (GADL), a training objective that ensures adherence to conditioned graphs by incorporating node modification tasks within the denoising process, leveraging synchronized edited video-graph pairs. Comprehensive evaluations demonstrate that incorporating graph-structured knowledge significantly enhances compositionality and the accurate portrayal of real-world interactions in generated videos.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表