The paper introduces the Graded Color Attribution (GCA) dataset, a benchmark that tests whether Vision‑Language Models (VLMs) and humans can articulate and follow a threshold rule for labeling objects by color. In experiments, humans consistently adhere to their stated rules, while VLMs—despite accurately estimating color coverage—often violate their own introspective rules, especially when world‑knowledge priors are present. This discrepancy highlights a miscalibration in VLM self‑knowledge that differs from human cognition.
By Jonathan Nemitz, Carsten Eickhoff, Junyi Jessy Li, Kyle Mahowald, Michal Golovanevsky, William Rudman
arXiv:2609.14495v1 Announce Type: new
Abstract: Image colorization is an inherently ill-posed task, since a single grayscale image may correspond to multiple plausible colorized results. Consequently...
By Yunkai Zhuang, Qihang Yan, Zicheng Zhang, Guangtao Zhai
Text-to-image diffusion models have achieved high visual fidelity and broad adoption, but remain vulnerable to safety violations when adversaries exploit them to synthesize illicit content. Existing alignment paradigms, from input sanitization to structural feature pruning, are largely organized around unsafe concepts explicitly exposed during filtering, editing, or localization.
arXiv:2606. 28386v1 Announce Type: cross Abstract: Image autoregressive models (IARs) have recently demonstrated remarkable capabilities in visual content generation, achieving photorealistic quality and rapid synthesis through the next-token prediction paradigm adapted from large language models.
By Bihe Zhao, Louis Kerner, Michel Meintz, Tameem Bakr, Franziska Boenisch, Adam Dziedzic
arXiv:2602.08059v2 Announce Type: replace-cross
Abstract: Text-to-image diffusion models can reproduce specific artists visual styles at extremely low cost, raising copyright and deployment safety co...
By Tong Zhang, Ru Zhang, Jianyi Liu
SafeStyle is a training‑free framework that injects calibrated style residuals into frozen diffusion models for reference‑guided stylization. It estimates style‑supported and content‑associated subspaces from small calibration sets, then transports purified style evidence across adaptive spatial granularity while limiting its influence with a residual‑norm budget. Experiments on texture‑ and geometry‑dominant styles show high DINO style similarity (0.432–0.474) with minimal semantic leakage (0.8%).
By Zhangping Yang, Min Li, Song Yan, Rong Gao, Xinliang Bi, Guanye Xiong, Yujie He
arXiv:2608. 10798v1 Announce Type: cross Abstract: Most image colorization systems operate in $Lab$ space by predicting chroma ($ab$) while preserving an input-derived luminance channel ($L$).
By Swarnim Maheshwari, Syed Imam Ali, Vineeth N. Balasubramanian
arXiv:2606. 09909v1 Announce Type: cross Abstract: With the growing concerns over copyright infringement in diffusion-based customization, adversarial attacks have emerged as a prominent defense strategy to prevent malicious content forgery in personalized image generation.
By Ziang Xu, Wenbo Yu, Hongyao Yu, Hao Fang, Jiawei Kong, Bin Chen, Hao Wu, Shu-Tao Xia, Zhiyong Wu
The paper introduces COMIC, a reference‑aware safety gate designed for multimodal large language models (MLLMs). COMIC detects the operation requested by a user, identifies visual targets through OCR and open‑vocabulary proposals, and evaluates safety on explicit operation‑target pairs, using max‑risk aggregation and quality‑aware routing to decide whether to allow or block a request. Experiments on several open‑source MLLMs and jailbreak benchmarks show that COMIC improves robustness while maintaining benign utility and efficiency.
By Md Abdullahil Oaphy, Anhao Xiang, Zongxing Xie, Huayue Gu, Chenyu Wang, Honghui Xu
arXiv:2607. 24101v1 Announce Type: cross Abstract: Concept unlearning is increasingly used to limit the reproduction of protected or unsafe visual concepts in text-to-image models.
By Keyu Li, Jin Gao, Jialing Zhang, Dequan Wang
arXiv:2606. 25375v2 Announce Type: replace-cross Abstract: With the rapid adoption of generative AI, synthetic medical images pose growing risks, including diagnostic deception and insurance fraud.
By Ching-Hao Chiu, Hao-Wei Chung, Gelei Xu, Xueyang Li, Pin-Yu Chen, John Kheir, Meysam Ghaffari, Carlos Morato, Ahmed Abbasi, Yiyu Shi
The paper introduces the concept of watermark laundering, where an attacker uses a single reconstruction prompt on public foundation image models to produce a visually faithful output that renders invisible watermarks undecodable. The authors evaluate this failure mode across six OpenAI and Google image editing models, three watermarking schemes, and 1,800 reconstructions, finding that OpenAI models cause the strongest payload disruption while Nano Banana 2 shows vulnerability of DwtDct under high-fidelity reconstruction. Prompt ablation experiments reveal that the disruption is driven by the reconstruction pathway itself rather than any specific removal instruction, highlighting prompt-conditioned reconstruction as a distinct attack interface.
By Jidong Yang, Qi Li, Wei Zong, Yang-Wai Chow, Willy Susilo, Huaike Yu, Chunpeng Wang, Suo Gao