arXiv:2506.09740v2 Announce Type: replace Abstract: Diffusion models excel at image generation. Recent studies have shown that these models not only generate high-quality images but also encode text-image alignment information through attention maps or loss functions. This information is valuable for various downstream tasks, including segmentation, text-guided image editing, and compositional image generation. However, current methods heavily rely on the assumption of perfect text-image alignment in diffusion models, which is not the case. In this paper, we propose using zero-shot referring image segmentation as a proxy task to evaluate the pixel-level image and class-level text alignment of popular diffusion models. We conduct an in-depth analysis of pixel-text misalignment in diffusion models from the perspective of training data bias. We find that misalignment occurs in images with small-sized, occluded, or rare object classes. Therefore, we propose ELBO-T2IAlign--a simple yet effective method to calibrate pixel-text alignment in diffusion models based on the evidence lower bound (ELBO) of likelihood. ELBO-T2IAlign is training-free and generic: it requires no additional annotations, model retraining, or architectural modifications, and it can be directly applied to different diffusion backbones. Extensive experiments on zero-shot referring image segmentation, text-guided image editing, and compositional image generation verify that the proposed calibration improves pixel-text alignment across complementary downstream tasks.
Science Journals
arXiv:2607.04255v1 Announce Type: new Abstract: With the development of low altitude intelligent systems, multiple unmanned aerial vehicles (UAVs) can collaboratively execute more complex tasks. Conventional task allocation methods usually regard tasks and UAVs as isolated entities, making it difficult to capture task dependencies and UAV communication relationships. To address this issue, this paper proposes a dual heterogeneous graph learning based UAV task allocation method. A directed task graph is constructed to represent task dependencies and encode task resource requirements, while an undirected UAV communication graph is built to model communication relationships and encode UAV resource states. The task allocation problem is formulated as a structural matching problem between the task graph and the UAV communication graph. A graph attention network based feature extraction method is introduced to learn structural representations from both graphs through message passing. A cross attention mechanism is further integrated with proximal policy optimization to optimize the matching between task nodes and UAV nodes for task allocation. Simulation results demonstrate that the proposed method achieves a higher task completion rate and shorter task completion time than benchmark methods under different evaluation settings. Furthermore, a UAV sensing and computing application is developed on the AirSim simulation platform. A large language model is employed to convert natural language task requirements into a structured task graph for autonomous UAV task execution, demonstrating the potential of the proposed framework for natural language driven UAV mission planning and execution.
arXiv:2607.04282v1 Announce Type: new Abstract: Classic models cast information seeking as iterative foraging: formulate a keyword query, scan results, reformulate, gather across sources, synthesize. We ask what happens when a conversational assistant is inserted into that episode. Linking real conversations with major assistants to the same users' searches and browsing in an opt-in cross-surface panel, and reconstructing the full episode rather than a single query, we find conversational AI changes the shape of information seeking, not merely its volume. AI episodes do not uniformly collapse; they bifurcate. Most terminate in place, with no onward search or content step in the observed trace, while roughly a third scaffold into longer multi-step journeys. Which shape occurs is governed less by task type than by articulation: collapse is statistically indistinguishable across lookup, learning, and comparison episodes, yet falls monotonically with opening-ask length, from 72% at one-to-three words to 48% beyond twenty. Roughly two-fifths of assistant episodes are workbench use--drafting, coding, editing--not information seeking at all, and these collapse most. Conversational AI also does not displace search: search remains woven through roughly three-quarters of within-episode transitions, after reading a page users return to the search box over the assistant 70/30, and within-user search share does not fall. Verification is rare: searches with explicit verification language follow roughly 1% of episodes, and citation-forward interfaces do not measurably increase checking. All of this is episode structure, a compositional object identifiable without a demand counterfactual. Conversational AI recomposes the seeking episode: it answers brief asks in place and anchors invested asks in longer journeys, adding a layer rather than replacing search.
arXiv:2506.13770v2 Announce Type: replace Abstract: We introduce Color Disentangled Style Transfer (CDST), a novel and efficient two-stream style transfer training paradigm which completely isolates color from style and forces the style stream to be color-blinded. With one same model, CDST unlocks universal style transfer capabilities in a tuning-free manner during inference. Especially, the characteristics-preserved style transfer with style and content references is solved in the tuning-free way for the first time. CDST significantly improves the style similarity by multi-feature image embeddings compression and preserves strong editing capability via our new CDST style definition inspired by Diffusion UNet disentanglement law. By conducting thorough qualitative and quantitative experiments and human evaluations, we demonstrate that CDST achieves state-of-the-art results on various style transfer tasks.
arXiv:2507.04221v3 Announce Type: replace Abstract: We introduce Context Tuning, a simple and effective method to significantly enhance few-shot adaptation of large language models (LLMs) without weight updates. In-Context Learning (ICL) forms a memory representation of the demonstrations in a single forward pass but cannot refine it when insufficient. Prompt-based methods offer lightweight adaptation by optimizing a trainable prompt or prefix but initialize it independently of the demonstrations. In contrast, Context Tuning leverages the model's inherent ICL ability to initialize a trainable memory representation from demonstrations, then refines it through gradient-based optimization. Extensive evaluations on benchmarks such as CrossFit, UnifiedQA, MMLU, BIG-Bench Hard, and ARC demonstrate that Context Tuning outperforms both ICL and traditional prompt-based adaptation methods while achieving competitive accuracy with Test-Time Training at significantly higher training efficiency.
arXiv:2508.02296v3 Announce Type: replace Abstract: Retrieval-Augmented Generation (RAG) systems are increasingly deployed in high-stakes domains, where safety depends not only on how a system answers, but also on whether a query should be answered given a knowledge base (KB). Out-of-domain (OOD) queries can cause dense retrieval to surface weakly related context and lead the generator to produce fluent but unjustified responses. We study lightweight, KB-aligned OOD detection as an always-on gate for RAG systems. Our approach applies PCA to KB embeddings and scores queries in a compact subspace selected either by explained-variance retention (EVR) or by a separability-driven t-test ranking. We evaluate geometric semantic-search rules and lightweight classifiers across 16 domains, including high-stakes COVID-19 and Substance Use KBs, and stress-test robustness using both LLM-generated attacks and an in-the-wild 4chan attack. We find that low-dimensional detectors achieve competitive OOD performance while being faster, cheaper, and more interpretable than prompted LLM-based judges. Finally, human and LLM-based evaluations show that OOD queries primarily degrade the relevance of RAG outputs, showing the need for efficient external OOD detection to maintain safe, in-scope behavior.
arXiv:2508.02932v2 Announce Type: replace Abstract: Low-Rank Adaptation (LoRA) has gained popularity as a fine-tuning approach for Large Language Models (LLMs) due to its low resource requirements and good performance. While numerous studies have investigated ways to improve LoRA serving efficiency by serving multiple LoRAs concurrently, existing methods assume that a wide range of LoRA adapters are available for serving. In our work, we conduct extensive empirical studies to show that current LoRA training paradigms do not efficiently utilize hardware resources and incur high overhead to obtain a performant LoRA adapter. Leveraging these insights, we propose PLoRA, which automatically orchestrates concurrent LoRA fine-tuning jobs under given hardware and model constraints and develops performant kernels to improve training efficiency. Across a range of LLMs and LoRA configurations, PLoRA improves training throughput by up to 12.8x and reduces the overall fine-tuning makespan by up to 7.52x compared to existing approaches.
arXiv:2508.04796v3 Announce Type: replace Abstract: Tokenization is the first -- and often least scrutinized -- step of most NLP pipelines. Standard algorithms for learning tokenizers rely on frequency-based objectives, which favor languages dominant in the training data and consequently leave lower-resource languages with tokenizations that are disproportionately longer, morphologically implausible, or even riddled with $ $ placeholders. This phenomenon ultimately amplifies computational and financial inequalities between users from different language backgrounds. To remedy this, we introduce Parity-aware Byte Pair Encoding (BPE), a variant of the widely-used BPE algorithm. At every merge step, Parity-aware BPE applies a fair-max rule that maximizes the compression gain of the currently worst-compressed language, trading a small amount of global compression for cross-lingual parity. We find empirically that Parity-aware BPE reduces tokenization inequality -- operationalized by the Gini coefficient of per-language token costs -- by up to 89% relative to Classical BPE. This comes with negligible impact on global compression rate and no evidence of systematic degradation in downstream LM performance.
arXiv:2508.16771v3 Announce Type: replace Abstract: Code Language Models (CodeLLMs) learn token importance from data correlations, whereas human developers attend selectively to semantically salient code. We present EyeMulator, a model-agnostic method that injects human visual-attention priors into CodeLLM fine-tuning without architectural changes. EyeMulator distills eye-tracking data into semantic salience and gaze-transition priors, then uses them to reweight token-level training losses. Across six backbones, two data regimes, and three CodeXGLUE tasks, the reported configurations yield positive matched-metric deltas in all 36 model-task-setting cells. Effects are largest for structure-preserving completion and translation, while summarization shows smaller but positive METEOR deltas. Session-mode and component-ablation analyses further show that reading, writing, semantic, and transition-derived priors provide complementary signal. Human-attention artifacts are available at https://zenodo.org/records/17205682.
arXiv:2508.17896v5 Announce Type: replace Abstract: We propose the Steiner Traveling Salesman Problem with Time Windows and Pickup and Delivery, an advanced and practical extension of classical routing models. This variant integrates the characteristics of the Steiner Traveling Salesman Problem with time-window constraints, pickup and delivery operations and vehicle capacity limitations. These features closely mirror the complexities of contemporary logistics challenges, including last-mile distribution, reverse logistics and on-demand service scenarios. To tackle the inherent computational difficulties of this NP-hard problem, we propose two specialized mathematical formulations: an arc-based model and a node-oriented model, each designed to capture distinct structural aspects of the problem. We further introduce a preprocessing reduction method that eliminates redundant arcs, significantly enhancing computational performance and scalability. We also introduce a benchmark instance generator designed to create diverse and scalable test scenarios, providing a flexible framework for assessing the performance and robustness of both classical and quantum optimization methods. Both formulations are implemented using classical and quantum optimization approaches. In particular, the classical models are solved with Gurobi, whereas the quantum implementation is carried out on D-Wave's LeapCQMHybrid platform, a hybrid quantum-classical environment that integrates quantum annealing with classical optimization techniques for constrained problem solving. Numerical experiments are conducted to validate the proposed formulations and the preprocessing reduction method. The analyses performed assess the structural properties of the two models, their computational behavior, and the impact of preprocessing on problem size and solution efficiency.
arXiv:2508.20054v2 Announce Type: replace Abstract: The study of categories abstracting the structural properties of relations has been extensively developed over the years, resulting in a rich and diverse body of work. In a previous paper we offered a survey providing a modern presentation of these ``categories for relations'' as instances of gs-monoidal categories, showing how they arise as Kleisli categories of suitable symmetric monoidal monads. The end result was a taxonomy that organised numerous related concepts in the literature, including in particular Markov and restriction categories. This paper further enriches the taxonomy: it proposes two categories that are once more instances of gs-monoidal categories, yet more abstract than Markov and restriction categories. They are characterised by an axiomatic notion of mass and domain of an arrow, the latter one of the key ingredients of restriction categories, which generalises the domain of partial functions. The paper then introduces mass and domain preserving monads, proving that the associated Kleisli categories in fact preserve the corresponding equations and that these monads arise naturally for the categories of semiring-weighted relations.
arXiv:2509.01251v3 Announce Type: replace Abstract: This paper presents a joint effort towards the development of a data-driven Social Robot Navigation metric to facilitate benchmarking and policy optimization for ground robots. We provide the motivations for our approach and describe our proposal to format and store rated social navigation trajectory datasets. Following these guidelines, we compiled a first version of the proposed dataset with 4427 trajectories -- 182 real and 4245 simulated -- and presented it to human raters, yielding a total of 4402 rated trajectories after data quality assurance. Notably, we provide the first all-encompassing learned social robot navigation metric (SN26), along qualitative and quantitative results, including the test loss achieved, a comparison against hand-crafted metrics, and an ablation study. All data, software, and model weights are publicly available.
arXiv:2509.03463v3 Announce Type: replace Abstract: Large Language Models (LLMs) show strong potential for automating model generation from natural-language descriptions. A common approach begins with an initial model generation, followed by an iterative critique-refine loop in which the model is evaluated for issues and refined based on those issues. This process needs to address: (1) structural correctness -- compliance with well-formedness rules -- and (2) semantic alignment -- accurate reflection of the intended meaning in the source text. We present LADEX (LLM-based Activity Diagram Extractor), a pipeline for deriving activity diagrams from natural-language process descriptions using an LLM-driven critique-refine process. Structural checks in LADEX can be performed either algorithmically or by an LLM, while alignment checks are performed by an LLM. We design five ablated variants of LADEX to study: (i) the impact of the critique-refine loop itself, (ii) the role of LLM-based semantic checks, and (iii) the comparative effectiveness of algorithmic versus LLM-based structural checks. To evaluate LADEX, we compare generated diagrams with expert ground truths using a trace-based behavioural and an LLM-based matcher. This enables automated measurement of correctness (whether the generated activity diagram includes the ground-truth nodes) and completeness (how many of the ground-truth nodes the generated activity diagram covers). Experiments on two datasets -- a public-domain dataset and an industry dataset from our collaborator, Ciena -- indicate: (1) Both matchers yield similar completeness and correctness comparisons. (2) The critique-refine loop improves structural validity, correctness, and completeness compared to single-pass generation. (3) Activity diagrams refined based on algorithmic structural checks achieve structural consistency, whereas those refined based on LLM-based checks often still show structural inconsistencies.
arXiv:2509.05191v3 Announce Type: replace Abstract: Feedback Linearisation (FBL) is a widely used technique that applies feedback laws to transform input-affine nonlinear control systems into linear control systems, allowing for the use of linear controller design methods such as pole placement. However, for problems with state constraints, controlling the linear system induced by FBL can be more challenging than controlling the original system. This is because simple state constraints in the original nonlinear system become complex nonlinear constraints in the FBL induced linearised system, thereby diminishing the advantages of linearisation. To avoid increasing the complexity of state constraints under FBL, this paper introduces a method to first augment system dynamics to capture state constraints before applying FBL. We show that our proposed augmentation method leads to ill-defined relative degrees at state constraint boundaries. However, we show that ill-defined relative degrees can be overcome by using a switching FBL controller. Numerical experiments illustrate the capabilities of this method for handling state constraints within the FBL framework.
arXiv:2607.04317v1 Announce Type: new Abstract: In the online Multi-Level Aggregation Problem (MLAP), requests arrive over time and are associated with nodes of a given weighted rooted tree of depth $D$. Each request must eventually be served by performing a service. Serving a request consists of selecting a rooted subtree that contains the request's node, incurring a service cost equal to the total weight of the selected subtree. To reduce service costs, multiple requests may be served simultaneously by selecting a single rooted subtree that spans all of them. In addition, each request is associated with a penalty function that specifies the cost incurred when the request is served at a particular time. The objective is to minimize the total cost, consisting of both service costs and penalty costs. Most previous work on MLAP assumes monotone non-decreasing penalty functions, commonly referred to as delay functions. Only very recent results consider penalty functions that initially decrease and subsequently increase, and even then only for the special cases of depths $D=1$ and $D=2$, namely the Joint Replenishment Problem (JRP). In this work, we extend previous results in two ways. First, we allow arbitrary penalty functions, which may decrease and increase multiple times. Second, we study the general MLAP with arbitrary tree depth $D$ under these arbitrary penalty functions. We present a randomized algorithm that is $O(D \log n \log(nDW))$-competitive, where $W$ is the maximum service window among all penalty functions after normalizing the Lipschitz parameter of each penalty function to 1 and the minimum positive edge weight incident to the root to 1, and $n$ is the number of requests. Our algorithm runs in polynomial time. Moreover, even for $D=1$, the problem admits an $\Omega(\log n)$ hardness of approximation for polynomial-time algorithms.
arXiv:2607.04319v1 Announce Type: new Abstract: A companion paper showed that a transformer's feed-forward layer can be rebuilt from explicit fuzzy set operations - intersection, set-difference, and a self-forgetting sequence quantifier - so its hidden units read as named logical operators at no cost to language-model quality. That left the other half of the transformer opaque. Here we carry the same idea into attention and join the two into one model. The mechanism is minimal: a head's value is passed through a sigmoid, so each value channel becomes a readable detector of whether a feature holds at a token. This adds no parameters and leaves the standard head otherwise untouched. A Boolean variant goes further, restructuring the value into an explicit within-token intersection and negation-capable set-difference. In both designs the output projection is left free, not tied to the vocabulary, which is the load-bearing decision: bounding what a head detects while leaving what it writes unconstrained yields selective detectors, whereas constraining the write does not. A bounded value is shaped into a readable detector by two selectivity pressures - one for sparse firing, one for decisive firing at the rails - and which a design wants is not universal. Across five specialized-attention designs at 125M parameters, 44 to 62 percent of value channels become crisp, contextually selective detectors, and their legibility rises with depth rather than crystallizing only on punctuation. Language-model quality is at parity with a conventional baseline. Finally, we couple the Boolean attention to the legible feed-forward layer and train an end-to-end legible-by-construction language model at benchmark parity: its feed-forward units are named set and quantifier operations throughout, and we can take a token it generates and read the named units that compose to produce it.
arXiv:2509.13742v4 Announce Type: replace Abstract: Science communication revision requires writers to dynamically balance scientific exposition and narrative engagement - a process where writers often struggle with competing directions. Existing LLM-assisted tools help with co-writing, but offer limited support for navigating this iterative, multi-directional revision process. To address this gap, we designed Spatial Balancing, an exploratory revision environment that maps rhetorical goals and revision strategies onto a two-dimensional spatial canvas for experienced science communication creators with domain expertise but lacking formal professional training. By building a design space of communication strategies and embedding them into a spatial exploratory canvas, our system treats feedback as navigational cues rather than prescriptive judgments. Our findings show that this integrated revision environment helps writers stay focused on writing goals, reason about revision as trajectories, and explore alternatives, which supports greater metacognitive control and confidence without increasing workload. This work highlights the value of spatially externalized revision environments for supporting iterative, reflective thinking during LLM-assisted writing.
arXiv:2607.04359v1 Announce Type: new Abstract: We investigate the dynamics of charged particles interacting with ultra-intense electromagnetic X-modes in strongly magnetized plasmas. We demonstrate that particle motion becomes chaotic for relative wave intensities $\delta = B_w/B_0 \gtrsim 0.25$ (not above the field reversal threshold $\delta \geq 1$). The transition to chaos occurs via the Chirikov resonance overlap mechanism and the related destruction of Kolmogorov-Arnold-Moser (KAM) tori. The maximum Lyapunov exponent increases logarithmically with $\delta$, even though the unmagnetized $\delta \to \infty$ limit is strictly integrable. In the $\delta \gg 1$ regime, incomplete re-laminarization of the phase space flow leads to two distinct populations: (i) the majority of particles undergoing stochastic diffusion, and (ii) a fraction of particles that become phase-locked with the wave, experiencing macroscopic intermittent surfing (L\'evy flights). The 1D Particle-In-Cell simulations using the EPOCH code in the highly magnetized ($\sigma \gg 1$) and under-dense regime are generally consistent with the Hamiltonian single-particle theory. The dissipation fraction of the initial EM energy remains mild.
arXiv:2607.04381v1 Announce Type: new Abstract: Research-based digital health interventions are often presented as potential solutions for extending health care in the real world. Yet the vast majority of these interventions fails to move beyond controlled studies. Existing frameworks offer valuable guidance for intervention development and testing, but provide less concrete support for translating these evidenced intervention mechanisms into sustained real-world applications. This paper introduces the ABC framework, referring to Accessibility, Buildability, and Continuity, as a practical model for a successful translation. Accessibility captures whether diverse users can find, understand, and begin using an application with minimal friction. Buildability refers to the development of an app that supports the iteration, integration, and personalization of features. Continuity describes both sustained user engagement and the operational capacity to maintain an application over time without disproportionate increases in cost, infrastructure, or human support. Different combinations of the ABC-dimensions make an application scalable (AB), automated (BC), and adherent (AC). By linking design decisions to these features, ABC offers a shared language for researchers, designers, and policymakers seeking to build or evaluate digital health interventions that work beyond trials and are viable applications in everyday life.
arXiv:2607.05189v1 Announce Type: new Abstract: Persistent personal agents combine long-term memory with access to users' external environments, enabling personalized foreground assistance and proactive background execution. This integration also creates a new path to compromise: untrusted external content can be silently written into persistent memory and later reused as trusted state. We study this threat as stealth memory injection, in which a remote black-box adversary delivers a single email payload that must induce the agent to write poisoned memory, stay hidden in the agent's response to the user, and affect future behavior. We introduce WhisperBench, a 108-case benchmark spanning five risk categories and both fact and preference poisoning. Built on a real IMAP/SMTP workflow and an authentic email agent skill, it enables full-cycle evaluation of stealth memory injection attacks. To enable this black-box attack under single-email delivery and without runtime feedback, we propose MemGhost, a one-shot payload generation framework. MemGhost uses an environment proxy to emulate persistent-agent execution and an objective proxy to convert memory adoption and conversational stealth into dense rubric-based rewards, then trains the attacker policy with supervised fine-tuning and reinforcement learning. Across 56 held-out test cases, MemGhost achieves 87.5% end-to-end success on OpenClaw with GPT-5.4 and 71.4% on Claude Code SDK with Sonnet 4.6. It also transfers across personal-agent architectures (NanoClaw and Hermes Agent) and memory backends (filesystem and vector-based Mem0), and remains effective against input-level, model-level, and system-level defenses. These results suggest that persistent memory can turn ordinary external processing into a practical pathway for long-term agent compromise.
arXiv:2607.04410v1 Announce Type: new Abstract: We present the AI Wizards submission to EXIST 2026 for multimodal sexism identification in memes. The task is composed of three, increasingly harder subtasks. We model them hierarchically as conditional soft-label prediction over empirical annotator distributions. Our system maps fixed Gemini Embedding 2 vision-language representations through a lightweight Gated MLP trained with KL divergence and homoscedastic uncertainty weighting. Our submissions ranked first on Task 2.3 and fourth on Tasks 2.1 and 2.2 on the official Soft-Soft leaderboards. The code is available at https://github.com/NLP-AI-Wizards/EXIST-2026
arXiv:2510.10355v2 Announce Type: replace Abstract: The semi-implicit (partly decoupled, also called staggered or fraction-step) time discretization is applied to compressible nonlinear dynamical models of viscoelastic solids in the Eulerian description, i.e.\ in the actual deforming configuration, formulated fully in terms of rates. The Kelvin-Voigt rheology and also, in the deviatoric part, the Jeffreys rheology are considered. The numerical stability and, considering the Stokes-type viscosity multipolar of the 2nd-grade, also convergence towards weak solutions are proved in three-dimensional situations, exploiting the convexity of the kinetic energy when written in terms of linear momentum. No (poly)convexity of the stored energy is required and some enhancements (specifically towards damage and diffusion models) are briefly outlined, too.
arXiv:2607.04425v1 Announce Type: new Abstract: Recent advances in multimodal foundation models and agent systems have driven GUI agents from single-platform task execution toward cross-platform interaction. However, building multi-platform GUI agents remains challenging. On one hand, high-quality and executable cross-platform interaction trajectories are still scarce, and existing data often suffer from limited platform coverage. On the other hand, different platforms exhibit distinct interaction conventions, making joint or continual training prone to behavioral pattern mixing, platform-specific capability degradation, and catastrophic forgetting. To address these challenges, we construct Uni-GUI, a high-quality cross-platform GUI interaction dataset, and propose UI-MOPD, the first method that incorporates multi-teacher on-policy distillation into continual learning for GUI agents. UI-MOPD dynamically selects a platform-specific teacher according to the current environment and transfers platform-specific behavioral priors to a shared policy through platform-conditioned distillation, enabling adaptation to new platforms while preserving capabilities on existing ones. Experiments on OSWorld and MobileWorld show that UI-MOPD achieves task success rates of 38.2% and 12.0%, respectively, demonstrating its effectiveness in balancing cross-platform capability retention and new-platform adaptation. Project page: https://elispectre.github.io/UI-MOPD/.
arXiv:2509.15206v3 Announce Type: replace Abstract: The high memory demands of generative language models have drawn attention to quantization, which reduces memory usage by mapping model weights to lower-precision integers. However, recent empirical studies show that, while efficient, quantization can increase the likelihood of generating biased outputs and degrade performance on fairness benchmarks. In this work, we draw new links between quantization and model fairness by adding explicit group-fairness constraints to the quantization objective and introduce Fair-GPTQ, the first quantization method explicitly designed to reduce unfairness in large language models. The added constraints guide the learning of the rounding operation toward less-biased text generation for protected groups. Specifically, we focus on stereotype generation involving occupational bias and discriminatory language spanning gender, race, and religion. Fair-GPTQ has minimal impact on performance, preserving at least 90% of baseline accuracy on zero-shot benchmarks, reduces unfairness relative to a half-precision model, and retains the memory and speed benefits of 4-bit quantization.
arXiv:2509.21541v3 Announce Type: replace Abstract: Hair simulation and rendering are challenging due to complex strand dynamics, diverse material properties, and intricate light-hair interactions. Recent video diffusion models can generate high-quality videos, but they lack fine-grained control over hair dynamics. We present ControlHair, a hybrid framework that integrates a physics simulator with conditional video diffusion to enable precise and controllable dynamic hair rendering. ControlHair adopts a three-stage pipeline: it first encodes physics conditions into per-frame geometry using a simulator, then extracts per-frame control signals, and finally feeds control signals into a video diffusion model to generate videos with desired hair dynamics. This cascaded design decouples physics reasoning from video generation, supports diverse physics, and makes training the video diffusion model easy. Trained on a curated 10K video dataset, ControlHair outperforms text- and pose-conditioned baselines, delivering precisely controlled hair dynamics. We also demonstrate three use cases of ControlHair, including dynamic hairstyle try-on, bullet-time effects, and cinemagraphic. Project page: https://linwk20.github.io/controlhair-web.