Fine-tuning-based adaptation is widely used to customize diffusion-based image generation, leading to large collections of community-created adapters that capture diverse subjects and styles. Adapters derived from the same base model can be merged linearly, enabling the synthesis of new visual results within a vast and continuous design space. To explore this space, current workflows rely on manual slider-based tuning, an approach that scales poorly and makes merging coefficient selection difficult, even when the candidate set is limited to 20–30 adapters. We propose GimmBO to support interactive exploration of adapter merging for image generation through Preferential Bayesian Optimization (PBO). Motivated by observations from real-world usage, including sparsity and constrained coefficient ranges, we introduce a two-stage BO backend that improves sampling efficiency and convergence in high-dimensional spaces. We evaluate our approach with simulated users and a user study, demonstrating improved convergence, high success rates, and consistent gains over BO and line-search baselines, and further show the flexibility of the framework through several extensions.
论文检索
输入标题、作者或关键词，从 5,418 篇学术成果中精准定位
Differentiable volume rendering enables gradient-based optimization of volumetric scenes, but unbiased estimators suffer from high gradient variance. We observe that the extinction gradients split into two components on structurally different integration domains: a scattering term evaluated at a single path vertex, and a transmittance term integrated along the ray segment. Because the domains are mismatched, existing estimators sample the two components at different locations, leaving the negative correlation between their opposite-signed contributions unexploited. We expose this overlooked correlation and exploit it through a principle we call sample matching: evaluate both components at shared sample locations. To enable this, we derive the first reformulation of the differential path integral that couples the two contributions within a single integrand, yielding an unbiased Monte Carlo estimator that ties them together by construction. For efficiency, the estimator reuses partially sampled light paths and amortizes in-scattering cost by evaluating gradients at multiple probe points per segment. On voxel-grid reconstruction, our estimator reduces gradient variance by up to 80% over differential ratio tracking (DRT), yielding faster convergence and higher reconstruction quality.
Sampling strategies in computer graphics have long been divided between local approaches, which optimize sample distributions independently at each pixel, and global approaches, which use high-dimensional low-discrepancy sequences to ensure uniformity across all dimensions, including adjacent pixels. While global samplers are most-commonly used thanks to generally having better convergence rates, it comes at a cost of very limited user control over the distribution of samples on subspaces, making it difficult to handle common aliasing artifacts. We introduce a novel modular meta-sampler architecture that bridges local and global sampling, allowing integrator designers to employ specialized low-dimensional samplers while still achieving high-dimensional uniformity. Our approach leverages high-dimensional low-discrepancy sequences to orchestrate sample generation by a collection of local samplers operating over power-of-two hierarchical intervals. We demonstrate how existing local sampling techniques, including stratified, blue noise, and dyadic nets sampling, can be reformulated to be used with this framework, enabling hybrid sampling strategies that combine the benefits of both paradigms.
Real-time path tracing increasingly operates under extremely low sampling budgets, often below one sample per pixel, as rendering complexity, resolution, and frame-rate requirements continue to rise. Superresolution is widely used in production because it reduces path-tracing cost by tracing rays on a coarser image grid and reconstructing missing details. This creates a uniform tradeoff between cost and spatial detail: every image region receives the same reduced ray budget, although path-tracing noise, reconstruction difficulty, and perceptual importance vary strongly across the image. Adaptive sampling offers a compelling alternative, but existing end-to-end approaches rely on approximations that break down in sparse regimes. We introduce an end-to-end adaptive sampling and denoising pipeline explicitly designed for the sub-1-spp regime. Our method uses a stochastic formulation of sample placement that enables gradient estimation despite discrete sampling decisions, allowing stable training of a neural sampler at low sampling budgets. To better align optimization with human perception, we propose a tone-mapping-aware training pipeline that integrates differentiable filmic operators and a state-of-the-art perceptual loss, preventing oversampling of regions with low visual impact. In addition, we introduce a gather-based pyramidal denoising filter and a learnable generalization of albedo demodulation tailored to sparse sampling. Our results show consistent improvements over uniform sparse sampling, with notably better reconstruction of perceptually critical details such as specular highlights and shadow boundaries, and demonstrate that adaptive sampling remains effective in the sub-1-spp regime.
We propose a system to optimize parametric designs subject to radiation pressure, i.e., the effect of light on the motion of objects. This is most relevant in the design of spacecraft, where radiation pressure presents the dominant non-conservative forcing mechanism, which is the case beyond approximately 800 km altitude. Despite its importance, the high computational cost of high-fidelity radiation pressure modeling has limited its use in large-scale spacecraft design, optimization, and space situational awareness applications. We enable this by offering three innovations in the simulation, in representation and in optimization: First, a practical computer graphics-inspired Monte-Carlo (MC) simulation of radiation pressure. The simulation is highly parallel, uses importance sampling and next-event estimation to reduce variance and allows simulating an entire family of designs instead of a single spacecraft as in previous work. Second, we introduce neural networks as a representation of forces from design parameters. This neural proxy model, learned from simulations, is inherently differentiable and can query forces orders of magnitude faster than a full MC simulation. Third, and finally, we demonstrate optimizing inverse radiation pressure designs, such as finding geometry, material or operation parameters that minimizes travel time, maximizes proximity given a desired end-point, minimize thruster fuel, trains mission control policies or allocated compute budget in extraterrestrial compute.
Multi-axis fused filament fabrication (FFF) reduces staircase artifacts and support material by adapting nozzle orientation. However, reliable and highly customizable printing requires layers and toolpaths combining a number of challenging properties: following prescribed fields, preserving constant bead geometry, avoiding stops, retractions, and collisions. We present AtomSlicer, which takes a 3D tool-orientation field and perlayer 2-RoSy tangent direction fields, to generate field-aligned non-planar layers of near-constant thickness. It then computes collision-free deposition toolpaths that are continuous within each layer, and as continuous as possible across layers. AtomSlicer encodes geometry with three orthogonal phase fields sampled into oriented atoms, partitions them into fabricable layers, reconstructs layer meshes, and synthesizes 2-RoSy-aligned toolpaths. Unlike prior methods it treats user- or optimizer-defined fields as constraints, producing a feasible toolpath or reporting non-fabricability rather than modifying the specification. We validate on varied shapes, including a large Thingi10k study, and 13 printed models using constant extrusion per unit path length under a fixed bead profile. AtomSlicer reduces non-extruding travel to a few percent, cuts travel moves by one to two orders of magnitude versus planar and Atomizer baselines, and is 9×-60× faster than Atomizer.
Reconstructing realistic, physically plausible garments from a single image remains a fundamental challenge. Template-free methods capture surface geometry but lack explicit sewing structure for simulation; while programmatic systems are simulation-ready but constrained by predefined templates. This reveals a fundamental representation gap between geometric reconstruction and structured garment construction. We present PatternGSL, a structured garment representation in the form of a template-free and learnable specification language that encodes complete sewing patterns, including panel boundaries, parameterized seams, and explicit stitch topology, in a compact and standardized form. PatternGSL preserves the physical rigor of pattern-based models while removing template dependence, elevating sewing structure as a first-class target for generative modeling. We further propose a vision-language framework that predicts PatternGSL specifications directly from a single image and decodes them into garments using lightweight deterministic validity handling, without optimization-based refinement or manual cleanup. In addition, we introduce PatternGSLData, the first large-scale image-to-GSL paired dataset comprising 300K samples with complete sewing pattern annotations, enabling supervised VLM training for structured garment reconstruction. Experiments demonstrate improved pattern accuracy over prior baselines, explicit sewing-structure recovery, reliable cloth simulation, and pattern-level editing through the same deterministic decoding pipeline. Code and data-processing scripts will be released at https://github.com/PatternGSL/PatternGSL.
We introduce Kinematic Kitbashing, an optimization framework that synthesizes articulated 3D objects by assembling reusable parts conditioned on an abstract kinematic graph. Given the graph and a library of articulated parts, our method optimizes per-part similarity transformations that place, orient, and scale each component into a coherent articulated object; optional graph edits further enable novel assemblies beyond the prescribed connectivity. Central to our method is an exemplar-based analogy for part placement: each reused component is paired with a single source asset that exemplifies how it attaches to its parent. We capture this attachment context using vector distance fields and measure consistency by integrating the matching error over the joint’s full motion range. This yields a kinematics-aware attachment energy that favors placements that preserve the exemplar’s local attachment neighborhood throughout articulation. To incorporate task-level functionality, we use this attachment energy as a prior in an annealed Langevin sampling framework, enabling gradient-free optimization of black-box functionality objectives. We demonstrate the versatility of kinematic kitbashing across diverse applications, including instantiating kinematic graphs from user-selected or automatically retrieved parts, synthesizing assemblies with user-defined functionality, and re-targeting articulations via graph edits.
We present Diceplay, a modular physical display for abstract visual composition built from a grid of identical dice. Each die has six faces with distinct geometric primitives, and images emerge through the placement and orientation of the dice. While this medium enables reusable and reconfigurable physical imagery, it poses a challenging design problem: images must be expressed through discrete, extremely low-resolution abstractions, making manual authoring difficult. To address this challenge, we introduce a computational design system that automatically generates Diceplay configurations from text prompts. Our key technical contribution is a grammar-based formulation that relaxes this discrete design space into a smooth optimization landscape, enabling gradient-based optimization using score distillation sampling. We show that our approach consistently produces meaningful abstractions for this medium, whereas state-of-the-art smoothing techniques fail in this extremely challenging regime. We demonstrate our method across a range of prompts and fabricated examples, showing how computationally generated abstractions can be realized as physical visual artifacts.
The low-poly style is a popular genre of vector graphics that depicts objects and scenes as flat-shaded meshes of low polygon count, often with a limited palette. In this paper, we propose a method to generate 2D low-poly meshes to abstract images. While it has been possible to achieve this look with general-purpose image generators and vector-based diffusion models, the resulting images are not guaranteed to be valid polygonal meshes. A key problem is that polygons overlap or intersect and, in the case of pixel-based image generators, the shapes are often not polygons. Moreover, the colors are not guaranteed to be constrained to a fixed palette. Aside from aesthetic considerations, this has practical consequences: it complicates editing and fabricating the results. We solve this problem by representing an image as a 2D polygonal mesh and optimizing the topology, geometry and coloring of the mesh using score distillation sampling, while enforcing geometric constraints, such as manifoldness and bijectivity. This presents unique challenges due to the discrete nature of the topology, which we handle using a fine-to-coarse strategy based on mesh simplification. By also constraining the colors to a fixed palette, we are able to produce various fabrications such as mosaics, embroidery, crocheting, patchwork and stencils from the resulting vector images. Code is at https://github.com/rubenwiersma/alps.
Low-discrepancy points have proven to be a useful source of samples for numerical integration via quasi-Monte Carlo integration. Low-discrepancy points are typically uniformly distributed in the unit hypercube. It is thus common to transform the input uniform points into target non-uniform samples via mapping, just like how uniform random numbers are used for generating non-uniform samples. Unfortunately, this common approach does not guarantee to carry over the low-discrepancy nature of the input points into the output non-uniform samples, giving the resulting non-uniform samples suboptimal discrepancy. The existing methods for generating low-discrepancy points focus only on uniform points, thus generating non-uniform and low-discrepancy samples has been an open problem. We introduce an optimization framework to approximate non-uniform low-discrepancy samples by numerically minimizing the discrepancy metric for non-uniform samples. We propose a differentiable numerical estimator of the generalized non-uniform discrepancy to enable gradient-based optimization. Our differentiable discrepancy estimator turns a challenging problem of generating non-uniform low-discrepancy samples into a simple numerical optimization problem. Since optimization can generate only one set of samples, we further propose a neural network architecture to encode the sample generation process, allowing us to generate multiple sets of samples from conditional targets without runtime optimization. We experimentally validate that our method reduces discrepancy below prior methods at least up to a certain sample size and consequently achieves lower integration error on numerical integration problems.
Imaging objects hidden outside the direct line of sight expands the effective field of view and is critical for applications such as autonomous driving and robotic perception. Despite impressive progress in time-of-flight (ToF)-based non-line-of-sight (NLOS) imaging, real-world deployment remains challenging because practical measurements are often collected over spatially limited, arbitrarily shaped relay regions—conditions that violate the planar-wall and dense-sampling assumptions made by most existing methods. To address these limitations, we propose a LOS-guided NLOS imaging pipeline that imposes no geometric assumptions on the relay surface and naturally supports both confocal and non-confocal configurations. Our method represents the hidden scene using 3D Gaussian primitives and couples them with an efficient, differentiable transient rendering model, enabling end-to-end optimization directly from measured transients. We validate our approach on real-world measurements from both a public dataset and a custom-built capture system. Across settings, our method achieves state-of-the-art reconstruction fidelity under spatially limited, sparsely sampled conditions, and significantly outperforms existing methods on complex, arbitrary relay surface geometries. We release our code and datasets at nlos-3d-gtr.
Achieving high-quality real-time rendering on mobile devices is critical for immersive 3D applications, yet 3D Gaussian Splatting (3DGS) faces significant challenges on resource-constrained mobile platforms, with sustaining 60 FPS at 2K resolution remaining difficult despite recent advances. Rather than further optimizing the core splatting pipeline, we explore 3DGS acceleration through post-rendering techniques leveraging gradient for super-sampling and frame interpolation. We propose Mobile3DGS3, a lightweight framework that alternates between super-sampling and frame interpolation. Our framework introduces three key innovations: Tangent Space Gradient Caching (TSGC), which converts gradient into cached geometric invariants enabling seamless integration with hardware-accelerated rasterization; Gradient-Aware Super-Sampling (GASS), which leverages Gaussian splat differentiability to upsample low-resolution images while preserving details; and Gradient-Aware Frame Interpolation (GAFI), which resamples gradient from adjacent frames with efficient motion vector upscaling for NPU-compatible processing. Extensive experiments demonstrate significant rendering efficiency improvements on mobile devices while maintaining high visual fidelity, presenting the first 3DGS acceleration paradigm on mobile devices that effectively leverages gradient in both super-sampling and frame interpolation. Code repository: https://github.com/SyouSanGin/Mobile3DGS3
Different shape representations facilitate different computations. Surface representations, in particular meshes, are often used for modeling, whereas volume representations are useful for spatial queries such as intersection or containment. Optimizing a surface representation based on a volumetric properties by gradient descent requires the derivatives of the volume relative to its bounding surface. We derive this gradient for winding numbers and show that it can be efficiently computed for volumetric values sampled on a regular grid (voxel representation) and surface parameters based on vertex sets (triangle meshes). This enables an efficient solution for a variety of optimization problems. We demonstrate the practical use of this approach at the examples of deforming meshes to resolve intersections, being manufacturable by cutting with a bandsaw from three directions, and creating shapes that are close to tiling 3D space.
Boundary Representation (B-rep) is the most commonly used data format in Computer-Aided Design (CAD) due to its analytical precision and direct support for parametric editing. However, its heterogeneous data structure— continuous parametric geometry with discrete topological graphs—poses fundamental challenges for deep learning models. Existing methods often directly predict the heterogeneous B-rep graph, relying on fixed-size padding or sequential tokenization to handle the varying cardinality of the geometric primitives. These approaches struggle with the combinatorial complexity of CAD models. The discrete, non-differentiable nature of the graph data structure prevents end-to-end optimization of the geometry and watertightness. In this work we introduce DualBrep, a novel continuous representation that unifies B-rep geometry and topology within a fully structured Euclidean domain. DualBrep encodes a CAD model using dual scalar fields: a Signed Distance Function (SDF) to represent the global shape geometry, and an Unsigned Distance Field (UDF) that implicitly encodes the topological structure via a Voronoi partitioning of the surface elements. Rather than processing these fields independently, we compress them into a single latent space. While the dual-field formulation alone already gives reconstruction a flexible, primitive-free segmentation signal that adapts to arbitrary face counts and surface types, the shared latent also makes generation tractable: a Flow Matching model can sample geometry and topology jointly from a single code, avoiding the error accumulation that plagues sequential or autoregressive B-rep predictors. Finally, we use a neural rebuilder to extract explicit B-rep models—comprising both prismatic and free-form primitives—directly from our continuous dual scalar fields. We demonstrate that DualBrep serves as a robust, unified backbone for CAD B-rep learning, achieving strong performance in both reverse engineering from raw point clouds and generative modeling via latent flow matching. Code is available at https://github.com/AutodeskAILab/DualBrep.
Large Language Models (LLMs) are now widely deployed across modern web services, but their safe and trustworthy use in real-world settings critically depends on accurate alignment with human preferences. Preference alignment is typically achieved using methods such as reinforcement learning or direct preference optimization (DPO), whose effectiveness in practice hinges on the quality of labeled preference data. However, a fundamental practical challenge remains: preference datasets inevitably contain noise. Through a systematic analysis of mainstream preference datasets, we find that roughly 25% of preference pairs show clear inconsistencies between reward-model evaluations and human annotations. Such inconsistent examples do not convey reliable preference signals; training directly on them therefore not only fails to improve alignment but can even degrade model behavior. To address this problem, we propose Noise-Aware Preference Alignment for LLMs via Confidence and Polarity Reweighting (AlignCP), a fully automated, human-free framework for noise-aware preference alignment. AlignCP derives two interpretable metrics from reward-model outputs: Confidence, which measures the reliability of each preference judgment, and Polarity, which evaluates whether the reward-model ranking agrees with the original human label. These metrics are combined to assign a training weight to each sample—amplifying high-confidence, label-consistent pairs while down-weighting or discarding low-confidence, contradictory, or otherwise noisy pairs. Unlike methods that rely on human re-inspection, repeated relabeling, or heavy computational reruns, AlignCP performs automated data-quality control with minimal overhead and no human intervention. Empirical results show that AlignCP substantially outperforms existing data-centric alignment approaches on standard preference benchmarks and remains more robust under noisy supervision.
Large Language Models (LLMs) have revolutionized Recommender Systems (RS) through advanced generative user modeling. However, LLM-based RS (LLM-RS) often inadvertently perpetuates bias present in the training data, leading to severe fairness issues. Addressing these fairness problems in LLM-RS faces two significant challenges. 1) Existing debiasing methods, designed for specific bias types, lack the generality to handle diverse or emerging biases in real-world applications. 2) Debiasing methods relying on retraining are computationally infeasible given the massive parameter scale of LLMs. To overcome these challenges, we propose FUDLR (Fast Unified Debiasing for LLM-RS). The core idea is to reformulate the debiasing problem as an efficient machine unlearning task with two stages. First, FUDLR identifies bias-inducing samples to unlearn through a novel bias-agnostic mask, optimized to balance fairness improvement with accuracy preservation. Its bias-agnostic design allows adaptability to various or co-existing biases simply by incorporating different fairness metrics. Second, FUDLR performs efficient debiasing by estimating and removing the influence of identified samples on model parameters. Extensive experiments demonstrate that FUDLR effectively and efficiently improves fairness while preserving recommendation accuracy, offering a practical path toward socially responsible LLM-RS. The code and data are available at https://github.com/JinLi-i/FUDLR.
As Large Language Models (LLMs) increasingly shape online content, how to remove targeted information from well-trained LLMs (also known as LLM unlearning) has become increasingly critical for web governance. A key challenge in LLM unlearning lies in the sample-wise imbalance within the forget set: different samples exhibit widely varying unlearning difficulty, leading to asynchronous forgetting speeds where some knowledge remains insufficiently erased while others become over-forgotten. To address this challenge, we propose BalDRO, a novel and efficient framework for balanced LLM unlearning. BalDRO formulates unlearning as a min–sup process, where the inner process identifies a worst-case data distribution that adaptively emphasizes hard-to-unlearn samples, while the outer process updates model parameters based on the worst-case data distribution. We instantiate this formulation through two efficient variants: BalDRO-G, a discrete GroupDRO-based approximation that focuses on high-loss subsets, and BalDRO-DV, a continuous Donsker–Varadhan dual method that enables smooth, adaptive weighting within standard LLM training pipelines. Extensive experiments on the TOFU and MUSE benchmarks demonstrate the effectiveness of our proposed BalDRO, yielding significant improvements in both forgetting quality and model utility over existing methods. For reproducibility, we have released the code for BalDRO. https://github.com/nxZhai/BalDRO.
Online advertising platforms serve as a critical bridge between advertisers and media, requiring precise prediction of user behaviors. Systematic underestimation or overestimation in these predictions can undermine the interests of both parties. Existing calibration methods fall short in addressing two key challenges. First, significant differences in CVR distributions across various targets lead to biased calibration when using global posterior statistics, causing some sample groups to be overestimated while others are underestimated. Second, current field-aware approaches are typically limited to single-field calibration and fail to account for field sensitivity. To overcome these limitations, we propose a Pareto Frontier-based Multi-field Personalized Calibration (PF-MPC) method which formulates multi-field calibration as a multi-objective optimization problem. PF-MPC identifies the optimal Pareto-efficient weight combinations to balance the conflicting calibration errors across different fields. We evaluate PF-MPC on both public calibration benchmark and a large-scale industrial dataset. Experimental results demonstrate that our method achieves significant improvements in calibration performance compared to existing approaches.
Most large-scale recommender systems follow a multi-stage cascade of retrieval, pre-ranking, ranking, and re-ranking. A key challenge at the pre-ranking stage arises from the heterogeneity of training instances sampled from coarse-grained retrieval results, fine-grained ranking signals, and exposure feedback. Our analysis reveals that prevailing pre-ranking methods, which indiscriminately mix heterogeneous samples, suffer from gradient conflicts: hard samples dominate training while easy ones remain underutilized, leading to suboptimal performance. We further show that the common practice of uniformly scaling model complexity across all samples is inefficient, as it overspends computation on easy cases and slows training without proportional gains. To address these limitations, this paper presents Heterogeneity-Aware Adaptive Pre-ranking (HAP), a unified framework that mitigates gradient conflicts through conflict-sensitive sampling coupled with tailored loss design, while adaptively allocating computational budgets across candidates. Specifically, HAP disentangles easy and hard samples, directing each subset along dedicated optimization paths. Building on this separation, it first applies lightweight models to all candidates for efficient coverage, and further engages stronger models on the hard ones, maintaining accuracy while reducing cost. This approach not only improves pre-ranking effectiveness but also provides a practical perspective on scaling strategies in industrial recommender systems. HAP has been deployed in the Toutiao production system for 9 months, yielding up to 0.4% improvement in user app usage duration and 0.05% in active days, without additional computational cost. We also release a large-scale industrial hybrid-sample dataset to enable the systematic study of source-driven candidate heterogeneity in pre-ranking.