arXiv subjects

Cristian Sbrolli

Publications and source records attributed to Cristian Sbrolli.

9 recordsLinked to original sources

arXiv2026

Beyond Bag-of-Words: Diagnosing Compositional Binding Failures in Vision-Language Models

Modern vision-language models struggle with basic compositional reasoning, failing to bind attributes to objects or relations to their referents. Existing benchmarks either rely on noisy real images that conflate confounding visual variables with the reasoning failure, or use simplistic synthetic scenes lacking the realism modern VLMs are tuned for. We introduce \textbf{Auto-Comp}, a fully automated, concept-driven pipeline that bridges this gap by generating photorealistic compositional benchmarks at scale. Its core innovation is a \textit{parallel A/B construction}: for each concept, the pipeline emits a \textit{Minimal} sample (template caption, isolated objects on a white background) and a \textit{Contextual} sample (LLM-rewritten caption, objects embedded in a realistic scene), isolating core binding ability from visio-linguistic complexity. We instantiate \textit{four} task families spanning the two canonical axes of compositional binding: \textit{Color} and \textit{Shape-Color} (attribute binding), and \textit{Position} and \textit{Relative Size} (relational binding). We evaluate over 25 VLMs spanning CLIP, SigLIP, hard-negative-trained, and frontier generative models. The findings are consistent across architectures and scales: every model exhibits a large Swap-vs-Confusion gap, with low-entropy distractors (e.g., repeated objects or colors) exposing failures \textit{beyond} the known bag-of-words limitations. We further uncover a task-dependent trade-off: visio-linguistic context aids relational reasoning but hinders attribute binding through visual clutter. We publicly release the pipeline and benchmarks.

cs.CV↗

arXiv2026

Synesthesia via Direct Latent Augmentation:Bypassing the Decode-Encode Loop for Cross-Modal Distillation

While multimodal integration significantly improves computer vision models, deploying them incurs prohibitive inference costs and requires scarce, perfectly paired datasets. Recent methods address this data bottleneck by synthesizing missing modalities via generative AI, yet they introduce a severe inefficiency: the Decode-Encode Loop. Specifically, information-rich generative latents are decoded into noisy raw signals, forcing the downstream classifier to waste capacity re-encoding them. To bypass this bottleneck, we propose Direct Latent Augmentation (DLA), utilizing undecoded generative latents directly as privileged information. Furthermore, to transfer this dense knowledge to a purely visual student, we introduce Multilayer Explicit Simulated Synesthesia (MESSy). Instead of enforcing rigid representation matching, which forces the student to distort its native visual features to accommodate complex multimodal topologies, MESSy uses a predictive objective to safely internalize these physical priors. Empirical results demonstrate that our framework significantly outperforms raw data augmentation and traditional distillation. Ultimately, our approach yields highly accurate unimodal students with "synesthetic" latent structures that are inherently aligned with modalities they have never directly observed.

cs.CV↗

arXiv2026

PolyGen: Fully Synthetic Vision-Language Training via Multi-Generator Ensembles

Synthetic data offers a scalable solution for vision-language pre-training, yet current state-of-the-art methods typically rely on scaling up a single generative backbone, which introduces generator-specific spectral biases and limits feature diversity. In this work, we introduce PolyGen, a framework that redefines synthetic data construction by prioritizing manifold coverage and compositional rigor over simple dataset size. PolyGen employs a Polylithic approach to train on the intersection of architecturally distinct generators, effectively marginalizing out model-specific artifacts. Additionally, we introduce a Programmatic Hard Negative curriculum that enforces fine-grained syntactic understanding. By structurally reallocating the same data budget from unique captions to multi-source variations, PolyGen achieves a more robust feature space, outperforming the leading single-source baseline (SynthCLIP) by +19.0% on aggregate multi-task benchmarks and on the SugarCrepe++ compositionality benchmark (+9.1%). These results demonstrate that structural diversity is a more data-efficient scaling law than simply increasing the volume of single-source samples.

cs.CV↗

arXiv2025

SCENEFORGE: Enhancing 3D-text alignment with Structured Scene Compositions

The whole is greater than the sum of its parts-even in 3D-text contrastive learning. We introduce SceneForge, a novel framework that enhances contrastive alignment between 3D point clouds and text through structured multi-object scene compositions. SceneForge leverages individual 3D shapes to construct multi-object scenes with explicit spatial relations, pairing them with coherent multi-object descriptions refined by a large language model. By augmenting contrastive training with these structured, compositional samples, SceneForge effectively addresses the scarcity of large-scale 3D-text datasets, significantly enriching data complexity and diversity. We systematically investigate critical design elements, such as the optimal number of objects per scene, the proportion of compositional samples in training batches, and scene construction strategies. Extensive experiments demonstrate that SceneForge delivers substantial performance gains across multiple tasks, including zero-shot classification on ModelNet, ScanObjNN, Objaverse-LVIS, and ScanNet, as well as few-shot part segmentation on ShapeNetPart. SceneForge's compositional augmentations are model-agnostic, consistently improving performance across multiple encoder architectures. Moreover, SceneForge improves 3D visual question answering on ScanQA, generalizes robustly to retrieval scenarios with increasing scene complexity, and showcases spatial reasoning capabilities by adapting spatial configurations to align precisely with textual instructions.

cs.CV↗

arXiv2025

Your Image Generator Is Your New Private Dataset

Generative diffusion models have emerged as powerful tools to synthetically produce training data, offering potential solutions to data scarcity and reducing labelling costs for downstream supervised deep learning applications. However, effectively leveraging text-conditioned image generation for building classifier training sets requires addressing key issues: constructing informative textual prompts, adapting generative models to specific domains, and ensuring robust performance. This paper proposes the Text-Conditioned Knowledge Recycling (TCKR) pipeline to tackle these challenges. TCKR combines dynamic image captioning, parameter-efficient diffusion model fine-tuning, and Generative Knowledge Distillation techniques to create synthetic datasets tailored for image classification. The pipeline is rigorously evaluated on ten diverse image classification benchmarks. The results demonstrate that models trained solely on TCKR-generated data achieve classification accuracies on par with (and in several cases exceeding) models trained on real images. Furthermore, the evaluation reveals that these synthetic-data-trained models exhibit substantially enhanced privacy characteristics: their vulnerability to Membership Inference Attacks is significantly reduced, with the membership inference AUC lowered by 5.49 points on average compared to using real training data, demonstrating a substantial improvement in the performance-privacy trade-off. These findings indicate that high-fidelity synthetic data can effectively replace real data for training classifiers, yielding strong performance whilst simultaneously providing improved privacy protection as a valuable emergent property. The code and trained models are available in the accompanying open-source repository.

cs.CV↗

arXiv2025

Neuro-Symbolic Scene Graph Conditioning for Synthetic Image Dataset Generation

As machine learning models increase in scale and complexity, obtaining sufficient training data has become a critical bottleneck due to acquisition costs, privacy constraints, and data scarcity in specialised domains. While synthetic data generation has emerged as a promising alternative, a notable performance gap remains compared to models trained on real data, particularly as task complexity grows. Concurrently, Neuro-Symbolic methods, which combine neural networks' learning strengths with symbolic reasoning's structured representations, have demonstrated significant potential across various cognitive tasks. This paper explores the utility of Neuro-Symbolic conditioning for synthetic image dataset generation, focusing specifically on improving the performance of Scene Graph Generation models. The research investigates whether structured symbolic representations in the form of scene graphs can enhance synthetic data quality through explicit encoding of relational constraints. The results demonstrate that Neuro-Symbolic conditioning yields significant improvements of up to +2.59% in standard Recall metrics and +2.83% in No Graph Constraint Recall metrics when used for dataset augmentation. These findings establish that merging Neuro-Symbolic and generative approaches produces synthetic data with complementary structural information that enhances model performance when combined with real data, providing a novel approach to overcome data scarcity limitations even for complex visual reasoning tasks.

cs.CV↗

arXiv2024

No Captions, No Problem: Captionless 3D-CLIP Alignment with Hard Negatives via CLIP Knowledge and LLMs

In this study, we explore an alternative approach to enhance contrastive text-image-3D alignment in the absence of textual descriptions for 3D objects. We introduce two unsupervised methods, $I2I$ and $(I2L)^2$, which leverage CLIP knowledge about textual and 2D data to compute the neural perceived similarity between two 3D samples. We employ the proposed methods to mine 3D hard negatives, establishing a multimodal contrastive pipeline with hard negative weighting via a custom loss function. We train on different configurations of the proposed hard negative mining approach, and we evaluate the accuracy of our models in 3D classification and on the cross-modal retrieval benchmark, testing image-to-shape and shape-to-image retrieval. Results demonstrate that our approach, even without explicit text alignment, achieves comparable or superior performance on zero-shot and standard 3D classification, while significantly improving both image-to-shape and shape-to-image retrieval compared to previous methods.

cs.CV↗

arXiv2024

Can Shape-Infused Joint Embeddings Improve Image-Conditioned 3D Diffusion?

Recent advancements in deep generative models, particularly with the application of CLIP (Contrastive Language Image Pretraining) to Denoising Diffusion Probabilistic Models (DDPMs), have demonstrated remarkable effectiveness in text to image generation. The well structured embedding space of CLIP has also been extended to image to shape generation with DDPMs, yielding notable results. Despite these successes, some fundamental questions arise: Does CLIP ensure the best results in shape generation from images? Can we leverage conditioning to bring explicit 3D knowledge into the generative process and obtain better quality? This study introduces CISP (Contrastive Image Shape Pre training), designed to enhance 3D shape synthesis guided by 2D images. CISP aims to enrich the CLIP framework by aligning 2D images with 3D shapes in a shared embedding space, specifically capturing 3D characteristics potentially overlooked by CLIP's text image focus. Our comprehensive analysis assesses CISP's guidance performance against CLIP guided models, focusing on generation quality, diversity, and coherence of the produced shapes with the conditioning image. We find that, while matching CLIP in generation quality and diversity, CISP substantially improves coherence with input images, underscoring the value of incorporating 3D knowledge into generative models. These findings suggest a promising direction for advancing the synthesis of 3D visual content by integrating multimodal systems with 3D representations.

cs.CV↗

arXiv2023

IC3D: Image-Conditioned 3D Diffusion for Shape Generation

In recent years, Denoising Diffusion Probabilistic Models (DDPMs) have demonstrated exceptional performance in various 2D generative tasks. Following this success, DDPMs have been extended to 3D shape generation, surpassing previous methodologies in this domain. While many of these models are unconditional, some have explored the potential of using guidance from different modalities. In particular, image guidance for 3D generation has been explored through the utilization of CLIP embeddings. However, these embeddings are designed to align images and text, and do not necessarily capture the specific details needed for shape generation. To address this limitation and enhance image-guided 3D DDPMs with augmented 3D understanding, we introduce CISP (Contrastive Image-Shape Pre-training), obtaining a well-structured image-shape joint embedding space. Building upon CISP, we then introduce IC3D, a DDPM that harnesses CISP's guidance for 3D shape generation from single-view images. This generative diffusion model outperforms existing benchmarks in both quality and diversity of generated 3D shapes. Moreover, despite IC3D's generative nature, its generated shapes are preferred by human evaluators over a competitive single-view 3D reconstruction model. These properties contribute to a coherent embedding space, enabling latent interpolation and conditioned generation also from out-of-distribution images. We find IC3D able to generate coherent and diverse completions also when presented with occluded views, rendering it applicable in controlled real-world scenarios.

cs.CV↗