Scaling LLaNA: Advancing NeRF-Language Understanding Through Large-Scale Training

Recent advances in Multimodal Large Language Models (MLLMs) have shown remarkable capabilities in understanding both images and 3D data, yet these modalities face inherent limitations in comprehensively representing object geometry and appearance. Neural Radiance Fields (NeRFs) have emerged as a promising alternative, encoding both geometric and photorealistic properties within the weights of a simple Multi-Layer Perceptron (MLP). This work investigates the feasibility and effectiveness of ingesting NeRFs into an MLLM. We introduce LLaNA, the first MLLM able to perform new tasks such as NeRF captioning and Q\&A, by directly processing the weights of a NeRF's MLP. Notably, LLaNA is able to extract information about the represented objects without the need to render images or materialize 3D data structures. In addition, we build the first large-scale NeRF-language dataset, composed by more than 300K NeRFs trained on ShapeNet and Objaverse, with paired textual annotations that enable various NeRF-language tasks. Based on this dataset, we develop a benchmark to evaluate the NeRF understanding capability of our method. Results show that directly processing NeRF weights leads to better performance on NeRF-Language tasks compared to approaches that rely on either 2D or 3D representations derived from NeRFs.

Perturb, Attend, Detect and Localize (PADL): Robust Proactive Image Defense

Image manipulation detection and localization have received considerable attention from the research community given the blooming of Generative Models (GMs). Detection methods that follow a passive approach may overfit to specific GMs, limiting their application in real-world scenarios, due to the growing diversity of generative models. Recently, approaches based on a proactive framework have shown the possibility of dealing with this limitation. However, these methods suffer from two main limitations, which raises concerns about potential vulnerabilities: i) the manipulation detector is not robust to noise and hence can be easily fooled; ii) the fact that they rely on fixed perturbations for image protection offers a predictable exploit for malicious attackers, enabling them to reverse-engineer and evade detection. To overcome this issue we propose PADL, a new solution able to generate image-specific perturbations using a symmetric scheme of encoding and decoding based on cross-attention, which drastically reduces the possibility of reverse engineering, even when evaluated with adaptive attack [31]. Additionally, PADL is able to pinpoint manipulated areas, facilitating the identification of specific regions that have undergone alterations, and has more generalization power than prior art on held-out generative models. Indeed, although being trained only on an attribute manipulation GAN model [15], our method generalizes to a range of unseen models with diverse architectural designs, such as StarGANv2, BlendGAN, DiffAE, StableDiffusion and StableDiffusionXL. Additionally, we introduce a novel evaluation protocol, which offers a fair evaluation of localisation performance in function of detection accuracy and better captures real-world scenarios.

Looking for Tiny Defects via Forward-Backward Feature Transfer

Motivated by efficiency requirements, most anomaly detection and segmentation (AD&S) methods focus on processing low-resolution images, e.g., $224\times 224$ pixels, obtained by downsampling the original input images. In this setting, downsampling is typically applied also to the provided ground-truth defect masks. Yet, as numerous industrial applications demand identification of both large and tiny defects, the above-described protocol may fall short in providing a realistic picture of the actual performance attainable by current methods. Hence, in this work, we introduce a novel benchmark that evaluates methods on the original, high-resolution image and ground-truth masks, focusing on segmentation performance as a function of the size of anomalies. Our benchmark includes a metric that captures robustness with respect to defect size, i.e., the ability of a method to preserve good localization from large anomalies to tiny ones. Furthermore, we introduce an AD&S approach based on a novel Teacher-Student paradigm which relies on two shallow MLPs (the Students) that learn to transfer patch features across the layers of a frozen vision transformer (the Teacher). By means of our benchmark, we evaluate our proposal and other recent AD&S methods on high-resolution inputs containing large and tiny defects. Our proposal features the highest robustness to defect size, runs at the fastest speed, yields state-of-the-art performance on the MVTec AD dataset and state-of-the-art segmentation performance on the VisA dataset.

LLaNA: Large Language and NeRF Assistant

Multimodal Large Language Models (MLLMs) have demonstrated an excellent understanding of images and 3D data. However, both modalities have shortcomings in holistically capturing the appearance and geometry of objects. Meanwhile, Neural Radiance Fields (NeRFs), which encode information within the weights of a simple Multi-Layer Perceptron (MLP), have emerged as an increasingly widespread modality that simultaneously encodes the geometry and photorealistic appearance of objects. This paper investigates the feasibility and effectiveness of ingesting NeRF into MLLM. We create LLaNA, the first general-purpose NeRF-language assistant capable of performing new tasks such as NeRF captioning and Q\&A. Notably, our method directly processes the weights of the NeRF's MLP to extract information about the represented objects without the need to render images or materialize 3D data structures. Moreover, we build a dataset of NeRFs with text annotations for various NeRF-language tasks with no human intervention. Based on this dataset, we develop a benchmark to evaluate the NeRF understanding capability of our method. Results show that processing NeRF weights performs favourably against extracting 2D or 3D representations from NeRFs.

Test Time Training for Industrial Anomaly Segmentation

Anomaly Detection and Segmentation (AD&S) is crucial for industrial quality control. While existing methods excel in generating anomaly scores for each pixel, practical applications require producing a binary segmentation to identify anomalies. Due to the absence of labeled anomalies in many real scenarios, standard practices binarize these maps based on some statistics derived from a validation set containing only nominal samples, resulting in poor segmentation performance. This paper addresses this problem by proposing a test time training strategy to improve the segmentation performance. Indeed, at test time, we can extract rich features directly from anomalous samples to train a classifier that can discriminate defects effectively. Our general approach can work downstream to any AD&S method that provides an anomaly score map as output, even in multimodal settings. We demonstrate the effectiveness of our approach over baselines through extensive experimentation and evaluation on MVTec AD and MVTec 3D-AD.

Multimodal Industrial Anomaly Detection by Crossmodal Feature Mapping

The paper explores the industrial multimodal Anomaly Detection (AD) task, which exploits point clouds and RGB images to localize anomalies. We introduce a novel light and fast framework that learns to map features from one modality to the other on nominal samples. At test time, anomalies are detected by pinpointing inconsistencies between observed and mapped features. Extensive experiments show that our approach achieves state-of-the-art detection and segmentation performance in both the standard and few-shot settings on the MVTec 3D-AD dataset while achieving faster inference and occupying less memory than previous multimodal AD methods. Moreover, we propose a layer-pruning technique to improve memory and time efficiency with a marginal sacrifice in performance.

Looking at words and points with attention: a benchmark for text-to-shape coherence

While text-conditional 3D object generation and manipulation have seen rapid progress, the evaluation of coherence between generated 3D shapes and input textual descriptions lacks a clear benchmark. The reason is twofold: a) the low quality of the textual descriptions in the only publicly available dataset of text-shape pairs; b) the limited effectiveness of the metrics used to quantitatively assess such coherence. In this paper, we propose a comprehensive solution that addresses both weaknesses. Firstly, we employ large language models to automatically refine textual descriptions associated with shapes. Secondly, we propose a quantitative metric to assess text-to-shape coherence, through cross-attention mechanisms. To validate our approach, we conduct a user study and compare quantitatively our metric with existing ones. The refined dataset, the new metric and a set of text-shape pairs validated by the user study comprise a novel, fine-grained benchmark that we publicly release to foster research on text-to-shape coherence of text-conditioned 3D generative models. Benchmark available at https://cvlab-unibo.github.io/CrossCoherence-Web/.

Semantic Image Synthesis via Class-Adaptive Cross-Attention

In semantic image synthesis, the state of the art is dominated by methods that use spatially-adaptive normalization layers, which allow for excellent visual generation quality and editing versatility. Granted their efficacy, recent research efforts have focused toward finer-grained local style control and multi-modal generation. By construction though, such layers tend to overlook global image statistics leading to unconvincing local style editing and causing global inconsistencies such as color or illumination distribution shifts. Also, the semantic layout is required for mapping styles in the generator, putting a strict alignment constraint over the features. In response, we designed a novel architecture where cross-attention layers are used in place of de-normalization ones for conditioning the image generation. Our model inherits the advantages of both solutions, retaining state-of-the-art reconstruction quality, as well as improved global and local style transfer. Code and models available at https://github.com/TFonta/CA2SIS.

Conditioning Diffusion Models via Attributes and Semantic Masks for Face Generation

Deep generative models have shown impressive results in generating realistic images of faces. GANs managed to generate high-quality, high-fidelity images when conditioned on semantic masks, but they still lack the ability to diversify their output. Diffusion models partially solve this problem and are able to generate diverse samples given the same condition. In this paper, we propose a multi-conditioning approach for diffusion models via cross-attention exploiting both attributes and semantic masks to generate high-quality and controllable face images. We also studied the impact of applying perceptual-focused loss weighting into the latent space instead of the pixel space. Our method extends the previous approaches by introducing conditioning on more than one set of features, guaranteeing a more fine-grained control over the generated face images. We evaluate our approach on the CelebA-HQ dataset, and we show that it can generate realistic and diverse samples while allowing for fine-grained control over multiple attributes and semantic regions. Additionally, we perform an ablation study to evaluate the impact of different conditioning strategies on the quality and diversity of the generated images.

Multimodal Side-Tuning for Document Classification

In this paper, we propose to exploit the side-tuning framework for multimodal document classification. Side-tuning is a methodology for network adaptation recently introduced to solve some of the problems related to previous approaches. Thanks to this technique it is actually possible to overcome model rigidity and catastrophic forgetting of transfer learning by fine-tuning. The proposed solution uses off-the-shelf deep learning architectures leveraging the side-tuning framework to combine a base model with a tandem of two side networks. We show that side-tuning can be successfully employed also when different data sources are considered, e.g. text and images in document classification. The experimental results show that this approach pushes further the limit for document classification accuracy with respect to the state of the art.