One Patch to Caption Them All: A Unified Zero-Shot Captioning Framework

Preprint ONE PATCH TO CAPTION THEM ALL: A UNIFIED ZEROSHOT CAPTIONING FRAMEWORK Lorenzo Bianchi*1,2 Giacomo Pacini*1,2 Fabio Carrara1Nicola Messina1 Giuseppe Amato1Fabrizio Falchi1 1CNR-ISTI 2Università di Pisa *Equal contribution ABSTRACT Zero-shot captioners are recently proposed models that utilize common-space vision-language representations to caption images without relying on paired imagetext data. To caption an image, they proceed by textually decoding a text-aligned image feature, but they limit their scope to global representations and whole-image captions. We present Patch-ioner, a unified framework for zero-shot captioning that shifts from an image-centric to a patch-centric paradigm, enabling the captioning of arbitrary regions without the need of region-level supervision. Instead of relying on global image representations, we treat individual patches as atomic captioning units and aggregate them to describe arbitrary regions, from single patches to non-contiguous areas and entire images. We analyze the key ingredients that enable current latent captioners to work in our novel proposed framework. Experiments demonstrate that backbones producing meaningful, dense visual features, such as DINO, are key to achieving state-of-the-art performance in multiple region-based captioning tasks. Compared to other baselines and state-of-the-art competitors, our models achieve better performance on zero-shot dense, region-set, and a newly introduced trace captioning task, highlighting the effectiveness of patchwise semantic representations for scalable caption generation. Project page at paciosoft.com/Patch-ioner/. 1 INTRODUCTION Image captioning is one of the most representative tasks in vision-language understanding and has reached incredible accuracy thanks to the availability of pre-trained vision-language backbones and large paired image-text datasets. In its basic formulation, a captioning model takes a full image as input and autonomously decides which elements must be described and up to what degree. To enable user guidance and produce more targeted descriptions, some previous works proposed region-level captioning methods (Johnson et al.,2016;Cornia et al.,2019), which take as an additional input a spatial indication — e.g., bounding boxes — specifying which image regions have to be described and, possibly, in which order. These region-level captioning methods require expensive manually labeled data to fully supervise the model. Indeed, each sequence or set of bounding boxes for a given image should correspond to a manually written ground-truth caption describing those objects. This fully supervised solution does not scale properly. In this paper, we propose a perspective shift that enables us to perform region-level captioning with arbitrary spatial granularity — from a single image patch up to the entire image — in a zero-shot fashion, i.e., without requiring any form of image-level or region-level supervision. Specifically, instead of relying on the idea that the subject of a captioning method is the image — then potentially conditioned on a set of sub-regions — we instead build on two straightforward yet powerful ideas: i) the simplest element that we could caption is a patch, the atomic element of an image representation in modern architectures based on vision transformers (Dosovitskiy et al.,2021), and ii) we can easily aggregate multiple patch representations to produce descriptions for arbitrarily large — and also potentially not contiguous — image regions. We present Patch-ioner, a zero-shot captioning framework implementing these ideas. Our formulation offers maximum flexibility in zero-shot captioning tasks, producing models that effortlessly generate captions for various aggregations of 1 arXiv:2510.02898v2 [cs.CV] 6 Oct 2025 Preprint image patches, ranging from individual patches to larger image regions, up to providing a caption for the entire image. Despite the powerful perspective change that defines the patch as the new captioning unit, the problem is now entangled in a simple yet critical question: how can we craft a model able to provide patch-level captions without relying on any direct patch-level ground truth supervision? In the last years, large pre-trained vision-language foundation models like CLIP (Radford et al., 2021;Jia et al.,2021;Li et al.,2023a) solved many downstream tasks in zero-shot or even trainingfree configurations. In particular, contrastively learned vision-language representations enabled impressive results in zero-shot settings in image classification (Radford et al.,2021;Zhai et al.,2022), open-vocabulary detection (Zhou et al.,2022;Minderer et al.,2023), and segmentation (Ghiasi et al., 2022;Liang et al.,2023), or text-image retrieval (Kordopatis-Zilos et al.,2025). Image captioning, however, cannot directly employ CLIP machinery at inference time to generate text, given that CLIP is inherently a discriminative — and not a generative — approach. Only recently, image captioning models became zero-shot by decoupling image encoding — where pre-trained discriminative models like CLIP are used to create proper image and text representations — from the actual generative module. This is the case for models like Nukrai et al. (2022); Li et al. (2023b); Gu et al. (2023); Zeng et al. (2024); Fei et al. (2023); Yan et al. (2025); Zeng et al. (2025); Tewel et al. (2022); Su et al. (2022), which i) employ CLIP to leverage a shared vision-language semantic space, and ii) train a text decoder on solely text samples to recover the text back from the CLIP textual feature. This requires nothing more than a pre-trained contrastive model and a large set of sole text samples to craft a powerful captioner. In this paper, we show that our framework can generalize this core idea and that, under this formulation, many zero-shot captioners, paired with the right components, can be easily restructured to perform zero-shot region-based captioning. Therefore, we identify and study in detail the most critical components of this novel zero-shot region-based captioning framework. Particularly, we focus our attention on the pre-trained vision-language contrastive backbone, which should be able, unlike CLIP, to create meaningful patch representations. To this aim, we largely explore DINO-based (Caron et al.,2021;Oquab et al.,2024) variants, having better localized capabilities than CLIP. We further expand the study of our framework, addressing multiple modality-gap mitigation strategies that help the text decoder to be trained only on text without having access to paired image-text features, as well as studying different patch aggregation methods. By studying existing components and employing vision backbones able to output patch-level meaningful representations like DINO, we show that we can enable many zero-shot captioners to reach state-of-the-art or comparable results in many zero-shot captioning task variants requiring captioning sub-parts of the entire image — dense captioning (Johnson et al.,2016), region-set captioning (Cornia et al.,2019), up to the standard image captioning (Tewel et al.,2022) where the region to caption extends over the entire image. To better showcase the effectiveness of our framework in extreme patch-based captioning scenarios, we also introduce the zero-shot trace captioning task, which requires generating a caption for a region within an image specified by a mouse trace. To summarize, we propose the following contributions: a) We reformulate captioning by shifting perspective from the image-to-caption approach to a patch-to-caption one, unifying local and global tasks in Patch-ioner, a single framework which does not require region-level supervision, b) we repurpose existing models to work within this novel framework, by analyzing the role of the key components, with a special attention to the vision backbone, c) we show the performance of these models on four zero-shot captioning tasks that span different region granularity, from captioning few patches to the whole image, showing the effectiveness of the proposed perspective shift proposed by our framework despite its overall simplicity. 2 RELATED WORK Language-aligned Dense Image Representations are crucial for our goal of captioning at patch level. Vision-language models (VLM) like CLIP (Radford et al.,2021) introduced a powerful approach to learning global modality representations in a shared space via contrastive learning, paving the way to solve several downstream tasks, including captioning (Mokady et al.,2021;Cornia et al.,2024). However, in zero-shot settings, CLIP-like representations are known to struggle with dense tasks 2 Preprint “a grassy field” VLM Visual Encoder Decoder Language-aligned Dense Features Modality Gap Mitigation Patch-centric Framework Patch Embedding Patch Aggregation (optional) “many cars are parked in a parking lot” Mouse Trace Captioning Patch Aggr. “a grassy field” Patch Captioning “a man in the middle of playing a game of frisbee” Dense Captioning Patch Aggr. “a dog jumping to catch a frisbee in a park with a man on field” Image Captioning Patch Aggr. A B “a dog is jumping to catch a frisbee in the air” Region-Set Captioning Patch Aggr. Region Embedding Zero-shot Text Decoding Figure 1: Patch-ioner: a patch-centric framework for unified zero-shot captioning.A. Overview of our framework. First, we extract language-aligned dense patch embeddings from the image using a VLM. Given a region, we select the underlying patches and aggregate their features to obtain a region representation. Finally, we obtain the region caption by applying a zero-shot text decoder, that is a) conditioned on the latent region representation, b) trained on text-only data, and c) equipped with a mechanism to handle the modality gap present in vision-language common spaces. This enables regional captioning without requiring region-level supervision. B. By aggregating patch-level features from arbitrary image regions, we can flexibly handle multiple captioning tasks across spatial granularities in a unique model. due to misalignment between local visual patches and fine-grained semantics (Zhong et al.,2022; Ranasinghe et al.,2023;Bica et al.,2024). On the other hand, visual-only self-supervised models (SSM) like DINO (Caron et al.,2021;Oquab et al.,2024) excel in local semantic modeling but lack a bridge with language. Recent works like SILC (Naeem et al.,2024) and DINO.txt (Jose et al., 2024) attempt to get the best of both worlds by combining DINOand CLIP-like training objectives, aiming to obtain language-aligned dense representations. INViTE (Chen et al.,2024) modifies CLIP’s visual encoder by zeroing the attention weights from each patch to all the others in the last layers, leading to more semantic patch features. DenseCLIP (Rao et al.,2022) and RegionCLIP (Zhong et al., 2022) extend CLIP with additional region-level supervision. RegionCLIP leverages a region-proposal network to construct region-text pairs from image-text datasets, while DenseCLIP introduces a pixel-to-text matching loss to strengthen the alignment between local regions and textual concepts. Other methods instead exploit already existing VLMs and SSMs to get the same properties with minimal or no training: Talk2DINO (Barsellotti et al.,2024) connects language to the DINOv2 space by mapping CLIP textual representations to DINOv2 patches. ProxyCLIP (Lan et al.,2024) instead leverages DINO’s attention maps to improve the local properties of the CLIP visual embeddings of patches. Zero-shot Image Captioning methods rely mostly on global CLIP representations to guide text generation. Early-guided decoding methods take CLIP visual features as input and introduce adaptation techniques to reduce the visual-textual modality gap (Liang et al.,2022). DeCap (Li et al., 2023b) projects CLIP visual features into a more text-aligned space using a memory of texts as basis, while CapDec (Nukrai et al.,2022) and CLOSE (Gu et al.,2023) inject noise during text-only training to enable decoding also from the CLIP visual space. To improve the generation ViECap (Fei et al., 2023), MeaCap (Zeng et al.,2024), MERCap (Zeng et al.,2025) and EntroCap (Yan et al.,2025) leverage external knowledge to condition the decoding together with the CLIP image representation. Late-guided decoding methods instead use CLIP as a scoring or optimization signal rather than direct input. ZeroCap (Tewel et al.,2022) leverages CLIP gradients to steer the cached context during 3 Preprint text generation, while MAGIC (Su et al.,2022) optimizes token selection based on CLIP similarity scores. However, all the above approaches rely on global representations, which are not well-suited for capturing localized semantic details, making them suboptimal for patch-level or region-level captioning in zero-shot settings. Region-level Captioning comprises several tasks in which models are asked to produce natural language descriptions based on sub-parts of an image. They pose additional challenges as naively captioning the cropped regions or feature maps often induces a loss of the global context of the image and, thus, misinterpretation of the region. For this reason, zero-shot solutions to this family of problems are still underexplored. For controllable captioning (Cornia et al.,2019) — the generation of an image caption controlled by a set or sequence of regions — and dense captioning (Johnson et al.,2016) — the localization and captioning of salient regions of an image — state-of-the-art solutions like CAG-Net (Yin et al.,2019), GRiT (Wu et al.,2024), ControlCap (Zhao et al.,2024), and FlexCap (Dwibedi et al.,2025) provide good performance but need supervision with ground-truth boxes. Recent works (Guo et al.,2024), (Hua et al.,2025) moved towards the direction of arbitrary regions captioning exploiting region-level supervision, and the Localized Narratives dataset PontTuset et al. (2020a) — comprising images, timed captions, and timed mouse tracks — provide the ingredients for evaluating captioning also at trackor patch-level. We propose a unique framework to tackle captioning at various granularities, from imageto patch-level, in a zero-shot setting. 3 A PATCH-CENTRIC FRAMEWORK 3.1 MOTIVATION AND FORMULATION Learned region fusion requires regional data. Traditional regional captioning (Johnson et al., 2016;Dwibedi et al.,2025;Zhao et al.,2024) follows an early injection of region specification in the model. Formally, given an image I and a region R , the region caption t is modeled as t=D(I, R) . Such models require region-caption annotations and often use dedicated models or losses per captioning task or granularity. A step forward can be moved by introducing a formulation which disentangles image encoding and postpones region specification, defining t=D(ψ(I), R) , where ψ(I) provides a visual representation of the image independent of the region R , and D performs late region selection and text decoding. In order to avoid training the whole pipeline using region-level labels, we further decompose the decoder module D(·) into two distinct modules: a parameter-free fixed aggregation aggR of patch-level representations, and an actual text decoder ϕ(·) not directly conditioned on regions R , so that t=ϕ(aggR(ψ(I), R)) . We will detail these two factorized components in the following paragraphs. Parameter-free patch aggregation. Let I∈RH×W×3 be an image split into non-overlapping patches of size P×P . Each patch is encoded with a vision backbone ψv , yielding a dense grid of patch-level embeddings V=ψv(I)={vi}∈RH P×W P×D , where vi∈RD . Assuming that ψv(I) extracts this spatial grid of patch embeddings, a region definition R selects a subset of patch features that we aggregate to obtain the region embedding vR= aggR(ψv(I)) . We describe this aggregation as vS=Pi∈Swivi , where S is the set of indices of patches that underlie the region, and wi are aggregation weights. While several aggregation functions can be considered (e.g., uniform, gaussian, attention-based), we found that the specific choice has limited impact in regional captioning (see SM§8) and report results with mean aggregation ( wi=1/|S| ). Using a set operator as aggregation gives us the flexibility to aggregate arbitrary sets of patches, and thus, define regions as boxes, masks, traces, single patches, or full-image grids. Empirically, we also find that not all ψv are suitable to extracting patch-level meaningful visual representations; transformer architectures pretrained with dense local contrastive objectives (such as DINO) are significantly more robust at the patch level, as validated in §4.2. Zero-shot decoding. We train a text decoder ϕ:RD→ T using a prefix language modeling approach, where the decoder reconstructs a caption t∈ T from its text embedding ψt(t) . If ψt(t) is aligned to the visual encoder ψv(I) , we could directly decode the visual embeddings using a text-only decoder trained on text-only data. In such a case, we can finally decode the region representation into a natural language caption t=ϕ(vR) . However, the assumption that ϕ(·) can digest features from ψv(I) while being trained to reconstruct features from ψt(t) is often too optimistic due to the 4 Preprint prominent modality gap (Liang et al.,2022). In fact, text and image representations, despite being in the same multimodal space, occupy different, separated subspaces. To obtain a decoder applicable to visual embeddings, we experimented with two mitigation strategies for this gap. The first strategy, following Li et al. (2023b), introduces a projection step at inference that maps visual features into the text subspace using a memory of text embeddings. The second, inspired by Gu et al. (2023) and Nukrai et al. (2022), trains the decoder under input perturbations, encouraging robustness so that it can directly process visual embeddings. We analyze the effects of the modality gap mitigation strategies in SM§9. Overall, our formulation offers three main advantages: a) it is zero-shot by design, in the sense that it only requires image-level textual descriptions to be trained and does not require paired image-text samples to train the text decoder, b) any region (e.g., whole image, box, mask, free-form trace, single point) is addressed identically, supporting a modular, general-purpose regional captioning pipeline, and c) our method requires only a single forward pass of the vision backbone to extract patch features for an entire image, which can then be reused to caption multiple regions without rerunning the full pipeline for each one. 3.2 FROM PATCHES TO REGIONS Building on patch-level zero-shot captioning, our framework can generate captions for arbitrary regions of an image. A region is defined as a set of patches, and its representation is obtained by averaging the embeddings of its constituent patches. This simple formulation unifies several existing region-level captioning tasks, which differ only in how the relevant patches are selected, allowing us to address them without task-specific modifications. In the following, we describe the tasks considered in our evaluation and their induced patch selections. Image Captioning involves generating a single caption that describes the entire image. To achieve this, we derive a global representation vI= avgi(vi) by aggregating the feature embeddings of all patches {vi}within the image I. Dense Captioning requires locating salient regions in an image and generating their descriptions. Following defined evaluation protocols (Johnson et al.,2016), we focus on captioning already defined boxes, effectively removing the localization subtask, which can be tackled using additional regionproposal models. Given a bounding box B and the set SB of indexes of patches that intersect with B , we obtain the representation of the region vB= avgi∈SB(vi). Region-set Captioning consists of generating a single caption for multiple regions within an image, where each region is specified by a distinct bounding box. Given an image I and a set of bounding boxes B={B1, B2, . . . , BK} , we define SBk as the set of patches that intersect with the k -th bounding box in B . To represent the entire set of regions, we aggregate the feature embeddings from all selected patches across all bounding boxes, which results in a combined region-level representation vB=1 PB∈B|SB|X B∈B X i∈SB vi.(1) Note that if a patch appears in more than one box, it is weighted more in the average. Trace Captioning. To demonstrate the flexibility of our approach, we introduce Trace Captioning, a novel task in which the region of interest is specified by a mouse trace T={p1, . . . , pL} with L points. Each point pj is mapped to the corresponding patch index ij , yielding the sequence ST= [i1, . . . , iL] . The trace-level representation is then obtained by averaging the embeddings of the selected patches vT=1 L L X j=1 vij.(2) Unlike box-based settings, this formulation allows for free-form, user-specified regions and thus enables interactive and fine-grained localized descriptions, expanding the scope of region-level captioning. 5 Preprint 4 EXPERIMENTS In this section, we first quantitatively assess the role of the backbone in our novel framework (§4.2). Then, we measure the performance of our new formulation over state-of-the-art zero-shot captioning methods, evaluating them region-specific zero-shot tasks (§4.3). Other interesting yet less influential studies — like the role of the modality-gap mitigation strategy and the patch aggregation operator aggR(·)— are available in SM §8and §9. To create a common ground, we first introduce the employed datasets and metrics in the following section. 4.1 DATASETS AND METRICS Trace Captioning. We build a benchmark for Trace Captioning exploiting Localized Narratives (Pont-Tuset et al.,2020b) — a dataset in which annotators vocally described objects in images while moving the mouse pointer over the described object. The dataset provides temporal annotated voice transcriptions and mouse traces for the images of many standard captioning datasets. We took the labeled COCO (Lin et al.,2014;Chen et al.,2015) with splits defined in Karpathy & Fei-Fei (2015) and Flickr30K 1 (Young et al.,2014) test subsets to build the trace captioning evaluation datasets. We split long traces and transcriptions for each image into sentences, and we discard the parts of the traces that are not temporally located between the start and the end of each sentence. We discarded noisy sentences — such as the ones describing image properties (The image is blurred,the image is edited, ...) and rewrote each sentence removing uncertainties typical of voice descriptions in a more concise and caption-like style through a few-shots prompted LLM (LLama 3 Dubey et al. (2024)). After annotation cleaning, 51 COCO images resulted without clean sentences and were discarded. Each sub-trace and relative sentence comprise an independent sample, that is, we ignore the temporal information of consecutive sentences and focus on evaluating the description of each sub-trace only. Samples and additional details are available in §10. Dense Captioning. We assess the performance on dense captioning tasks following the evaluation procedure of Johnson et al. (2016), omitting the bounding box proposal and evaluating only the bounding box captioning task, using ground-truth boxes as input for the models. In addition to standard caption metrics, for this task, we also report the mAP as originally defined by Johnson et al. (2016). We use the Visual Genome (VG) v1.2 (Johnson et al.,2016;Krishna et al.,2017) and VG-COCO test splits (Li et al.,2019). The former comprises 5000 images from VG, while the latter contains 2476 images present in both VG and COCO. Both contain multiple bounding box annotations per image with descriptions. Region-Set Captioning We follow the evaluation protocol of Cornia et al. (2019) that originally introduced region-set captioning. We use the Flickr30K Entities (Plummer et al.,2015) and the COCO Entities (Cornia et al.,2019) datasets. Each record comprises an image, a set of bounding boxes of variable length, and a ground-truth controlled caption. We evaluate on the test splits, comprising of images in the Karpathy & Fei-Fei (2015) test splits, that consist of 3569 and 1000 images for COCO and Flickr30k versions, respectively. Image Captioning We follow the standard evaluation pipeline for zero-shot image captioning, generating captions for the 5000 images in Karpathy’s COCO test split. We compare with these state-of-the-art models: DeCap (Li et al.,2023b), CLOSE (Gu et al.,2023), ZeroCap (Tewel et al., 2022), MAGIC (Su et al.,2022), ViECap (Fei et al.,2023), CapDec (Nukrai et al.,2022), EntroCap (Yan et al.,2025), MeaCap 2 (Zeng et al.,2024), and MERCap (Zeng et al.,2025). For DeCap, ViECap, and MeaCap, we also report the best results reproduced by us. Metrics. All datasets used in our evaluation provide a ground-truth caption for each annotation. We therefore adopt standard captioning metrics to assess the similarity between generated and reference 1 Licenses for COCO and Visual Genome are CC-BY 4.0. Flickr30k images are under the terms of Flickr’s user agreement. 2 We picked the MeaCap InvLM , which uses a GPT-2 decoder and achieves the highest scores in the zero-shot captioning task. 6 Preprint Captioning Task: (Dataset) Trace (COCO) Dense (VG v1.2) Region-Set (COCO Entities) Image (COCO) Encoder Backbone C P C P C P C P CLIP-S CLIP CLIP B/16 10.9 75.0 10.9 74.2 41.6 78.8 42.1 84.0 66.2 DenseCLIP CLIP B/16 18.6 75.3 19.9 75.2 51.0 77.6 28.0 77.0 57.3 INViTE CLIP B/16 13.8 76.4 16.8 77.3 43.3 78.9 21.3 79.1 60.6 ProxyCLIP DINO B/8 + CLIP B/16 16.7 75.7 15.7 76.0 41.2 78.4 28.7 79.0 61.7 ProxyCLIP DINOv2 B/14 + CLIP B/16 16.5 75.7 15.5 76.0 40.6 78.5 27.4 78.6 61.0 DINO.txt DINOv2 B/14 23.2 78.8 23.4 78.5 91.8 86.3 67.8 87.2 70.8 Talk2DINO DINOv2 B/14 27.9 78.7 31.9 78.8 109.1 87.5 69.2 87.4 72.8 Table 1: Vision-Language Backbones. CIDEr (C) and RefPAC-S (P) across four captioning tasks. captions. In particular, we focus on CIDEr (C) (Vedantam et al.,2015), which captures syntactic overlap, and RefPAC Score (P) (Sarto et al.,2023), a more recent metric that quantifies semantic similarity independently of caption phrasing. For completeness, results with other traditionally used metrics — BLEU@4 (Papineni et al.,2002), METEOR (Banerjee & Lavie,2005), ROUGE-L (Lin, 2004), and SPICE (Anderson et al.,2016) — are reported in supplementary materials, as they follow the same trends. For the image captioning task, we additionally report CLIP-Score (Hessel et al., 2021), which measures the alignment between an image and its generated caption in the joint CLIP vision–language space. This metric is not applicable to region-set, trace, or dense captioning tasks, where captions describe local regions rather than the entire image. 4.2 BACKBONE SELECTION The choice of the visual backbone is crucial for our patch-centric framework, as the quality and semantic richness of the patch features directly impact the captioning performance. We tested several state-of-the-art vision-language models pre-trained without region-level supervision and evaluated their effectiveness within the framework. We tested vanilla CLIP (Radford et al.,2021), three CLIP adaptations for dense tasks — DenseCLIP (Rao et al.,2022), INViTE (Chen et al.,2024), and ProxyCLIP (Lan et al.,2024) —, and two methods with visual encoders based on DINOv2 (Oquab et al.,2023)—DINO.txt (Jose et al.,2025) and Talk2DINO (Barsellotti et al.,2024). Patches are aggregated as described in §3.2, while for the zero-shot decoder, we align with the setting of Li et al. (2023b), and use a prefix GPT-2 style decoder (SM§6reports implementation details), with a memory-based latent projection approach as mitigation strategy for handling the modality gap. Specifically, before decoding, the region representation v is projected into the text embedding space as a similarity-weighted linear combination of memory elements, vproj =M α with α= softmax1 τM⊤v , where M= [m1,...,mN] stores the text embeddings mj=ψt(tj) and τ > 0 controls the sharpness of the weighting distribution. This choice enables us to be directly comparable with other works using the same architecture in the all the following experiments, besides also performing the best among the tested zero-shot decoder methods (see SM§9). In SM§8, we also study how the choice of the memory bank M affects the captioning performance, providing an upper bound to metrics. Table 1 shows that backbone effectiveness in our framework is closely tied to capturing fine-grained local semantics. Standard CLIP performs poorly, indicating that its patch tokens lack the spatial detail needed for our tasks (Mukhoti et al.,2023;Ranasinghe et al.,2023;Bica et al.,2024). Backbones that strengthen CLIP local representations, such as INViTE and DenseCLIP, achieve stronger results, supporting our hypothesis. The best performance comes from DINOv2-based models, including DINO.txt and Talk2DINO, with the latter emerging as the most effective encoder. This underscores the importance of semantically rich patch-level features for high-quality region-level captions. For this reason, we show results using Talk2DINO as the default backbone in subsequent experiments. 4.3 COMPARISON WITH SOTA Despite significant advances in zero-shot image captioning, we are unaware of any prior methods specifically tailored for zero-shot regional captioning tasks. Existing zero-shot captioners are usually evaluated only at the level of whole images, without any mechanisms to natively attend to arbitrary 7 Preprint regions. To rigorously quantify the benefit of our Patch-ioner framework, in Table 2, we compare against both state-of-the-art zero-shot image captioners and adapted baselines: (i) state-of-the-art whole-image zero-shot captioners in their standard setting or applied to region crops, simulating regional captioning by isolating local content, and (ii) region-supervised encoders, thus outside our no-region-label setting, that leverage mask-based (AlphaCLIP, Sun et al. (2024)) or crop-based (RegionCLIP, Zhong et al. (2022)) attention coupled with the same zero-shot decoder, allowing them to attend to specific regions. This design ensures that our evaluation covers the strongest available baselines for both image-level and region-level zero-shot captioning. Note that we do not compare with large multimodal models tackling regional understanding and captioning, such as Guo et al. (2024); Hua et al. (2025), as they are outside of our assumptions on the available supervision by training on region-level data. In addition to our strongest model (i.e., Talk2DINO with the memory-based mitigation strategy), we also report other combinations that express existing zero-shot captioning approaches (CLOSE (Gu et al.,2023), CapDec (Nukrai et al.,2022), VieCap (Fei et al.,2023), and MeaCap (Zeng et al.,2024)) but replacing the original CLIP backbone. Specifically, CLOSE and CapDec can be expressed by choosing the noise-injection mitigation strategy and using the standard decoder pipeline described in §4.2. The same applies to ViECap and MeaCap, although with the addition of extra knowledge in the text decoding step: ViECap uses external entity-aware prompts, while MeaCap leverages also structured concept retrieval from a knowledge base. We use two datasets for each task — a COCO-derived dataset and an additional dataset such as Visual Genome (Krishna et al.,2017) or Flickr30k (Young et al.,2014). Figure 6shows qualitative results. Patch-centric captioning excels in local, fine-grained tasks. In the trace and dense captioning tasks, which emphasize local visual content, our patch-centric framework significantly outperforms all baselines across metrics. For trace captioning (Table 2, first group), our patch-based formulation outperforms whole-image captioners and crop-based adaptations. Models relying on global CLS representations fail to capture the precise objects and attributes under the trace. Even AlphaCLIP, that is a region-supervised backbone that can be naturally applied to traces, lags behind, underlying intrinsic limitations of the pretrained CLIP backbone. Dense captioning shows a similar trend (Table 2, second group), with our models outperforming baselines. For this task, we report crop-based adaptations for DeCap, ViECap, and MeaCap, as they provide a stronger baseline with respect to the same models applied to the global CLS (see Table 4 in SM). Although isolating regional content, these models discard broader contextual cues that are crucial for coherent dense descriptions. Patch aggregation extends seamlessy to context-aware captioning. On region-set captioning (Table 2, third group), our models once again achieves state-of-the-art results, outperforming both zero-shot baselines and region-supervised models. Note that the region-set captioning task tends to align more closely with image-level captioning rather than strictly focusing on localized regions (see Figure 6 and Figure 7 in SM), as regions are intended to control an image-level caption3. Thus, global models also tend to perform well on those tasks, showing a narrower gap with respect to our new state of the art. By aggregating patch embeddings from arbitrary and possibly disjoint sets of regions, the model produces coherent and contextually rich captions that align with the collective semantics of the chosen areas. In contrast, whole-image methods, including ViECap and DeCap, cannot naturally incorporate regional cues, limiting their effectiveness in this setting. Importantly, our patch-based aggregation even surpasses AlphaCLIP, which relies on explicit mask supervision. Patch-centric models deliver comparable performance on whole-image captioning. In wholeimage captioning (Table 2, fourth group), our results remain competitive with the strongest zero-shot captioners but are slightly behind dedicated image-centric architectures such as MERCap (Zeng et al.,2025) and EntroCap (Yan et al.,2025). For this task, we report the performance of our models using an attention-based weighting, when helpful, as it usually performs marginally better than the standard average patch aggregation and thus represents the best available model for this task and reaches the smaller gap with state-of-the-art models (see SM§8). Notably, adding structured external knowledge or filtered retrieval (as in ViECap or MeaCap) on noise-based decoders improves fluency 3 This is expected since the ground-truth captions in the COCO Entities dataset originate from the image-level annotations of COCO, as stated by Cornia et al. (2019). 8 Preprint Trace Captioning Dense Captioning Region-Set Captioning Image Captioning COCO Flickr30k VG v1.2 VG-COCO COCO Entities Flickr30k Entities COCO Flickr30k Model C P C P mAP C P mAP C P C P C P C P CLIP-S C P CLIP-S Whole-image Zero-shot Captioners ZeroCap (Tewel et al.,2022)CVPR’22 - - - - - - - - - - - - - - 14.6 - - - - - MAGIC (Su et al.,2022)ArXiv’22 - - - - - - - - - - - - - - 49.3 - - 17.5 - - CLOSE (Gu et al.,2023)ICCV’23 - - - - - - - - - - - - - - 81.2 - - - - - CapDec (Nukrai et al.,2022)EMNLP’22 - - - - - - - - - - - - - - 91.8 - - 35.7 - - EntroCap (Yan et al.,2025)NEUCOM’25 - - - - - - - - - - - - - - 94.3 - - 41.5 - - MERCap (Zeng et al.,2025)AAAI’25 - - - - - - - - - - - - - - 96.0 - - 45.6 - - ViECap†(Fei et al.,2023)ICCV’23 24.3 74.4 12.0 68.8 14.90⋄26.4⋄74.3⋄15.2⋄26.6⋄74.3⋄102.7 85.0 31.8 74.9 89.7 88.5 75.6 29.8 80.6 70.5 MeaCap†(Zeng et al.,2024)CVPR’24 22.5 74.4 12.6 69.8 15.01⋄28.6⋄75.1⋄16.0⋄28.9⋄75.1⋄97.9 85.2 38.6 76.4 86.0 88.6 77.8 40.1 82.8 73.9 DeCap†(Li et al.,2023b)ICLR’23 20.5 75.3 11.2 71.0 17.75⋄24.6⋄77.8⋄17.8⋄24.9⋄77.7⋄95.1 87.4 39.4 78.8 87.4 90.6 79.3 40.0 84.8 76.3 With Region-level Supervision RegionCLIP (Zhong et al.,2022)CVPR’22 + Mem. (≃DeCap) - - - - 15.85 21.7 76.7 16.01 21.0 75.4 - - - - 93.4 91.2 77.5 38.8 84.5 73.6 AlphaCLIP (Sun et al.,2024)CVPR’24 + Mem. (≃DeCap) 21.3 75.4 11.8 71.0 14.63 19.1 73.9 14.82 19.4 73.8 95.1 87.4 39.5 78.8 89.7 91.1 78.2 40.9 85.4 75.0 Patch-ioner (Our Patch-based Framework) T2D + Mem. (≃DeCap) 27.9 78.7 18.8 77.0 21.31 31.9 78.8 21.53 32.3 78.7 109.1 87.5 44.1 79.1 88.5♢90.2♢76.0♢39.3♢84.2♢71.8♢ T2D + Noise (≃CLOSE, CapDec) 29.3 78.1 19.3 75.6 20.26 26.3 77.0 20.33 26.4 76.9 97.5 85.6 37.1 76.5 65.5†86.2†70.9†27.8†80.8†67.0† T2D + Noise + External knowledge (≃ViECap) 28.2 78.2 18.5 76.2 18.43 30.3 77.8 18.43 30.7 77.7 109.3 86.7 37.8 77.8 88.5♢89.2♢73.7♢34.1♢82.8♢69.9♢ T2D + Noise + Filtered knowledge (≃MeaCap) 27.4 78.8 20.3 77.3 18.66 31.9 78.9 19.43 32.3 78.7 104.4 86.9 42.3 78.6 83.0♢89.6♢74.8♢39.4♢84.4♢71.4♢ †: reproduced by us. ⋄: Model applied to image crops. ♢: attention-based weighting. Table 2: Comparison of our Patch-ioner framework, using Talk2DINO (T2D), with state-of-theart zero-shot captioning methods on trace, dense, region-set, and image captioning tasks. Our approach consistently outperforms whole-image and region-supervised baselines in local, fine-grained captioning tasks, while achieving competitive results on whole-image captioning. The table reports CIDEr (C), RefPAC (P), mean average precision (mAP) for dense captioning, and CLIP-Score (CLIPS) when applicable; best and second-best results are highlighted in bold and underline , respectively. Patch Trace Dense Region-Set Image DeCap: a cat is sleeping on a cluttered desk. Ours (CLIP + Mem.): a cat is sitting on the bed and it’s contents. Ours (Talk2DINO + Mem.): a plant in a vase sitting on a table. DeCap: a giraffe in a zoo with a city in the background. Ours (CLIP + Mem.): there are some people that are out by a lot of trees. Ours (Talk2DINO + Mem.): a view of a city with a sky in the background. GT: A sky. DeCap: a woman squatting on a bench with a cat. Ours (CLIP + Mem.): there is a person that is out on the kitchen. DeCap (Crop): a a close up of a person standing by a person holding a phone. Ours (Talk2DINO + Mem.): a black cat is sitting on a black bench. GT: black cat sitting on a bench. DeCap: a baseball player at bat getting ready to hit the ball. Ours (CLIP + Mem.): some baseball players are on the field playing baseball. Ours (Talk2DINO + Mem.): a baseball player is swinging his bat as a crowd watches. GT: a man swinging a baseball bat as another looks on. DeCap: a bench sits on the beach next to the ocean. ZeroCap: a beachfront bench. ViECap: A wooden bench sitting on top of a sandy beach. Ours (Talk2DINO + Mem.): a bench sitting on the beach next to the ocean. GT: A wooden bench sitting on a beach. Figure 2: Qualitative results from finer (left) to coarser (right) tasks. Note the discrepancy of predicted and ground-truth captions when an image-level (DeCap, DeCap (Crop)) or a CLIP-based regional (CLIP + Mem.) captioner is applied, with respect to our Talk2DINO-based model. and informativeness, suggesting such modules are complementary than substitutes for strong regional semantics. However, they compare similarly to the model using the memory-based decoder. 5 CONCLUSIONS We introduced Patch-ioner, a novel zero-shot captioning framework that shifts from an imagecentric to a patch-centric approach, enabling caption generation for individual patches and arbitrary aggregations without any region or even image supervision. We rely on the strong spatial awareness of DINOv2, whose local image patches have been effectively bridged with the text modality. Thanks to the disentangled training of the decoder network, this flexible and scalable method sets the new state of the art on various regional captioning tasks, including dense and region-based captioning, as well as our newly proposed trace captioning. Despite its simplicity, results show that our patch-centric approach can effectively bridge the gap between local and global understanding in image captioning, providing a unified framework for multigranularity captioning tasks in a zero-shot setting. Moreover, our models require a single backbone forward pass to caption multiple regions, facilitating practical viability in interactive applications. Limitations and Future Work. Despite strong zero-shot performance, our model still lags behind fully supervised, task-specific approaches. The contextual scope of each patch is fixed by the backbone and is not adjusted to meet the user intents. On top of that, the modality jump introduces noise that can cause hallucinations. Future work could incorporate weak supervision, e.g., image-level 9 Preprint Model # Patches Backbone Input Weighting B M R C S P Image-based DeCap@224 7 CLIP B32 CLS - 2.1 9.7 21.7 21.1 8.8 75.2 DeCap@224 14 CLIP B16 CLS - 2.2 9.8 21.8 21.3 8.7 75.4 DeCap@592 37 CLIP B16 CLS - 2.0 9.6 21.5 20.5 8.7 75.3 ViECap@224 7 CLIP B32 CLS - 2.5 9.8 22.4 24.8 9.3 74.1 ViECap@224 14 CLIP B16 CLS - 2.5 9.9 22.3 24.7 9.5 74.5 ViECap@592 37 CLIP B16 CLS - 2.3 9.6 22.1 24.3 9.5 74.4 MeaCap@224 7 CLIP B32 CLS - 2.3 9.4 21.5 23.1 8.9 74.2 MeaCap@224 14 CLIP B16 CLS - 2.3 9.3 20.8 23.4 9.0 74.6 MeaCap@592 37 CLIP B16 CLS - 2.2 9.1 20.3 22.5 9.0 74.4 Patch-based (Our Framework) T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches uniform 2.5 10.7 23.2 27.9 12.6 78.7 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches attention 2.4 10.4 22.7 27.6 12.0 78.1 Table 3: Trace Captioning results on COCO test set. Model # Patches Backbone Input Weighting mAP M B R C S P Image-based DeCap@224 7 CLIP B32 CLS - 0.15 8.40 0.94 15.61 19.38 9.38 73.71 DeCap@224 14 CLIP B16 CLS - 0.14 8.48 0.95 15.70 19.11 9.40 73.94 DeCap@592 37 CLIP B16 CLS - 0.15 8.37 0.92 15.67 18.53 9.26 73.91 ViECap@224 7 CLIP B32 CLS - 0.13 8.25 1.02 16.06 24.18 9.97 73.03 ViECap@224 14 CLIP B16 CLS - 0.14 8.30 1.01 15.86 23.81 9.91 73.49 ViECap@592 37 CLIP B16 CLS - 0.14 8.17 1.00 15.86 23.26 9.82 73.35 MeaCap@224 7 CLIP B32 CLS - 0.13 8.04 0.98 15.40 23.22 9.66 72.97 MeaCap@224 14 CLIP B16 CLS - 0.13 8.01 1.03 15.15 23.37 9.49 73.55 MeaCap@592 37 CLIP B16 CLS - 0.13 7.86 0.99 15.13 22.77 9.43 73.50 Region-based DeCap@224 Crop 7 CLIP B32 CLS - 0.17 10.03 1.35 18.20 23.61 10.90 77.09 DeCap@224 Crop 14 CLIP B16 CLS - 0.18 10.33 1.40 18.44 24.56 11.28 77.76 DeCap@592 Crop 37 CLIP B16 CLS - 0.14 8.30 1.05 16.39 17.20 7.78 75.47 ViECap@224 Crop 7 CLIP B32 CLS - 0.15 9.32 1.42 17.79 26.40 10.07 74.34 ViECap@224 Crop 14 CLIP B16 CLS - 0.16 9.59 1.46 18.03 27.13 10.43 75.62 ViECap@592 Crop 37 CLIP B16 CLS - 0.12 7.83 1.14 16.02 20.20 7.44 73.26 MeaCap@224 Crop 7 CLIP B32 CLS - 0.15 9.64 1.46 18.00 28.62 10.98 75.08 MeaCap@224 Crop 14 CLIP B16 CLS - 0.16 10.03 1.57 18.45 30.53 11.51 76.35 MeaCap@592 Crop 37 CLIP B16 CLS - 0.12 7.93 1.19 16.17 21.32 7.86 73.69 Patch-based (Our Framework) T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches uniform 0.21 10.63 1.36 18.59 31.94 15.03 78.82 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches gaussian 0.22 10.82 1.43 18.82 32.80 15.48 79.14 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches attention 0.21 10.31 1.27 18.17 30.58 14.72 78.69 Table 4: Dense Captioning results on VG v1.2 test set. Model # Patches Backbone Input Weighting B M R C S P Image-based DeCap@224 7 CLIP B32 CLS - 10.1 19.0 38.0 94.4 26.4 86.9 DeCap@224 14 CLIP B16 CLS - 10.0 19.4 38.3 95.1 26.8 87.4 DeCap@592 37 CLIP B16 CLS - 9.6 18.6 37.5 91.4 25.9 86.7 ViECap@224 7 CLIP B32 CLS - 11.2 18.2 38.9 102.7 27.0 85.0 ViECap@224 14 CLIP B16 CLS - 11.3 18.3 38.6 102.2 26.9 85.4 ViECap@592 37 CLIP B16 CLS - 10.8 17.8 37.9 99.2 26.5 85.0 MeaCap@224 7 CLIP B32 CLS - 10.4 17.7 37.0 97.9 25.9 85.2 MeaCap@224 14 CLIP B16 CLS - 10.1 17.5 35.5 96.5 25.7 85.4 MeaCap@592 37 CLIP B16 CLS - 9.3 16.9 34.6 91.1 25.5 85.0 Patch-based (Our Framework) T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 CLS - 9.1 16.9 35.0 89.4 25.4 85.5 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches uniform 11.5 19.3 38.8 109.1 29.4 87.5 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches gaussian 11.6 19.6 39.3 111.6 30.1 87.7 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches attention 11.0 19.0 38.3 107.0 29.3 87.4 Table 5: Region-Set Captioning results for COCO Entities test set. 3 Preprint Model # Patches Backbone Input Weighting B M R C S P Image-based DeCap@224 7 CLIP B32 CLS - 23.46 25.12 50.06 87.40 19.14 90.58 DeCap@224 14 CLIP B16 CLS - 23.89 25.51 50.34 89.64 19.52 91.05 DeCap@592 37 CLIP B16 CLS - 22.43 24.64 49.25 84.57 18.66 90.36 ViECap @224 7 CLIP B32 CLS - 26.70 23.99 50.85 89.67 17.54 88.45 ViECap@224 14 CLIP B16 CLS - 26.3 24.0 50.3 89.5 17.6 88.8 ViECap @592 37 CLIP B16 CLS - 25.60 23.38 49.52 86.84 17.08 88.33 MeaCap@224 7 CLIP B32 CLS - 24.57 23.12 47.68 86.66 17.27 88.76 MeaCap@224 14 CLIP B16 CLS - 23.6 22.7 45.5 85.1 17.3 89.0 MeaCap@592 37 CLIP B16 CLS - 22.01 21.87 44.72 80.84 16.69 88.41 Patch-based (Our Framework) T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches central patch 15.68 18.46 40.84 55.53 12.66 84.26 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches uniform 19.52 21.49 44.88 69.19 15.59 87.36 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches gaussian 21.17 22.62 46.62 76.79 16.73 88.36 T2D + Mem. (≃DeCap) @518 37 DINOv2 B14 Patches attention 23.64 23.93 48.54 88.46 18.21 90.21 T2D + Mem. (≃DeCap) @518 GT Memory 37 DINOv2 B14 CLS - 23.58 23.54 47.71 85.67 17.86 89.53 T2D + Mem. (≃DeCap) @518 GT Memory 37 DINOv2 B14 Patches attention 25.66 24.77 49.83 93.87 19.09 90.70 Table 6: Image Captioning results on COCO test set. Image Captioning. In Table 6, we report the results of standard zero-shot image captioning. In addition to the already described weighting schemes, we test one additional configuration for our framework that is central patch, where the decoding is applied to the central patch of the image. We can observe that the most effective strategy for the image captioning task is attention. This is coherent with results from Barsellotti et al. (2024), where they suggest the attention-weighted patch means to use Talk2DINO for global tasks such as image-text retrieval. Memory Bank. Considering that in the memory-based model of our framework (that is similar to Li et al. (2023b)) we tackle the modality gap through a projection based on a collection of texts, we tested how much the selection of the texts in the memory bank influences performance. In Table 6, we also report the results obtained by that model when in its memory bank there are also ground-truth captions of the test set (rows marked with GT Memory). This provides a sort of upper bound to performance when varying the text collection used as memory. We observe that in this configuration, the performance only slightly improves (+0.5%), indicating that the model is robust to the choice of the memory bank. 9 MODALITY GAP: PROJECTION TO TEXTUAL SPACE VS TRAINING WITH NOISE In this section, we quantitatively assess the performance of two state-of-the-art solutions to overcome the modality gap. In particular, we compared the configuration based on a memory bank of texts — the one introduced in §3— with an alternative solution based on noise injection during the decoder training. Additionally, we include in our comparison a baseline with no modality gap mitigation (no mitig.), to highlight the benefits brought by each strategy. Training with Noise. Various works Gu et al. (2023); Nukrai et al. (2022) proposed zero-shot image captioning solutions based on noise injection during the training of the text decoder. Through this strategy, the trained decoders are more effective in understanding semantic representations, even when those are not coming from the text modality. To implement this strategy in our framework, we trained the textual decoder on the same collection of captions as for the memory bank-based configuration. We adopted Talk2DINO Barsellotti et al. (2024) textual space for the decoder input space, which is aligned to DINOv2 Oquab et al. (2023) with registers Jose et al. (2024). Following the setting of Gu et al. (2023), we added Gaussian noise with σ2= 0.08 to the textual embeddings while leaving the other parameters unchanged (as defined in §6). In the next paragraphs, we report and compare the results for each task of Talk2DINO within our framework with the memory bank (Memory) and with the training with noise (Noise). In Table 7, we compare the two modality gap mitigation strategies across multiple captioning tasks, and also report the performance of a baseline without any mitigation (no mitig.). The baseline consistently underperforms compared to both the Memory and N oise configurations, indicating that, like other contrastively learned image-text encoders Liang et al. (2022), Talk2DINO is also affected by the modality gap. These results highlight the importance of explicitly addressing this gap to 4 Preprint Table 7: Mitigation of Modality Gap. Comparison of Memory-based Projection ( Memory ) vs Noise-trained Decoder (Noise) across tasks. Trace Captioning (COCO) Dense Captioning (VG v1.2) Region-Set Captioning (COCO Entities) Image Captioning (COCO) Mitigation B M R C S P mAP M B R C S P B M R C S P B M R C S P CLIP-S no mitig. 1.2 9.1 18.3 14.7 8.5 75.1 0.18 9.7 0.7 15.9 17.8 10.2 75.2 5.0 15.0 29.4 59.4 21.1 82.2 9.9 17.7 36.8 43.7 12.3 82.2 69.6 Noise 3.0 11.5 24.7 29.3 12.3 78.1 0.20 10.4 1.2 17.8 26.3 12.6 77.0 10.5 18.4 37.2 97.5 26.7 85.6 19.6 21.5 45.4 65.5 15.5 86.2 70.9 Memory 2.5 10.7 23.2 27.9 12.6 78.7 0.21 10.6 1.4 18.6 31.9 15.0 78.8 11.5 19.3 38.8 109.1 29.4 87.5 19.5 21.5 44.9 69.2 15.6 87.4 72.8 achieve strong captioning performance. For Trace Captioning, the Memory method is slightly more effective in the semantic metric RefPAC-S, while the N oise variant achieves marginally better scores in CIDEr, ROUGE-L, METEOR, and BLEU@4, with a minimal gap between the two approaches. In Dense Captioning, the Memory model consistently outperforms the N oise model across all metrics. Similarly, for Region-Set Captioning, both methods achieve strong results, but the Memory method shows a clearer advantage, particularly in tasks closer to the patch level. Finally, in Image Captioning, the performance gap between the two architectures narrows, especially on the Flickr30k test split. In this scenario, the Memory method performs significantly better when applied to the CLS token, whereas patch aggregation produces comparable results. However, the metrics reveal conflicting trends across different datasets. Chosen Strategy. Based on the observed results, we selected the projection-based approach ( Memory ) as the primary strategy for overcoming the modality gap in our framework. While the noise injection method ( N oise) yielded competitive performance across multiple tasks, the Memory method demonstrated superior performance in dense captioning and region-set captioning, as well as a clear advantage when applied to the CLS token in image captioning. Given these trends, and considering the stability of the projection-based approach across different evaluation settings, we adopted Memory as the default configuration for our framework. 10 TRACE CAPTIONING BENCHMARK GENERATION We construct our Trace Captioning dataset from the Localized Narratives dataset Pont-Tuset et al. (2020a). This dataset consists of mouse traces and their corresponding speech transcriptions, where annotators describe objects in images while moving the mouse pointer over them. The initial dataset samples include timestamped mouse traces and are composed of multiple sentences that thoroughly describe the trace, with the generated descriptions following the order of the mouse movement. However, our task does not require strict temporal coherence. Instead, we aim to generate a single, concise caption that describes the specific area covered by the localized trace, rather than a multi-sentence description. To achieve this, we split the descriptions into individual sentences and align the traces accordingly. We then refine the traces by removing intermediate periods caused by transitions between sentences, which often occur when the annotator moves to a different region of the image. Specifically, we trim each trace by removing the first and last 15% of points, eliminating these transitional segments. Furthermore, we refine the captions by prompting the Llama3 8B model to rephrase the sentences, removing vague or subjective phrases such as "there is," "we can see," or "on the left of the image," and replacing them with concise, objective descriptions that refer specifically to the region covered by the trace. This rephrasing is crucial to ensure that each caption adheres to the standard format of image-captioning datasets and focuses only on the precise part of the image that the trace corresponds to. The LLM also helps identify and remove irrelevant sentences (e.g., "the image is blurred," "the image is edited"), which are then discarded along with their associated traces from the final benchmark. Figure 4shows the full prompt used to guide the Llama model in refining and cleaning the descriptions. Figure 5illustrates how the initial narrative samples are transformed into final trace captioning samples through the process of trace splitting and caption rephrasing. 5 Preprint 11 MORE QUALITATIVE RESULTS Additional qualitative results are shown in Figures 6and 7. Note that the first rows of Figures 6and 7 contain also qualitative results for single patch captioning, for which we do not have annotated data to report quantitative results. As can be noticed in Figures 6and 7, the Region-Set Captioning task tends to align more closely with image-level captioning rather than strictly focusing on localized regions. This is expected since the ground-truth captions in the COCO Entities dataset originate from the image-level annotations of COCO, as stated in Cornia et al. (2019). 6 Preprint I have image descriptions derived from spoken narratives. These need to be rewritten as concise, stand-alone captions in the style of the image-caption datasets. Follow these rules: ,→ ,→ - Remove unnecessary narrative phrases like "we can see," "there is," "in this image," etc.,→ - Ensure the caption is standalone and descriptive. - Use simple, objective language that highlights key elements. - Keep it concise--just a single phrase. - Follow the classical style of caption datasets. - If the description is vague, subjective, or does not describe a concrete visual element (e.g., "The image is taken indoor," "This image is blurred"), return `<INVALID>`. ,→ ,→ - Wrap the output in `{}`and add nothing else. ### **Examples:** -**Input:** "We can see a young elephant stands which is near the water in a wooded area.",→ **Output:** {A young elephant stands near the water in a wooded area.} -**Input:** "In this image I can see some young children kicking a soccer ball in a field.",→ **Output:** {A group of young children kicking a soccer ball around a field.} -**Input:** "In the left of the image, we see a pole that has two green street signs on it.",→ **Output:** {A pole has two green street signs on it.} -**Input:** "We can see two surfboards which are stuck in the sand along the seashore.",→ **Output:** {Two surfboards stuck in the sand along the seashore.} -**Input:** "This image consists of a man which rides a wakeboard behind a boat.",→ **Output:** {A man rides a wakeboard behind a boat.} -**Input:** "In the background, there are a bunch of sticky notes and a pair of scissors.",→ **Output:** {A bunch of sticky notes and a pair of scissors.} -**Input:** "It looks like a sepia-toned photograph of a motorcycle underneath the shadow of a,→ tree." **Output:** {A sepia-toned photograph of a motorcycle underneath the shadow of a tree.},→ -**Input:** "There is a sky" **Output:** {A sky.} -**Input:** "She is smiling." **Output:** {A smiling girl.} -**Input:** "The image is taken indoor." **Output:** {<INVALID>} -**Input:** "This image is edited." **Output:** {<INVALID>} -**Input:** "The image is blurred." **Output:** {<INVALID>} -**Input:** "I think he is about to jump." **Output:** {<INVALID>} Now, rewrite the following captions accordingly. Wrap each in `{}`and add nothing else:,→ <INPUT CAPTION> Figure 4: LLM Prompt for rephrasing trace captions. 7 Preprint (a) Localized Narrative In this picture I can observe a dog running on the land. I can observe water and grass on the ground. The background is blurred. (b) Track 1 Original: In this picture I can observe a dog running on the land. Processed: A dog runs on the land. (c) Track 2 Original: I can observe water and grass on the ground. Processed: Water and grass on the ground. (d) Track 3 Original: The background is blurred. Processed:<INVALID> In this image there is a person wearing a helmet is on a vehicle. At the bottom of the image there are side mirrors. The background of the image is blurred. Original: In this image there is a person wearing a helmet is on a vehicle. Processed: A person wearing a helmet rides a vehicle. Original: At the bottom of the image there are side mirrors. Processed: Side mirrors. Original: The background of the image is blurred. Processed:<INVALID> This image is taken outdoors. In this image we can see the green grass on the ground. In the middle of the image we can see there are two dogs. Original: This image is taken outdoors. Processed:<INVALID> Original: In this image we can see the green grass on the ground. Processed: Green grass on the ground. Original: In the middle of the image we can see there are two dogs. Processed: Two dogs. Figure 5: Narrative vs. Trace Samples. The first column displays sample images from the Localized Narrative dataset Pont-Tuset et al. (2020b). The remaining three columns show the corresponding mouse traces, along with the captions generated by the LLM. Captions marked with <INVALID> are removed from the dataset. 8 Preprint PATCH DeCap a cat is sleeping on a cluttered desk. a cat is sleeping on a cluttered desk. a tennis player is playing tennis on the court for a serve. a few people are skiing on a snowy mountain. Ours (CLIP + Mem.) a cat is sitting on the bed and it’s contents. a cat is sitting at a table with a full laptop . a couple of people are in the middle of a tennis court. a few people are skiing in a snowy mountain. Ours (Talk2DINO + Mem.) a plant in a vase sitting on a table. office supplies , pens , toys , and other items on desk. a street light in front of a large building. a cloudy sky is seen in this cloudy day. TRACE GT Two giraffes, rocks, and a fence. A sky. A flag. People walking on a walkway. DeCap a giraffe in a zoo with a city in the background. a giraffe in a zoo with a city in the background. a man on a skateboard who is holding onto a skateboard. a park filled with people sitting on benches near trees. Ours (CLIP + Mem.) there are some people that are in a lot by a tree. there are some people that are out by a lot of trees. there are some people that are in the water with a couple of them. there are several traffic lights out in the wild. Ours (Talk2DINO + Mem.) two giraffes standing in a fenced area. a view of a city with a sky in the background. a flag is flying high in the air. a large group of people walking on a sidewalk. DENSE GT light shining through the trees. bench sitting in the woods. a clock at a train station. black cat sitting on a bench. DeCap a bench sits in the middle of a wooded area. a bench sits in the middle of a wooded area. a train traveling along the platform of a public train. a woman squatting on a bench with a cat. DeCap (Crop) a person in a tree is standing in the wild near trees. a bench sitting in the middle of a wooded area. a black cat is leaning on a black cat. a a close up of a person standing by a person holding a phone. Ours (CLIP + Mem.) a bear is in the woods among the trees. there are many trees that are standing in the woods. a train is on the tracks and going by. there is a person that is out on the kitchen. Ours (Talk2DINO + Mem.) sun shining through the trees at sunset. a park bench sitting in the middle of a wooded area. a clock on a train station platform above a train. a black cat is sitting on a black bench. REGION-SET GT an elderly man in a cap sitting on a bench. an old man sitting on a bench with a purse. a man performing a trick near a fire hydrant. a man swinging a baseball bat as another looks on. DeCap a man sitting on a bench while holding a door. a man sitting on a bench while holding a door. a man on a skateboard doing a trick. a baseball player at bat getting ready to hit the ball. Ours (CLIP + Mem.) a bathroom has a blue floor and it is very clean. a bathroom has a blue toilet and the walls. there are many cars driving down the street corner. some baseball players are on the field playing baseball. Ours (Talk2DINO + Mem.) a man in a hat sitting on a bench. a man sits on a wooden bench with a bag on his back. a fire hydrant on a sidewalk next to a street pole. a baseball player is swinging his bat as a crowd watches. IMAGE GT A black cat rubbing against a bottle of wine. A man in a wetsuit rides a wave. A wooden bench sitting on a beach. A wooden table with a plate of cake and coffee. DeCap a black cat standing next to a bottle of wine glasses a man on a surf board riding a wave in the water a bench sits on the beach next to the ocean a slice of cake on a plate with a cup of cake ZeroCap a Wine dro Pet Cat. a man surfing in the area 0. a beachfront bench. a sunny cake with tea. CLOSE a cat sitting on the counter of a green bottle. a man on a surf board riding a wave in the ocean. a wooden bench sitting in the sand near the ocean. and a cake is sitting on a white plate. Ours (Talk2DINO + Mem.) a black cat sitting on a chair next to a bottle of wine. a man on a surfboard riding a wave. a bench sitting on the beach next to the ocean. a piece of cake on a plate with a cup of coffee. Figure 6: Qualitative results. We report four predictions of our model and compare baselines from the finer (top) to the coarser (bottom) task. For trace captioning examples, the trace time is color-coded from start (red) to end (yellow). DeCap = DeCap applied on the whole image. DeCap (Crop) = DeCap applied on cropped box. ZeroCap = ZeroCap Tewel et al. (2022) applied to the whole image. CLOSE = CLOSE Gu et al. (2023) applied to the whole image. Ours (CLIP + Mem.) = Our patch-based framework using CLIP as backbone and the projection as modality gap mitigation strategy. Ours (Talk2DINO + Mem.) = Our patch-based framework using Talk2DINO as backbone and the projection as modality gap mitigation strategy. GT = ground-truth caption. 9 Preprint PATCH DeCap a group of people in a kitchen are cooking food. a table with a cup of coffee and plates of silverware. a small bed is curled up in a cluttered room. a police car is parked on the side of a street. Ours (CLIP + Mem.) a couple of people that are standing around each other. a bunch of people are sitting at the table together. aa baby is in a bedroom with a white sink and toilet. there are a few street signs in the middle of the neighborhood. Ours (Talk2DINO + Mem.) a forest with trees in the background. a cup of coffee with a spoon sitting on a plate. a dog laying on a rug in a living room. a fence that is next to a road. TRACE GT Clouds and the sun in the sky. A person wearing a cap. A Christmas tree decorated with balls and toys. Few people on a boat. DeCap a couple of people are sitting on a bench looking at the ocean. a woman at a table putting food in a pot. two people posing with a man and woman having a glass of wine. a man on a boat in a body of water with other boats. Ours (CLIP + Mem.) a couple of people are on a boat by the ocean. a couple of people are in a kitchen making food. there are two people in a kitchen with a red sweater. there are some people in the water by a boat. Ours (Talk2DINO + Mem.) a sunset in the distance in the sun a person wearing a hat looking at something in the background the christmas tree is decorated for christmas a boat with many people on it DENSE GT a white ceiling fan hanging in the kitchen. a plane flying in the sky. two sandwiches on a plate. potted plant on the ledge. DeCap a kitchen with a large refrigerator , cabinets and stove. a building is flying under a traffic light in the air near a building. a sandwich and a plate of soup on a table. a bird that is perched on a small bird. DeCap (Crop) a bathroom sink with a variety of toilet above the wall. a large airplane is in flight on the airport. a sandwich on a plate containing a sandwich. a a close up of a person standing by a person holding a phone. Ours (CLIP + Mem.) a kitchen has a lot of fridge and a stove in it. a lot of a building is outside of a yellow car. the couple of food are in the kitchen with a meal. there is a man that is about to take a trick. Ours (Talk2DINO + Mem.) a ceiling fan is hanging in the kitchen. there is a plane flying high in the sky. a plate topped with two sandwiches on a table. a potted plant sitting on a ledge. REGION SET GT Dogs near the edge of water . A soccer player is running while kicking a ball . A brown-haired woman is pushing a baby stroller . The child lays on the hardwood floor . DeCap a dog and his dogs are wading in the muddy water. a soccer player in the soccer uniform tries to kick the ball. a man and a child walking in the street while holding a stroller. a young boy sitting on the floor in a room. Ours (CLIP + Mem.) there are many things that are out in the water. there are some people on a baseball field playing a game. there are some cars and a man about to go down the street. the kitchen is very clean and has an open door. Ours (Talk2DINO + Mem.) two dogs near one another near water. a soccer player getting ready to kick the ball. a woman pushing a stroller with a child inside. a young child is laying on the floor. IMAGE GT Four birds are chasing another bird which has a piece of food in its mouth. Brown-haired girl wearing a green tank top, talking on a cellphone. A woman with blond-hair is sitting in a booth with a drink working on her laptop. A young girl in a blue shirt is in a bowling alley, and is casting her ball down a lane. DeCap a flock of birds flying over the water. a woman talking on a cell phone while on a street. a woman sitting at a table using a laptop. a young girl playing a bowling game on wii. ZeroCap a gull mating. a man in the back of a pickup truck with blood on the back. a reader’s writing on a laptop on deskmounted computer. a view hitting the deck pin at the end of the row stretch. CLOSE a group of birds flying over a body of water. a woman looking at her cell phone while standing in a street. a woman sitting at a table with a laptop and a drink. on a person on a skateboard doing a trick. Ours (Talk2DINO + Mem.) a flock of birds flying in the sky. a woman talking on a cell phone in a market. a woman sitting at a cafe using her laptop. a woman playing a bowling game on the bowling. Figure 7: Qualitative results. We report four predictions of our model and compare baselines from the finer (top) to the coarser (bottom) task. For trace captioning examples, the trace time is color-coded from start (red) to end (yellow). DeCap = DeCap applied on the whole image. DeCap (Crop) = DeCap applied on cropped box. ZeroCap = ZeroCap Tewel et al. (2022) applied to the whole image. CLOSE = CLOSE Gu et al. (2023) applied to the whole image. Ours (CLIP + Mem.) = Our patch-based framework using CLIP as backbone and the projection as modality gap mitigation strategy. Ours (Talk2DINO + Mem.) = Our patch-based framework using Talk2DINO as backbone and the projection as modality gap mitigation strategy. GT = ground-truth caption. 10