MSLA is the first physically deployable attack that uses adversarial lighting to break semantic alignment in VLMs such as CLIP, LLaVA, and BLIP, causing classification failures and hallucinations in real scenes.
hub Canonical reference
Demystifying CLIP Data
Canonical reference. 71% of citing Pith papers cite this work as background.
35
Pith papers citing it
Background
71% of classified citations
abstract
Contrastive Language-Image Pre-training (CLIP) is an approach that has advanced research and applications in computer vision, fueling modern recognition systems and generative models. We believe that the main ingredient to the success of CLIP is its data and not the model architecture or pre-training objective. However, CLIP only provides very limited information about its data and how it has been collected, leading to works that aim to reproduce CLIP's data by filtering with its model parameters. In this work, we intend to reveal CLIP's data curation approach and in our pursuit of making it open to the community introduce Metadata-Curated Language-Image Pre-training (MetaCLIP). MetaCLIP takes a raw data pool and metadata (derived from CLIP's concepts) and yields a balanced subset over the metadata distribution. Our experimental study rigorously isolates the model and training settings, concentrating solely on data. MetaCLIP applied to CommonCrawl with 400M image-text data pairs outperforms CLIP's data on multiple standard benchmarks. In zero-shot ImageNet classification, MetaCLIP achieves 70.8% accuracy, surpassing CLIP's 68.3% on ViT-B models. Scaling to 1B data, while maintaining the same training budget, attains 72.4%. Our observations hold across various model sizes, exemplified by ViT-H achieving 80.5%, without any bells-and-whistles. Curation code and training data distribution on metadata is made available at https://github.com/facebookresearch/MetaCLIP.
hub tools
citation-role summary
background 5
method 2
citation-polarity summary
representative citing papers
Embedding norms in contrastive models encode semantic properties via optimization dynamics under scale-invariant losses.
ImageAuditor is the first MIA for IRAG that achieves over 80% AUROC with four queries by using reward-guided policy optimization for cross-modal retrieval and task-specific prompting for signal extraction.
Gromov-Wasserstein distance between modalities provides a stronger, inference-only predictor of final VLM performance than conventional encoder metrics, backed by theory linking it to cross-modal learnability and verified across 60+ training runs.
DouC fuses an OG-CLIP branch for patch reliability via inference-time token gating with an FADE-CLIP branch for structural priors via proxy attention, outperforming prior training-free methods on eight benchmarks.
A wrinkle-field perturbation method creates photorealistic non-rigid image changes that degrade state-of-the-art VLMs on image captioning and VQA more effectively than prior baselines.
SAM 3 introduces promptable concept segmentation that doubles accuracy of prior systems on images and videos while improving standard SAM segmentation performance.
MMSearch-R1 uses reinforcement learning to train multimodal models for on-demand multi-turn internet search with image and text tools, outperforming same-size RAG baselines and matching larger ones while cutting search calls by over 30%.
On-policy self-distillation with sampled demonstrations reduces rollout diversity by amplifying existing probability gaps in the base model, unlike ideal RL which preserves ratios among correct outputs.
RaysUp reconstructs high-resolution VFM feature maps via Spatially Decoupled Guidance Encoder, Any-Resolution Cross-Attention, RayPE with 6D Plucker rays, and Geometry-Aware Neighborhood Attention, claiming SOTA results at 16% parameters and 7x speed of AnyUp.
LIVE uses language to generate task-centric vision embeddings at inference, reducing hallucinations by 34 points on MMVP, outperforming larger VLMs on VQA, and generalizing to unseen tasks.
SynIB is an information-theoretic objective that adds a penalty for unimodal confidence to standard task loss, improving accuracy on synergy-dependent examples by up to 7.8% across synthetic XOR tasks and five real-world multimodal benchmarks.
Data curation alone raises VLM accuracy by more than 11 points on average across many benchmarks while cutting required training compute by up to 87 times.
Exploiting linear structure in VLM embeddings, a synthetic-data pre-training method yields background-invariant representations that exceed 90% worst-group accuracy on Waterbirds even under 100% spurious correlation with no minority examples in training.
LLaVA-UHD v4 reduces visual-encoding FLOPs by 55.8% for high-resolution images in MLLMs via slice-based encoding plus intra-ViT early compression while matching or exceeding baseline performance on document, OCR, and VQA benchmarks.
GenLIP pretrains ViTs to generate language tokens from images via LM objective without contrastive batches or extra decoders, matching baselines on less data and improving on OCR after multi-resolution continued pretraining.
CLIP models understand 360-degree textual semantics via explicit identifiers but show limited comprehension of visual semantics under horizontal circular shifts, which a LoRA fine-tuning approach improves with a noted trade-off in original task performance.
ViTs exhibit lazy aggregation by relying on irrelevant background patches for global semantics, and selectively integrating patch features into the CLS token reduces this effect and improves results across label-, text-, and self-supervision.
Frozen features from vision foundation models enable a linear probe to outperform specialized AIGI detectors by over 30% on in-the-wild data due to emergent forgery knowledge from pre-training.
MetaEmbed trains fixed learnable Meta Tokens to produce granularity-organized multi-vector embeddings that support test-time scaling in multimodal retrieval.
LeakyCLIP reconstructs images from CLIP embeddings with over 258% SSIM gain versus baselines and enables membership inference from reconstruction metrics on LAION-2B data.
Fine-tuned image classifiers, led by RegNetY-16GF at about 99% top-1 accuracy, can sort historical archaeological page scans into 11 content categories, and the authors release the dataset, code, and model weights.
GLM-4.5V reaches state-of-the-art results on 42 multimodal benchmarks among open-source models of similar size by applying reinforcement learning with curriculum sampling to a strong vision foundation model.
An iterative bootstrapped self-filtering approach selects balanced clean and diverse subsets from noisy vision-language datasets to train improved CLIP models.
citing papers explorer
-
Challenging Vision-Language Models with Physically Deployable Multimodal Semantic Lighting Attacks
MSLA is the first physically deployable attack that uses adversarial lighting to break semantic alignment in VLMs such as CLIP, LLaVA, and BLIP, causing classification failures and hallucinations in real scenes.
-
Optimization Dynamics Imprint Semantic Specificity in Contrastive Embedding Norms
Embedding norms in contrastive models encode semantic properties via optimization dynamics under scale-invariant losses.
-
ImageAuditor: Membership Inference Attack against Image-based Retrieval-Augmented Generation
ImageAuditor is the first MIA for IRAG that achieves over 80% AUROC with four queries by using reward-guided policy optimization for cross-modal retrieval and task-specific prompting for signal extraction.
-
Rethinking Model Selection in VLM Through the Lens of Gromov-Wasserstein Distance
Gromov-Wasserstein distance between modalities provides a stronger, inference-only predictor of final VLM performance than conventional encoder metrics, backed by theory linking it to cross-modal learnability and verified across 60+ training runs.
-
DouC: Dual-Branch CLIP for Training-Free Open-Vocabulary Segmentation
DouC fuses an OG-CLIP branch for patch reliability via inference-time token gating with an FADE-CLIP branch for structural priors via proxy attention, outperforming prior training-free methods on eight benchmarks.
-
When Surfaces Lie: Exploiting Wrinkle-Induced Attention Shift to Attack Vision-Language Models
A wrinkle-field perturbation method creates photorealistic non-rigid image changes that degrade state-of-the-art VLMs on image captioning and VQA more effectively than prior baselines.
-
SAM 3: Segment Anything with Concepts
SAM 3 introduces promptable concept segmentation that doubles accuracy of prior systems on images and videos while improving standard SAM segmentation performance.
-
MMSearch-R1: Incentivizing LMMs to Search
MMSearch-R1 uses reinforcement learning to train multimodal models for on-demand multi-turn internet search with image and text tools, outperforming same-size RAG baselines and matching larger ones while cutting search calls by over 30%.
-
On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity
On-policy self-distillation with sampled demonstrations reduces rollout diversity by amplifying existing probability gaps in the base model, unlike ideal RL which preserves ratios among correct outputs.
-
RaysUp: Ultra-light Universal Feature Upsampling via Geometry-Aware Ray Representation
RaysUp reconstructs high-resolution VFM feature maps via Spatially Decoupled Guidance Encoder, Any-Resolution Cross-Attention, RayPE with 6D Plucker rays, and Geometry-Aware Neighborhood Attention, claiming SOTA results at 16% parameters and 7x speed of AnyUp.
-
Language-Instructed Vision Embeddings for Controllable and Generalizable Perception
LIVE uses language to generate task-centric vision embeddings at inference, reducing hallucinations by 34 points on MMVP, outperforming larger VLMs on VQA, and generalizing to unseen tasks.
-
SynIB: Informational Bottleneck for Maximizing Synergy in Multimodal Learning
SynIB is an information-theoretic objective that adds a penalty for unimodal confidence to standard task loss, improving accuracy on synergy-dependent examples by up to 7.8% across synthetic XOR tasks and five real-world multimodal benchmarks.
-
20/20 Vision Language Models: A Prescription for Better VLMs through Data Curation Alone
Data curation alone raises VLM accuracy by more than 11 points on average across many benchmarks while cutting required training compute by up to 87 times.
-
Birds of a Feather Flock Together: Background-Invariant Representations via Linear Structure in VLMs
Exploiting linear structure in VLM embeddings, a synthetic-data pre-training method yields background-invariant representations that exceed 90% worst-group accuracy on Waterbirds even under 100% spurious correlation with no minority examples in training.
-
LLaVA-UHD v4: What Makes Efficient Visual Encoding in MLLMs?
LLaVA-UHD v4 reduces visual-encoding FLOPs by 55.8% for high-resolution images in MLLMs via slice-based encoding plus intra-ViT early compression while matching or exceeding baseline performance on document, OCR, and VQA benchmarks.
-
Let ViT Speak: Generative Language-Image Pre-training
GenLIP pretrains ViTs to generate language tokens from images via LM objective without contrastive batches or extra decoders, matching baselines on less data and improving on OCR after multi-resolution continued pretraining.
-
Probing CLIP's Comprehension of 360-Degree Textual and Visual Semantics
CLIP models understand 360-degree textual semantics via explicit identifiers but show limited comprehension of visual semantics under horizontal circular shifts, which a LoRA fine-tuning approach improves with a noted trade-off in original task performance.
-
Vision Transformers Need More Than Registers
ViTs exhibit lazy aggregation by relying on irrelevant background patches for global semantics, and selectively integrating patch features into the CLS token reduces this effect and improves results across label-, text-, and self-supervision.
-
Simplicity Prevails: The Emergence of Generalizable AIGI Detection in Visual Foundation Models
Frozen features from vision foundation models enable a linear probe to outperform specialized AIGI detectors by over 30% on in-the-wild data due to emergent forgery knowledge from pre-training.
-
MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interaction
MetaEmbed trains fixed learnable Meta Tokens to produce granularity-organized multi-vector embeddings that support test-time scaling in multimodal retrieval.
-
LeakyCLIP: Extracting Training Data from CLIP
LeakyCLIP reconstructs images from CLIP embeddings with over 258% SSIM gain versus baselines and enables membership inference from reconstruction metrics on LAION-2B data.
-
Page image classification for content-specific data processing
Fine-tuned image classifiers, led by RegNetY-16GF at about 99% top-1 accuracy, can sort historical archaeological page scans into 11 content categories, and the authors release the dataset, code, and model weights.
-
GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning
GLM-4.5V reaches state-of-the-art results on 42 multimodal benchmarks among open-source models of similar size by applying reinforcement learning with curriculum sampling to a strong vision foundation model.
-
Data Selection Through Iterative Self-Filtering for Vision-Language Settings
An iterative bootstrapped self-filtering approach selects balanced clean and diverse subsets from noisy vision-language datasets to train improved CLIP models.
-
ShellfishNet: A Domain-Specific Benchmark for Visual Recognition of Marine Molluscs
ShellfishNet is a new benchmark of 8,691 images across 32 mollusc taxa for evaluating vision models on real-world underwater ecological monitoring tasks including robustness to degradation.
-
From Codebooks to VLMs: Evaluating Automated Visual Discourse Analysis for Climate Change on Social Media
VLMs recover reliable population-level trends in climate change visual discourse on social media even when per-image accuracy is only moderate.
-
Concrete Jungle: Towards Concreteness Paved Contrastive Negative Mining for Compositional Understanding
Using lexical concreteness to guide contrastive negative mining and a new margin-based Cement loss, the Slipform framework reaches state-of-the-art on compositional benchmarks for vision-language models.
-
Human-Inspired Context-Selective Multimodal Memory for Social Robots
A new memory system for social robots selectively stores multimodal memories by emotional salience and novelty, achieving 0.506 Spearman correlation in selectivity and up to 13% better Recall@1 in multimodal retrieval.
-
Revealing Physical-World Semantic Vulnerabilities: Universal Adversarial Patch for Infrared Vision-Language Models
A single physically deployable curved-grid patch, optimized with differential evolution against the visual encoder's subspace, degrades classification, captioning, and VQA across several infrared vision-language models.
-
Adapting Vision-Language Foundation Model for Next Generation Medical Ultrasound Image Analysis
Introduces Hybrid Tuning adapter with frequency filtering and noise estimation to adapt CLIP for ultrasound segmentation and classification, claiming outperformance on six multi-center datasets.
-
Movie Gen: A Cast of Media Foundation Models
A 30B-parameter transformer and related models generate high-quality videos and audio, claiming state-of-the-art results on text-to-video, video editing, personalization, and audio generation tasks.
-
ReasonCLIP-58M: Visually Grounded Commonsense Reasoning Supervision for CLIP
ReasonCLIP-58M applies continual pretraining with visually grounded reasoning captions on 58M examples to improve CLIP-style models on commonsense and compositional reasoning tasks.
-
Page image classifier fine-tuned on century-spanning archives of scanned documents for further content-specific processing
Fine-tuned RegNetY-16GF reaches 99.16% accuracy classifying 48k century-old Czech archival pages into 11 visual content types, beating a 75% hand-crafted feature baseline, with models and data released publicly.
-
Multilingual OCR-Aware Fine-Tuning and Prompt-Guided Chain-of-Thought Reasoning for Multimodal Large Language Models
An OCR-aware multilingual framework combining synthetic data generation, LoRA SFT, and visual CoT prompting improves text extraction and translation robustness in multimodal LLMs on degraded images.
-
Rethinking the Global Knowledge of CLIP in Training-Free Open-Vocabulary Semantic Segmentation
GCLIP improves training-free CLIP semantic segmentation by fusing intermediate attention maps and suppressing an abnormal FFN channel.