Dimension d = O(m^{-2} log n) nearly achieves the optimal margin m^rd(+∞, A) for retrieval embeddings, with matching lower bounds showing d = O(k log(n/k)) suffices and is necessary for m = Θ(k^{-1/2}) on k-sparse query matrices.
hub Mixed citations
Perception Encoder: The best visual embeddings are not at the output of the network
Mixed citation behavior. Most common role is background (67%).
40
Pith papers citing it
Background
67% of classified citations
abstract
We introduce Perception Encoder (PE), a state-of-the-art vision encoder for image and video understanding trained via simple vision-language learning. Traditionally, vision encoders have relied on a variety of pretraining objectives, each tailored to specific downstream tasks such as classification, captioning, or localization. Surprisingly, after scaling our carefully tuned image pretraining recipe and refining with our robust video data engine, we find that contrastive vision-language training alone can produce strong, general embeddings for all of these downstream tasks. There is only one caveat: these embeddings are hidden within the intermediate layers of the network. To draw them out, we introduce two alignment methods: language alignment for multimodal language modeling, and spatial alignment for dense prediction. Together, our PE family of models achieves best-in-class results on a wide variety of tasks, including (1) zero-shot image and video classification and retrieval, simultaneously obtaining 86.6 average zero-shot ImageNet robustness and 76.9 zero-shot Kinetics-400 video classification; (2) document, image, and video Q&A, enabling 94.6 DocVQA, 80.9 InfographicVQA, and 82.7 PerceptionTest with an 8B LLM; and (3) spatial tasks such as detection, tracking, and depth estimation, setting a new COCO state-of-the-art of 66.0 box mAP. To foster further research, we release our models, code, and novel dataset of synthetically and human-annotated videos: https://github.com/facebookresearch/perception_models
hub tools
citation-role summary
background 6
method 3
citation-polarity summary
representative citing papers
The paper introduces SP-VTP as a new setting for egocentric manipulation, releases the EgoSPT dataset with first-frame spatial annotations, and proposes the SPOT model that outperforms non-prompted baselines on cross-scene trajectory prediction.
Optimal INR freeze depth matches highest weight stable rank layer; SAEs reveal SIREN atoms are localized while FFMLP atoms trace cohort contours with causal impact on PSNR.
VAnim creates open-domain text-to-SVG animations via sparse state updates on a persistent DOM tree, identification-first planning, and rendering-aware RL with a new 134k-example benchmark.
Pretraining on 1M wild videos followed by post-training on curated data yields high-fidelity feedforward 3D avatars that generalize across identities, clothing, and lighting with emergent relightability and loose-garment support.
Text-only contrastive fine-tuning of an MLLM with hard negatives produces embeddings that handle temporal, negation, and multimodal nuances in video retrieval and achieves SOTA performance.
SAM 3 introduces promptable concept segmentation that doubles accuracy of prior systems on images and videos while improving standard SAM segmentation performance.
MoTIF adds temporal self-attention and automatic VLM-based concept discovery to concept bottleneck models for interpretable video classification, showing gains over prior global CBMs on benchmarks.
Introduces LOES, a constructive spectral method to select task-discriminative subspaces from intermediate layer embeddings, and GeoReg for enforcing simplicial class geometry during fine-tuning, with reported gains increasing with model depth across modalities.
RADAR is a geometrically grounded metric that predicts cross-domain transferability by comparing layer-wise representation trajectory distributions in foundation models.
DySink uses adaptive retrieval of relevant historical frames plus a sink anomaly gate to improve dynamic degree and temporal quality in minute-long autoregressive video generation.
RAE v2 reaches gFID 1.06 on ImageNet-256 in 80 epochs by combining multi-layer encoder sums, complementary REPA targets, and free guidance via output reparameterization.
SparseSAM achieves 2x faster inference and 2.8x memory reduction in SAM with only 0.004 mIoU loss at 0.4 density via Stripe-Sort Attention and Residual-Consistency MLP.
SegRAG is a training-free retrieval-augmented framework that extracts class-specific point prompts from a filtered DINOv3 feature bank to boost SAM3 semantic segmentation performance on standard and agricultural benchmarks.
OProver-32B achieves top Pass@32 scores on MiniF2F, ProverBench, and PutnamBench by combining continued pretraining with iterative agentic proving, retrieval, SFT on repairs, and RL on unresolved cases using a 6.86M-proof dataset.
Qwen-Image-VAE-2.0 achieves state-of-the-art high-compression image reconstruction and superior diffusability for diffusion models, with a new text-rich document benchmark.
CAMFusion fuses multiview 2D vision-language embeddings via cross-attention and multiview consistency self-supervision to produce better 3D semantic and instance representations, outperforming averaging and reaching SOTA on benchmarks including zero-shot out-of-domain cases.
A timing-aware pre-quantization fusion approach integrates visual cues into audio tokenizers along the temporal axis, maintaining reconstruction quality while outperforming audio-only and prior multimodal baselines on downstream tasks.
TIPSv2 improves dense patch-text alignment in vision-language pretraining through distillation and iBOT++ modifications, yielding models on par with or better than recent baselines on 9 tasks across 20 datasets.
A vision-language-aligned world model turns visuomotor MPC into a language-following planner that reaches 87% success on 288 unseen semantic tasks where standard VLAs drop to 22%.
UniversalVTG is a lightweight foundation model for video temporal grounding that achieves state-of-the-art results across five benchmarks while being over 100 times smaller than recent MLLM-based methods.
Q-Zoom achieves up to 4.39x inference speedup in high-resolution MLLM scenarios via query-aware gating and region localization, matching or exceeding baseline accuracy on document and high-res benchmarks.
Telescope uses learnable hyperbolic foveation to deliver a 76% relative mAP gain (0.185 to 0.326) for objects beyond 250 meters while keeping overhead low.
VideoStir introduces a spatio-temporal graph-based structure and intent-aware retrieval for long-video RAG, achieving competitive performance with SOTA methods via a new IR-600K dataset.
citing papers explorer
-
Is Dimensionality a Barrier for Retrieval Models?
Dimension d = O(m^{-2} log n) nearly achieves the optimal margin m^rd(+∞, A) for retrieval embeddings, with matching lower bounds showing d = O(k log(n/k)) suffices and is necessary for m = Θ(k^{-1/2}) on k-sparse query matrices.
-
Spatially Prompted Visual Trajectory Prediction for Egocentric Manipulation
The paper introduces SP-VTP as a new setting for egocentric manipulation, releases the EgoSPT dataset with first-frame spatial annotations, and proposes the SPOT model that outperforms non-prompted baselines on cross-scene trajectory prediction.
-
What Cohort INRs Encode and Where to Freeze Them
Optimal INR freeze depth matches highest weight stable rank layer; SAEs reveal SIREN atoms are localized while FFMLP atoms trace cohort contours with causal impact on PSNR.
-
VAnim: Rendering-Aware Sparse State Modeling for Structure-Preserving Vector Animation
VAnim creates open-domain text-to-SVG animations via sparse state updates on a persistent DOM tree, identification-first planning, and rendering-aware RL with a new 134k-example benchmark.
-
Large-scale Codec Avatars: The Unreasonable Effectiveness of Large-scale Avatar Pretraining
Pretraining on 1M wild videos followed by post-training on curated data yields high-fidelity feedforward 3D avatars that generalize across identities, clothing, and lighting with emergent relightability and loose-garment support.
-
Adapting MLLMs for Nuanced Video Retrieval
Text-only contrastive fine-tuning of an MLLM with hard negatives produces embeddings that handle temporal, negation, and multimodal nuances in video retrieval and achieves SOTA performance.
-
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.
-
Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification
MoTIF adds temporal self-attention and automatic VLM-based concept discovery to concept bottleneck models for interpretable video classification, showing gains over prior global CBMs on benchmarks.
-
Uncovering the Latent Potential of Deep Intermediate Representations
Introduces LOES, a constructive spectral method to select task-discriminative subspaces from intermediate layer embeddings, and GeoReg for enforcing simplicial class geometry during fine-tuning, with reported gains increasing with model depth across modalities.
-
RADAR: Relative Angular Divergence Across Representations
RADAR is a geometrically grounded metric that predicts cross-domain transferability by comparing layer-wise representation trajectory distributions in foundation models.
-
DySink: Dynamic Frame Sinks for Autoregressive Long Video Generation
DySink uses adaptive retrieval of relevant historical frames plus a sink anomaly gate to improve dynamic degree and temporal quality in minute-long autoregressive video generation.
-
Improved Baselines with Representation Autoencoders
RAE v2 reaches gFID 1.06 on ImageNet-256 in 80 epochs by combining multi-layer encoder sums, complementary REPA targets, and free guidance via output reparameterization.
-
SparseSAM: Structured Sparsification of Activations in Segment Anything Models
SparseSAM achieves 2x faster inference and 2.8x memory reduction in SAM with only 0.004 mIoU loss at 0.4 density via Stripe-Sort Attention and Residual-Consistency MLP.
-
SegRAG: Training-Free Retrieval-Augmented Semantic Segmentation
SegRAG is a training-free retrieval-augmented framework that extracts class-specific point prompts from a filtered DINOv3 feature bank to boost SAM3 semantic segmentation performance on standard and agricultural benchmarks.
-
OProver: A Unified Framework for Agentic Formal Theorem Proving
OProver-32B achieves top Pass@32 scores on MiniF2F, ProverBench, and PutnamBench by combining continued pretraining with iterative agentic proving, retrieval, SFT on repairs, and RL on unresolved cases using a 6.86M-proof dataset.
-
Qwen-Image-VAE-2.0 Technical Report
Qwen-Image-VAE-2.0 achieves state-of-the-art high-compression image reconstruction and superior diffusability for diffusion models, with a new text-rich document benchmark.
-
Cross-Attentive Multiview Fusion of Vision-Language Embeddings
CAMFusion fuses multiview 2D vision-language embeddings via cross-attention and multiview consistency self-supervision to produce better 3D semantic and instance representations, outperforming averaging and reaching SOTA on benchmarks including zero-shot out-of-domain cases.
-
Why Your Tokenizer Fails in Information Fusion: A Timing-Aware Pre-Quantization Fusion for Video-Enhanced Audio Tokenization
A timing-aware pre-quantization fusion approach integrates visual cues into audio tokenizers along the temporal axis, maintaining reconstruction quality while outperforming audio-only and prior multimodal baselines on downstream tasks.
-
TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment
TIPSv2 improves dense patch-text alignment in vision-language pretraining through distillation and iBOT++ modifications, yielding models on par with or better than recent baselines on 9 tasks across 20 datasets.
-
Grounded World Model for Semantically Generalizable Planning
A vision-language-aligned world model turns visuomotor MPC into a language-following planner that reaches 87% success on 288 unseen semantic tasks where standard VLAs drop to 22%.
-
UniversalVTG: A Universal and Lightweight Foundation Model for Video Temporal Grounding
UniversalVTG is a lightweight foundation model for video temporal grounding that achieves state-of-the-art results across five benchmarks while being over 100 times smaller than recent MLLM-based methods.
-
Q-Zoom: Query-Aware Adaptive Perception for Efficient Multimodal Large Language Models
Q-Zoom achieves up to 4.39x inference speedup in high-resolution MLLM scenarios via query-aware gating and region localization, matching or exceeding baseline accuracy on document and high-res benchmarks.
-
Telescope: Learnable Hyperbolic Foveation for Ultra-Long-Range Object Detection
Telescope uses learnable hyperbolic foveation to deliver a 76% relative mAP gain (0.185 to 0.326) for objects beyond 250 meters while keeping overhead low.
-
VideoStir: Understanding Long Videos via Spatio-Temporally Structured and Intent-Aware RAG
VideoStir introduces a spatio-temporal graph-based structure and intent-aware retrieval for long-video RAG, achieving competitive performance with SOTA methods via a new IR-600K dataset.
-
Learning Robust Visual Features in Computed Tomography Enables Efficient Transfer Learning for Clinical Tasks
VoxelFM learns robust 3D CT visual features via DINO self-distillation that transfer effectively to seven clinical task categories using frozen backbones and lightweight heads, outperforming prior CT foundation models even on report generation.
-
From Local Matches to Global Masks: Template-Guided Instance Detection and Segmentation in Open-World Scenes
L2G-Det detects and segments novel object instances in open scenes by using local template patch matches to generate points that prompt an augmented SAM for global masks.
-
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.
-
Chorus: Multi-Teacher Pretraining for Holistic 3D Gaussian Scene Encoding
Chorus pretrains a shared 3D Gaussian scene encoder via multi-teacher distillation to capture holistic features from high-level semantics to fine-grained structure, with strong transfer on segmentation and point-cloud tasks using far fewer scenes.
-
VFM-VAE: Vision Foundation Models Can Be Good Tokenizers for Latent Diffusion Models
VFM-VAE uses a frozen VFM directly as LDM tokenizer via a custom decoder, reaching gFID 2.22 in 80 epochs and 1.62 after 640 epochs.
-
Deepfake Detection that Generalizes Across Benchmarks
GenD achieves state-of-the-art average cross-dataset AUROC in deepfake detection by parameter-efficient adaptation of a foundational vision encoder with hyperspherical manifold enforcement via L2 normalization and metric learning.
-
V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
V-JEPA 2 pre-trained on massive unlabeled video achieves strong results on motion understanding and action anticipation, SOTA video QA at 8B scale, and enables zero-shot robotic planning on Franka arms using only 62 hours of unlabeled robot video.
-
SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics
SmolVLA is a small efficient VLA model that achieves performance comparable to 10x larger models while training on one GPU and deploying on consumer hardware via community data and chunked asynchronous action prediction.
-
Enabling High Error Tolerance in Satellite Video Transmissions by Generative Semantic Communication
A generative semantic communication method for satellite video achieves 2.5 dB higher PSNR than conventional semantic comms at 45% error rate and remains functional above 80% error by combining semantic encoding with generative reconstruction.
-
Sapiens2
Sapiens2 improves pretraining, data scale, and architecture over its predecessor to set new state-of-the-art results on human pose estimation, body-part segmentation, normal estimation, and new tasks like pointmap and albedo estimation.
-
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.
-
Revisiting Compositionality in Dual-Encoder Vision-Language Models: The Role of Inference
Dual-encoder VLMs gain robust compositional generalization by learning localized alignments from frozen patch and token embeddings instead of using global similarity.
-
SegEarth-OV3: Exploring SAM 3 for Open-Vocabulary Semantic Segmentation in Remote Sensing Images
SAM 3 can be applied training-free to remote sensing open-vocabulary segmentation and change detection by fusing its semantic and instance heads and filtering with presence scores.
-
Position: Life-Logging Video Streams Make the Privacy-Utility Trade-off Inevitable
Life-logging video streams create an inevitable privacy-utility trade-off that is a foundational challenge for always-on AI systems.
-
CoCo-SAM3: Harnessing Concept Conflict in Open-Vocabulary Semantic Segmentation
CoCo-SAM3 improves SAM3 by aligning evidence from synonymous prompts for concept consistency and then running inter-class competition on a unified scale to reduce mask overlaps.
-
Adaptive Forensic Feature Refinement via Intrinsic Importance Perception
I2P adaptively selects the most discriminative layers from visual foundation models for synthetic image detection and constrains task updates to low-sensitivity parameter subspaces to improve specificity without harming generalization.