The 19th European Conference on Computer Vision arrives in Malmö, Sweden this Monday, carrying a program that may define the shape of AI research for years to come. When ECCV 2026 opens at Malmö Arena on September 8, it will bring together five days of workshops, tutorials, keynotes, paper presentations, and live demos organized around three interlocking ideas that have moved from fringe speculation to the field’s dominant preoccupation: world models, Gaussian splatting as infrastructure, and embodied AI.
ECCV — the European Conference on Computer Vision — is one of only three A*-ranked venues in the field worldwide, alongside CVPR and ICCV, and because it meets only biennially, each edition represents a two-year snapshot of where the discipline stands rather than an annual progress report. This year’s edition is the largest in the conference’s history: 10,473 papers entered peer review, 2,883 were accepted at a 27.5 percent acceptance rate, and proceedings will be published by Springer Science+Business Media. Thousands of researchers from academia and industry are expected to attend in person; virtual registrants can access keynotes and oral sessions for $175 ($116 for students).
The five-day program divides across two adjacent venues in Malmö’s Hyllie district. Malmömässan hosts workshops, tutorials, the exhibition hall, and demo sessions on September 8–9. Malmö Arena hosts the opening ceremony, three keynotes, two panel discussions, and all main-conference sessions on September 10–12. Both venues sit within a short walk of each other and are easily reached from Copenhagen Airport (CPH)
What ECCV 2026 Is Really About: Three Ideas Taking Over the Field
Before cataloguing the program, it’s worth naming what ECCV 2026 is intellectually organized around — because the session titles, workshop tracks, and paper clusters all point toward the same convergence.
World models are no longer a research agenda; they are the research agenda. Jürgen Schmidhuber introduced the term “world model” in machine learning in 1990, and Yann LeCun revived it in a landmark 2022 position paper arguing that intelligence requires predictive models of the physical world rather than pattern matching alone. By ECCV 2026, that argument has been validated not by a single breakthrough but by institutional consensus: workshops explicitly dedicated to world models outnumber those from any prior edition of this conference, and every major track — from embodied AI to autonomous driving to 3D scene generation — frames its work in world-model terms. The ECCV 2026 workshop program reflects this dominance at a glance.
Gaussian splatting has become infrastructure. 3D Gaussian splatting — a technique that represents 3D scenes as millions of small, semi-transparent ellipsoids with learnable positions, orientations, and colors — was revitalized by the Inria research group’s landmark 2023 paper at SIGGRAPH and has since spread into every downstream pipeline it could touch. At ECCV 2026, the first poster session alone lists more than 80 splatting-related papers, spanning SLAM, underwater scenes, aerial views, thermal-infrared sensors, dynamic 4D scenes, compressed streaming, and medical CT reconstruction. The field is no longer debating whether Gaussian splatting works. It is engineering it into every pipeline.
These two trends may be answering the field’s hardest open question. Analysts who have studied world model investment have identified a structural problem with the analogy between world models and large language models: text has a universal token (the word or sub-word unit) that allows LLMs to represent all language in a single representational space. The physical world has no equivalent, as prior TechTimes analysis has documented. The convergence visible at ECCV 2026 — Gaussian splatting as a universal 3D substrate combined with world models as the prediction-and-planning layer — suggests the field may be implicitly building the answer: 3D Gaussians as the physical token that world models can operate on. No researcher has announced this convergence as a solved problem. But the papers presented at ECCV 2026 collectively describe its architecture.
September 8–9: Workshops and Tutorials
ECCV 2026 will open its workshop program with 86 accepted satellite workshops — the largest in the conference’s history — organized into thematic tracks: Embodied AI, Agents & World Models (13 workshops); 3D Vision & Geometry (11 workshops); Trustworthy & Responsible AI (10 workshops); Generative Models & Content Creation (6 workshops); Multimodal & Foundation Models (4 workshops); plus Medical & Biological Vision, Autonomous Driving, and Efficiency. The full ECCV 2026 workshop listing details all 86 events.
Selected workshops to watch on September 8:
ViLMa — Visual Localization and Mapping: From Optimization to 3D Foundation Models will examine the transition from classical optimization pipelines to large-scale learned representations — a shift that parallels the Gaussian splatting story. Agent in World: Living Worlds with Interactive Agents focuses on simulated environments where AI agents can act and be evaluated in open-ended physical settings. The afternoon track includes On-Device Embodied World Models, which addresses the engineering challenge of running learned world models on edge hardware — the deployment bottleneck for robotics — and World Models in the Loop: Towards Application-Driven World Model Evaluation, which asks how world model quality should be measured when the downstream application is the yardstick, not a benchmark leaderboard.
3D Human Understanding: Towards Human-Centric World Models examines body pose estimation, shape reconstruction, and action understanding as building blocks for world models centered on human behavior. A full-morning practical tutorial, The 5th Hands-on Egocentric Research Tutorial with Project Aria from Meta, uses Meta’s Aria glasses platform for hands-on egocentric perception research — grounding the abstract embodied perception literature in a concrete deployed hardware system.
DriveX — Foundation Models for Autonomous Driving is one of the most anticipated workshops, covering how vision-language and world models are being adapted for self-driving systems. Safe World Models for Trustworthy Embodied AI asks how world models can be made reliable enough for deployment in physical systems — the safety-alignment question for robots. UniWorld: Universal Representations for Perception, Reasoning, and World Modeling proposes unified frameworks serving perception, language-grounded reasoning, and predictive world modeling simultaneously.
The afternoon session on September 9 includes 3D in the Era of World Models — a flagship workshop where Apple researchers will present RayRoPE: Projective Ray Positional Encoding for Multi-view Attention — and How to Build Effective World Models for Embodied AI, a practical workshop covering training recipes, evaluation protocols, and open research questions. Multimodal Reasoning and Slow Thinking in the Large Model Era: Towards System 2 and Beyond draws on dual-process cognitive theory to develop AI that reasons deliberately rather than pattern-matching — a direct conversation with the world-model agenda.
September 10: Main Conference Opens
The main conference at Malmö Arena will open on September 10 with an opening ceremony at 8:00 a.m. local time (2:00 a.m. ET). Three parallel tracks of long orals and spotlights will begin at 9:00 a.m.
Long Oral: Computational Imaging, Shape Recovery and Camera Geometry will cover six papers on wavefront sensing, wide-field imaging with computational mirrors, polarization-guided surface normal estimation, depth from focus, scale-aware visual odometry (PRISM-VO), and cross-view yaw estimation.
Spotlight: 3D Reconstruction, Gaussian Splatting & Neural Rendering will feature fifteen talks, including DreamWorld: Geometry-Grounded Video Diffusion for 3D-Consistent World Modeling — which directly bridges video generation and coherent 3D structure — and TetraSDF: Analytic Isosurface Extraction with Multi-resolution Tetrahedral Grid.
Spotlight: Vision-Language Models & Visual Foundation Models will address evaluation, robustness, and grounding, including Molmo-Point: Better Pointing for VLMs with Grounding Tokens and work on spatial reasoning and token compression.
The first poster session will run from 10:30 a.m. to 12:30 p.m. alongside six live demos, including VIGS-SLAM: Visual Inertial Gaussian Splatting SLAM, Triangle Splatting SLAM, and interactive biometric demonstrations. Naver Labs Europe’s Sparse Auto-Regressive Modeling for Scene Generation from Multi-View Images will appear here — an autoregressive approach to generating coherent 3D scenes from multi-view input using token-by-token prediction analogous to LLM generation, with direct applications for robotics simulation and synthetic training data.
Keynote 1: Kristen Grauman — 3:00 p.m. local time (9:00 a.m. ET)
Kristen Grauman, Professor of Computer Science at the University of Texas at Austin and former Director at Meta’s Fundamental AI Research lab (FAIR), will deliver the conference’s first keynote. The 2026 Hill Prize in Artificial Intelligence recipient — awarded $500,000 by Lyda Hill Philanthropies for research on video understanding models that help people acquire physical and procedural skills — Grauman’s work sits at the intersection of egocentric video understanding, embodied perception, and scalable first-person AI systems. She and her collaborators have been recognized with the 2026 Hill Prize from EurekAlert, the 2011 Marr Prize, the 2017 Helmholtz Prize, three EgoVis Distinguished Paper awards, and the 2025 Huang Prize. Her keynote is expected to address video-grounded embodied perception — the challenge of learning from first-person visual experience at scale.
Following Grauman’s keynote, a 30-minute panel discussion will be moderated by Anna Rogers (IT University of Copenhagen), Roberta Sinatra (University of Copenhagen and ITU, co-lead of the Pioneer Centre for AI in Copenhagen), and Isabelle Augenstein (University of Copenhagen, Karen Spärck Jones Award recipient) on the topic “The Age of LLMs: Who Is the Researcher Now?” — examining how large language models are transforming research workflows, where AI assistance ends, and where intellectual contribution begins. The full panel information is available on the conference site.
September 11: LeCun on World Models
September 11 will run oral and spotlight sessions in the morning, including Steerable Visual Representations, Understanding Geometric Representations in Self-Supervised Vision Transformerseading Concept Circuits of Vision Transformers in the long oral track — papers that probe the internal representations of large vision models with implications for interpretability and controllable generation
The spotlight on geometry and matching will include RoMa v2: Harder Better Faster Denser Feature Matching, a significant update to the dense feature matching line of work with direct downstream applications to 3D reconstruction and localization. The video generation spotlight will include Live Avatar, a system for streaming real-time audio-driven avatar generation.
Keynote 2: Yann LeCun — “World Models: Enabling the Next AI Revolution” — 3:00 p.m. local time (9:00 a.m. ET)
The marquee event of the week. Yann LeCun, ACM Turing Award laureate (with Geoffrey Hinton and Yoshua Bengio), Executive Chairman of AMI Labs, and Jacob T. Schwartz Professor at NYU’s Courant Institute, will deliver his ECCV 2026 keynote directly on the theme pervading this year’s program. LeCun left Meta in November 2025 after twelve years as its Chief AI Scientist and co-founded AMI Labs, which raised $1.03 billion in March 2026 — the largest seed round in European startup history at the time — to build world models that learn from physical reality rather than text. AMI Labs backers include Bezos Expeditions, NVIDIA, Cathay Innovation, and Greycroft. LeCun is also a recipient of the 2025 Queen Elizabeth Prize for Engineering and the 2022 Princess of Asturias Award.
LeCun has been a vocal proponent of the Joint Embedding Predictive Architecture (JEPA) — a framework that trains AI systems to predict abstract representations of future states from observations rather than to reconstruct raw pixels — as a more tractable path to physical intelligence than large language model scaling. His 2022 position paper laid out this research agenda in detail. His keynote at ECCV 2026 is expected to lay out a research agenda for learning-based world models as the foundation for general embodied intelligence, tying together threads visible across dozens of workshops and papers in the program.
An evening reception and concert at Malmö Arena will close the second main conference day. Full registration includes admission.
September 12: Shotton on Autonomous Driving as Embodied AI’s First Test
The final day of the conference opens at 9:00 a.m. with Keynote 3: Jamie Shotton — “From Visual Recognition to Embodied AI.”
Jamie Shotton, Chief Scientist at Wayve and leader of Wayve Labs — the group building foundation models for embodied intelligence including GAIA and LINGO — will close the keynote program with a retrospective that doubles as a research agenda. His Shotton ECCV keynote page traces the arc from object recognition work at ECCV 2006 through Kinect body tracking and HoloLens hand- and eye-tracking at Microsoft to the current challenge of building systems that can act intelligently in the physical world. His work on Kinect received the Royal Academy of Engineering’s MacRobert Award in 2011; he received the Academy’s Silver Medal in 2020 and was elected a Fellow in 2021. Wayve closed a $1.2 billion Series D in 2026 at an $8.6 billion valuation, with investors including Microsoft, SoftBank, and NVIDIA.
Shotton’s ECCV 2026 abstract offers three principles from twenty years of computer vision progress: learning rather than engineering, betting on data and compute scaling, and not confusing today’s constraints with tomorrow’s limits. He will use autonomous driving as the first proving ground for embodied intelligence and will outline open questions on the path toward more general physical AI systems.
At 10:00 a.m., the closing panel “Can AI Build New Knowledge?” will ask whether future AI systems can genuinely discover new knowledge from experience — identifying knowledge gaps, acquiring evidence, forming abstractions, and updating themselves without forgetting. Panelists include Dima Damen (Professor at the University of Bristol, Senior Research Scientist at Google DeepMind, and creator of EPIC-KITCHENS), Yilun Du (Assistant Professor at Harvard’s Kempner Institute, formerly of OpenAI and Google DeepMind, focused on generative AI for embodied decision-making), and Viorica Patraucean (Research Scientist at Google DeepMind, lead organizer of the Perception Test challenge).
The final day’s poster and demo sessions close the conference through 6:00 p.m. local time. The long oral session 3D Reconstruction, Registration and Scene Modeling is among the most technically dense of the program. The Embodied AI, Robotics & Autonomous Driving spotlights will include the Naver Labs Europe navigation paper demonstrating that a frozen Vision Transformer with one scalar per image patch can navigate physical environments without retraining — a result that suggests the barrier to deploying high-quality robot vision is structurally lower than frontier-compute narratives imply.
Selected Notable Papers
A handful of accepted papers illustrate the program’s range:
Keep It Simple: Multi-Key Episodic Memory Retrieval for Ultra-Long Video Understanding proposes a memory mechanism for reasoning over video sequences far longer than any transformer context window — addressing a fundamental constraint in building AI systems that can learn from long continuous experience.
RoMa v2: Harder Better Faster Denser Feature Matching improves matching accuracy and speed for dense visual correspondence, with downstream impact on 3D reconstruction and localization.
OmniForcing: Unleashing Real-time Joint Audio-Visual Generation jointly generates synchronized audio and video in real time, a step toward fully multimodal generative applications.
Grounding World Simulation Models in a Real-World Metropolis connects learned world models to actual city-scale observations, directly addressing the gap between simulation-based training and real deployment conditions.
EgoCogNav: Cognition-Aware Human Egocentric Navigation applies cognition-aware representations to the navigation problem, bridging the egocentric perception work Grauman’s group has pioneered and the deployment focus Shotton will address in his keynote.
Does Gaussian Splatting Solve the Universal Token Problem for World Models?
This question will not be explicitly answered at ECCV 2026. But it is the implicit question behind the largest cluster of papers in the program.
Theorists of world models have identified a structural gap between language modeling and physical modeling: text has a universal discrete token that allows a single architecture to operate across all language domains. The physical world has no equivalent — no agreed-upon unit that captures physics, appearance, geometry, and dynamics in a jointly learnable space. This has made the “train a world model the way you train an LLM” analogy technically suspect, even as it proved enormously influential as an investment thesis. Prior TechTimes analysis of world model investment documented this structural problem in June 2026.
Gaussian splatting offers a candidate answer. A 3D Gaussian is a fully differentiable, compact, learnable representation of a spatial region — with position, orientation, scale, color, and opacity as learnable parameters. The explosion of Gaussian splatting papers at ECCV 2026 covers not just static reconstruction but dynamic 4D splatting, SLAM integration, compression for streaming, autonomous driving simulation via Gaussian-based world models, and medical imaging. What these applications share is that they treat Gaussians as a universal substrate for 3D scene state — the thing a world model should predict, update, and reason over.
Whether Gaussians are ultimately the right universal 3D token is an open question the conference will surface but not settle. The answer depends on whether Gaussian representations scale favorably with scene complexity, whether they can represent sufficiently abstract physical relationships (not just geometry), and whether learned world models can manipulate Gaussian scenes coherently during planning. The collective signal at ECCV 2026 is that the field has decided to find out by building on Gaussians rather than waiting for a theoretical resolution.
How Can I Follow ECCV 2026 Remotely?
Virtual attendees can access keynotes and oral sessions through the ECCV virtual platform at eccv.ecva.net immediately after each session concludes. Pre-recorded five-minute poster videos are already available on individual paper pages. The Cvent Events mobile app — available on iOS and Android as of September 5 — supports personal schedule building, networking, and live Q&A during keynotes for both in-person and virtual participants.
Virtual registration costs $175 for attendees and $116 for students. In-person registration (standard rate, early-bird having ended July 17) is $1,045 for full access and $557 for students. One-day workshop and tutorial passes are $250 ($125 for students).
The proceedings will be available through Springer and the ECCV virtual platform following the conference’s conclusion on September 12.
Frequently Asked Questions
What is ECCV 2026, and why is it significant this year?
ECCV — the European Conference on Computer Vision — is the 19th edition of one of the three most selective research venues in AI and computer vision worldwide, alongside CVPR and ICCV. It meets only every two years, which means each edition represents a genuine two-year accumulation of the field’s best work. The 2026 edition is significant because it arrives at the moment world models — AI systems that build internal representations of physical environments and predict how they change over time — have moved from a theoretical agenda to the organizing frame of an entire research community. With 2,883 accepted papers, 86 workshops (the most in the conference’s history), and three keynotes from researchers who have spent careers building toward embodied physical intelligence, ECCV 2026 is a snapshot of a field that has decided what it is for. Full program details are available at the ECCV 2026 official website.
What is Gaussian splatting, and why does it matter for robotics?
Gaussian splatting is a rendering technique that represents 3D scenes as millions of small semi-transparent ellipsoids — “Gaussians” — each with learnable position, size, orientation, color, and opacity. The technique was revitalized in a landmark 2023 paper from Inria, which demonstrated real-time photorealistic rendering from a sparse set of ordinary photographs. It matters for robotics because it provides a differentiable, compact, physically grounded representation of 3D scenes that robot navigation and manipulation systems can update in real time. SLAM systems using Gaussian splatting can simultaneously localize a robot and build a scene model at speeds and qualities that prior methods could not match. The more than 80 Gaussian splatting papers appearing in ECCV 2026’s first poster session alone signal that the field is treating this as infrastructure — the substrate on which other capabilities are built — rather than as a research novelty.
Can I attend ECCV 2026 without traveling to Malmö?
Yes. Virtual registration costs $175 for general attendees and $116 for students, and provides access to keynotes and oral sessions through the ECCV virtual platform at eccv.ecva.net, viewable immediately after each session concludes on-site. Pre-recorded five-minute poster videos are already publicly accessible on individual paper pages. The Cvent Events mobile app supports live Q&A during keynote sessions for virtual participants. For practitioners following the field without the budget for in-person attendance, the virtual track gives access to all three keynotes — Grauman on September 10, LeCun on September 11, and Shotton on September 12 — along with the oral presentations.
Does the convergence of world models and Gaussian splatting mean physical AI has a “universal token”?
Not yet confirmed, but the question is increasingly central to how the field positions itself. Large language models work because text can be represented as discrete tokens — words or sub-word units — that allow a single architecture to operate across all language domains. The physical world has no agreed-upon equivalent, which has made the analogy between LLM scaling and world model scaling theoretically suspect. Gaussian splatting offers a candidate: a 3D Gaussian is a compact, differentiable, fully learnable unit that encodes position, geometry, color, and opacity — a building block that world models could predict, update, and reason over the way LLMs manipulate tokens. ECCV 2026 does not resolve this question, but the breadth of the Gaussian splatting program — spanning robotics, SLAM, autonomous driving simulation, dynamic scene generation, and medical imaging — suggests the field is testing Gaussians as that universal physical substrate in practice, even without a settled theoretical answer.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
