CVPR 2026 Recap: What Actually Happened in Denver

TL;DR: CVPR 2026 ran June 3-7, 2026, at the Colorado Convention Center in Denver, Colorado - not June 2-6 as the ACF date field on this page previously showed. The 2026 edition of the IEEE/CVF Conference on Computer Vision and Pattern Recognition fielded a record 16,092 paper submissions, up 24% year over year, with 4,089 accepted for a 25.42% acceptance rate. Google DeepMind, University College London, and the University of Oxford took the Best Paper award for D4RT, a 4D scene-reconstruction model that replaces an entire pipeline of specialized tools with one query interface. Three keynotes covered AI-designed drugs, quantum-centric supercomputing, and the gap between how deep networks and human brains actually see.

I track AI research events the same way I test AI SaaS tools before recommending them: check what actually happened against primary sources, not whatever a directory site guessed months in advance. CVPR is not a corporate keynote conference like NVIDIA GTC - it is the biggest peer-reviewed academic venue in computer vision, and its findings tend to show up as shipping features in AI tools 12 to 24 months later. That is why it sits on my AI events hub.

This is a recap, not a “should you attend” guide - CVPR 2026 already happened over a month ago. Here is the real dates and venue, what the keynotes covered, who won the paper awards, and what the submission numbers say about where computer vision research is heading. Every claim below traces to CVPR’s own site, the IEEE Computer Society, or the project pages behind the award-winning papers, and where I could not verify a number independently, I say so instead of rounding it into something cleaner than it is.

Key Takeaways

• The real dates were June 3-7, 2026, not June 2-6. CVPR’s own conference site confirms a Wednesday-to-Sunday run at the Colorado Convention Center in Denver: workshops and tutorials ran June 3-4, the main technical program ran June 5-7. The “June 2-6” date previously attached to this page does not match the official CVPR 2026 site.
• Submissions hit a record 16,092, up 24% over 2025. Of those, 4,089 papers were accepted, a 25.42% acceptance rate that CVPR’s own program co-chairs describe as “consistently in the low-to-mid 20% range” even as submission volume has more than doubled in five years.
• Google DeepMind won Best Paper for D4RT, a 4D scene-reconstruction model built with University College London and the University of Oxford. Microsoft’s TRELLIS.2 team took Best Student Paper, and Meta’s Segment Anything lineage picked up a Best Paper Honorable Mention with SAM 3D.
• Three keynotes, zero product launches. CVPR does not run a single vendor keynote the way NVIDIA GTC does. Instead, three separate named speakers covered AI-designed drug candidates (Latent Labs), quantum-centric supercomputing (IBM), and why deep networks still fail basic vision tests humans pass easily (Brown University).
• No confirmed official attendance figure exists for the 2026 edition specifically. CVPR’s own boilerplate says the conference is “annually attended by more than 10,000 scientists and engineers,” but that is a generic multi-year claim, not a 2026 headcount, and I could not find a post-event number that CVPR published for this specific edition.

What Is CVPR?

CVPR, the IEEE/CVF Conference on Computer Vision and Pattern Recognition, is the largest annual academic conference in computer vision and pattern recognition, co-sponsored by the IEEE Computer Society and the nonprofit Computer Vision Foundation (CVF). It is a peer-reviewed research venue, not a trade show: most of the program is built from submitted papers that pass blind review, plus workshops, tutorials, and a doctoral consortium around the main sessions.

That distinction matters for reading this recap. There was no Jensen Huang-style two-hour keynote compressing a year of product news into one video - just three separate, unrelated keynotes on three separate days, each tied to a specific piece of research. CVPR’s headline “announcements” are not product launches; they are the Best Paper and Best Student Paper awards, decided by a committee delegated by the program chairs.

CVPR’s academic weight is real and independently measurable: past proceedings rank number two in Google’s 2025 Scholar Metrics, ahead of most traditional journals, and Research.com ranks CVPR as the top-cited venue for computer science, image processing and computer vision, and machine learning and AI combined. If you build AI products that touch images, video, or spatial understanding, CVPR is where the underlying research gets published before it shows up in the tools reviewed on my best AI tools list.

CVPR rotates host cities each year. The 2026 edition landed in Denver; the next one, CVPR 2027, is already confirmed for Seattle in June 2027.

What Happened at CVPR 2026: Dates, Venue, and Scale

CVPR 2026 ran June 3-7, 2026, at the Colorado Convention Center in Denver, Colorado, according to CVPR’s own conference site. Workshops and tutorials occupied the first two days, June 3-4, with the main technical program - paper presentations, keynotes, and awards - running June 5-7. The Expo, where more than 100 companies exhibited, ran June 5-7 inside the same venue’s Exhibit Hall. That corrects the “June 2-6, 2026” date previously attached to this page, which starts and ends a day earlier than the actual Wednesday-through-Sunday run.

Quick facts:

Detail	Information
Event	CVPR 2026 (43rd IEEE/CVF Conference on Computer Vision and Pattern Recognition)
Dates	June 3-7, 2026 (workshops/tutorials Jun 3-4; main conference Jun 5-7)
Venue	Colorado Convention Center, Denver, Colorado, USA
Organizers	IEEE Computer Society (CS) and the Computer Vision Foundation (CVF)
Format	Onsite plus virtual track
Submissions	16,092 (up 24% over CVPR 2025)
Accepted papers	4,089 (25.42% acceptance rate)
Program tracks	31 workshop tracks, 11 tutorials
Expo	100+ exhibiting companies, June 5-7, Exhibit Hall
Keynotes	3 (Latent Labs, IBM, Brown University)
Best Paper	“Efficiently Reconstructing Dynamic Scenes One D4RT at a Time” (Google DeepMind, UCL, Oxford)
Next edition	CVPR 2027, June 20-24, 2027, Seattle, Washington
Official recap source	CVPR’s official technical program announcement

The submission growth is the number that tells you something about the field, not just the venue. Program co-chair Alexander G. Schwing, an associate professor at the University of Illinois Urbana-Champaign, put it this way in CVPR’s own release: “CVPR submissions have more than doubled over the past five years, but the acceptance rate has remained highly competitive, consistently in the low-to-mid 20% range.” That is a conference getting harder to get into every year, not easier, even as AI research funding and headcount have both expanded.

Keynote Recap: Biology, Quantum Computing, and Why AI Still Can’t See Like Humans

CVPR 2026 ran exactly three keynotes, all in the Bluebird Ballroom, one per day across the three main-conference days. There was no unifying theme beyond “cutting-edge research a computer vision audience would find genuinely interesting” - a different bar than a corporate keynote clears, and no reason to assume the sessions were low-value just because there was no single headline reveal.

June 5: Simon Kohl, founder and CEO of Latent Labs, “Programmable Biology: Generative AI for Molecular Design.” Kohl previously co-led DeepMind’s protein design team and was a senior research scientist on the Nobel Prize-winning AlphaFold2 project before founding Latent Labs. His talk argued that biology is becoming programmable in the same sense vision became learnable: instead of screening millions of existing compounds, Latent Labs’ models design molecules from scratch. He cited Latent-X and Latent-X2, generative models producing lab-validated antibody and peptide candidates with hit rates matching traditional screening, and previewed Latent-Y, which the company describes as the first autonomous biologics design AI agent - pitched to the CVPR audience as sharing deep structural similarities with multi-modal conditioning and agentic reasoning problems computer vision has spent a decade solving.

June 6: Jerry Chow, IBM Fellow and CTO for Quantum-Centric Supercomputing, “Transforming Computing with Quantum-Centric Supercomputing.” Chow, a physicist who co-led the 2016 launch of IBM Quantum Experience - the first cloud-accessible quantum computer - argued that quantum computing is moving past the “scientific curiosity” phase into results comparable to leading classical methods, particularly when quantum systems integrate directly into supercomputing environments. This keynote sits furthest from typical CVPR subject matter, and that is deliberate: the program committee increasingly uses keynote slots to expose a vision-focused audience to adjacent fields.

June 7: Thomas Serre, Professor of Cognitive and Psychological Sciences and Computer Science at Brown University, “Scaling Laws vs. Neural Laws: Toward More Natural Artificial Vision.” This was the keynote most relevant to CVPR’s core audience, and arguably the most quietly damning. Serre’s argument: deep networks now match or exceed human accuracy on benchmarks like ImageNet, but the resemblance to human vision is fragile - on simple probes drawn from cognitive science, even the largest models drop to near-chance performance, and the gap widens, not narrows, as models scale up. His proposed fix is not more scale but closer alignment with “neural laws”: developmental learning principles and architectural constraints borrowed from biological vision. He presented early evidence that naturalistic video training and brain-inspired recurrent state-space models in place of transformer self-attention close some of that gap - a talk that undercuts the industry’s default assumption that bigger models straightforwardly mean better vision.

Best Paper, Best Student Paper, and the Honorable Mentions

CVPR’s paper awards, not a keynote, are the closest thing this conference has to a headline moment, and the official awards list maintained by the IEEE Computer Society’s Technical Committee on Pattern Analysis and Machine Intelligence confirms all of the following for 2026.

Best Paper: “Efficiently Reconstructing Dynamic Scenes One D4RT at a Time.” A 14-researcher team from Google DeepMind, University College London, and the University of Oxford - including well-known DeepMind names Andrew Zisserman, Zoubin Ghahramani, and Raia Hadsell - won for D4RT, short for Dynamic 4D Reconstruction and Tracking. Recovering the geometry and motion of a moving scene from video has historically meant chaining together separate depth, optical-flow, and camera-pose models, then fusing their outputs with optimization steps that break down exactly on moving objects. D4RT replaces that pipeline with one transformer that encodes a video once, then answers a single kind of query - a pixel location plus a source timestamp, target timestamp, and camera frame - that returns a 3D position, with the same model producing depth maps, point tracks, camera parameters, or a full point cloud depending on which timestamps you vary. Per the team’s own project page, D4RT sets a new state of the art across every 4D reconstruction benchmark tested and outperforms the widely used VGGT method specifically on scenes with moving objects, which VGGT either skips or handles through expensive iterative refinement. As one independent technical breakdown put it, D4RT “refuses to treat dynamic objects as a special case.” Model weights had not been publicly released as of this writing.

Best Student Paper: “Native and Compact Structured Latents for 3D Generation.” An 11-author team led by Jianfeng Xiang won for what shipped publicly as TRELLIS.2, the successor to Microsoft’s popular open-source TRELLIS 3D-generation model. Its O-Voxel representation generates 3D assets with arbitrary topology - open, non-watertight surfaces as easily as closed meshes - while also encoding full physically based rendering (PBR) materials, not just surface color, so the model outputs game-ready assets directly without a separate texturing pass.

Best Paper Honorable Mention (three papers, all named on the official award list):

• “ChordEdit: One-Step Low-Energy Transport for Image Editing” - a six-author team’s approach to fast, single-step image editing.
• “SAM 3D: 3Dfy Anything in Images” - a large Meta FAIR team, including Segment Anything Model veterans Piotr Dollár, Georgia Gkioxari, Matthijs Feiszli, and Jitendra Malik, extending the Segment Anything lineage from 2D segmentation into full 3D object generation from single images.
• “NitroGen: An Open Foundation Model for Generalist Gaming Agents” - a collaboration led by NVIDIA researchers with Stanford, Caltech, the University of Chicago, and UT Austin. Per CVPR’s own technical program announcement, NitroGen is “a vision-action foundation model for generalist gaming agents…trained on 40,000 hours of gameplay videos across more than 1,000 games” - worth a look next to my breakdown of the best MCP servers, which tracks the agent-infrastructure ecosystem this kind of foundation model tends to feed into.

The pattern across all five award papers matches what CVPR’s own program committee flagged: the highest-submission-volume areas in 2026 were image and video synthesis, vision-language and reasoning, multi-modal learning, 3D reconstruction from multiview sensors, and medical or biological vision. None of the award winners are classic detection, segmentation, or tracking work - the categories that defined CVPR a decade ago.

Industry Presence and Research Themes at CVPR 2026

CVPR is not a single-vendor product event, but industry showed up in Denver in a way that tells you something concrete about where computer vision is heading commercially. Per CVPR’s own Expo press release, more than 100 companies exhibited across the June 5-7 Expo floor, spanning a genuinely wide range of what “computer vision” now means commercially: Tesla (vision-only autonomous driving, Cybertruck robotics), Waymo (sensor-fusion work scaling its robotaxi service), NVIDIA (its newly released Nemotron 3 Nano Omni model unifying vision, audio, and language into one multimodal agent system), Adobe (Firefly AI Assistant workflows across Creative Cloud), Ultralytics (YOLO26, a lightweight real-time vision model for edge deployment), and Weights & Biases (its MLOps platform for training large-scale vision models in production).

That exhibitor list matches the research-side shift in independent analysis of the 4,089 accepted papers. A research-trend breakdown found multimodal language and vision-language model papers grew from roughly 4.9% to 10.6% of a highlighted-paper sample year over year - the largest swing in the dataset, enough to make vision-language work the largest individual theme at the conference for the first time. Video generation and world models came second (3.8% to 8.8%), and embodied AI and robotics grew from 2.9% to 6.2%. Classic computer vision work - detection, segmentation, and tracking, the categories that built CVPR’s 2010s reputation - fell from 3.8% to 1.2%. Worth flagging: that breakdown draws from a roughly 500-paper highlight sample, not the full 4,089-paper population, so the direction is reliable but the exact percentages are a snapshot, not a census.

Program co-chair Chen Change Loy, a professor at Nanyang Technological University in Singapore, framed the trend this way in CVPR’s own release: “As fundamental concepts of computer vision permeate new applications, we’re seeing a rise in submitted research that corresponds with particular disciplines.” Medical and biological vision and cell microscopy specifically grew “substantially” this year, in his words, though that work remains in earlier stages than the more mature generative and multimodal categories.

Attendance and Scale: How Big Was CVPR 2026, Really?

This is a section where I have to be upfront about a gap. CVPR’s own boilerplate “about” language, reused across its press materials, describes the conference as “annually attended by more than 10,000 scientists and engineers from across the globe.” That is a standing, multi-year claim, not a number specific to the 2026 Denver edition, and I could not find an official CVPR-published post-event headcount for 2026 specifically.

The closest thing to a number I found came from a research-trend report published before the conference started, which estimated “9,000+ in person” expected attendance based on the scale of the accepted-paper program and prior-year patterns. That is a pre-event third-party estimate, not a confirmed count from CVPR itself, and I am not going to present it as more solid than it is.

What I can confirm with certainty: 16,092 paper submissions, 4,089 accepted papers, 31 workshop tracks, 11 tutorials, and more than 100 exhibiting companies across a three-day Expo. Submissions alone are a real, auditable proxy for scale, and on that measure CVPR 2026 was the largest edition in the conference’s history, consistent with Schwing’s “more than doubled in five years” framing. I would rather tell you plainly that an exact headcount is not yet public than round a pre-event estimate into something it is not.

Was CVPR 2026 Worth It? My Honest Take

For the audience CVPR is actually built for - computer vision and machine learning researchers, robotics and autonomous-vehicle engineers, and applied vision-product teams tracking where research is heading before it ships - CVPR 2026 delivered real substance. The Best Paper award went to a genuinely clever piece of engineering in D4RT, not a benchmark-chasing incremental result, and the fact that it comes from a cross-institution team with model weights not yet public tells you this is still frontier research, not something already commoditized into a product.

The honest caveat applies to every academic conference this size: unless you are actively publishing or reviewing vision research, the value is diffuse, not concentrated in one keynote you can watch in two hours. NVIDIA GTC compresses a year of hardware news into a single free livestream. CVPR spreads its substance across 4,089 papers, three unrelated keynotes, and a 100-company Expo floor, and pulling a coherent narrative out of that takes real digging - which is exactly what the research-trend breakdown above surfaces.

If your work touches computer vision, multimodal AI, robotics perception, or 3D generation directly, CVPR 2026’s paper program was worth tracking closely. If you are reading about it purely for general AI awareness, the Best Paper and Best Student Paper writeups above cover the two results most likely to matter outside a narrow academic audience.

Frequently Asked Questions

When did CVPR 2026 take place?

CVPR 2026 ran June 3-7, 2026, at the Colorado Convention Center in Denver, Colorado. Workshops and tutorials ran June 3-4, and the main technical program ran June 5-7. This corrects an earlier “June 2-6, 2026” date that did not match CVPR’s own official conference site.

What was the Best Paper award at CVPR 2026?

“Efficiently Reconstructing Dynamic Scenes One D4RT at a Time,” from a 14-author team at Google DeepMind, University College London, and the University of Oxford, won Best Paper. D4RT replaces separate depth, tracking, and camera-pose models with one unified query interface for 4D scene reconstruction.

How many papers were submitted and accepted at CVPR 2026?

CVPR 2026 received a record 16,092 paper submissions, a 24% increase over 2025, and accepted 4,089 for a 25.42% acceptance rate. Submission volume has more than doubled over the past five years, per CVPR’s own program co-chairs.

How many people attended CVPR 2026?

CVPR has not published an official post-event attendance figure specific to 2026 as of this writing. Its standing boilerplate describes the conference as “annually attended by more than 10,000 scientists and engineers,” a generic multi-year claim, and a pre-event third-party estimate put expected in-person attendance at 9,000-plus - but neither figure is a confirmed 2026 headcount.

Who gave keynotes at CVPR 2026?

Three speakers presented across the three main-conference days: Simon Kohl, founder and CEO of Latent Labs and former DeepMind protein-design researcher, on generative AI for drug discovery; Jerry Chow, IBM Fellow and CTO for Quantum-Centric Supercomputing, on integrating quantum computing with AI infrastructure; and Thomas Serre, a Brown University professor, on the gap between deep-network vision and human vision.

Is CVPR good for people outside academic computer vision research?

CVPR is built primarily for researchers, PhD students, and engineers directly applying published computer vision research. It is a strong fit for robotics, autonomous-vehicle, and applied vision-product teams, but the paper-heavy program assumes technical fluency and is not designed as a beginner-friendly AI conference.

The Bottom Line

CVPR 2026 ran June 3-7, 2026, in Denver, Colorado, not June 2-6 as this page previously stated - and that correction matters less than what actually happened once the conference started: a record 16,092 submissions, a genuinely inventive Best Paper in D4RT from Google DeepMind, and a research program that has visibly tilted toward vision-language models, video generation, and embodied AI at the expense of the detection-and-segmentation work that used to define this conference.

The insight worth taking with you: check the primary source - CVPR’s own technical program release and the official award list - before repeating whatever a secondary aggregator claims. The submission and acceptance numbers are real and auditable. The attendance figure is not yet public, and I would rather say that plainly than invent a clean-sounding round number.

Your concrete next step: if D4RT or TRELLIS.2 are relevant to what you are building, both project pages and their arXiv papers are public today even though D4RT’s weights are not yet released. And if you are tracking where CVPR goes next, CVPR 2027 is already confirmed for Seattle, June 20-24, 2027. Want a heads-up the moment new AI event coverage and deal alerts go live? Subscribe here.