CVPR 2026 Recap: What Actually Happened in Denver

TL;DR: CVPR 2026 ran June 3-7, 2026 at the Colorado Convention Center in Denver. The 43rd IEEE/CVF Conference on Computer Vision and Pattern Recognition fielded a record 16,092 paper submissions and accepted 4,089 for a 25.42% acceptance rate (checked 2026-08-24, cvpr.thecvf.com/Conferences/2026/News/Technical_Program). Google DeepMind, UCL, and Oxford took Best Paper for D4RT, a 4D scene-reconstruction model that replaces a whole depth-plus-flow-plus-pose pipeline with one query interface. CVPR 2027 lands in Seattle, June 20-24, 2027.

I track CVPR the same way I track AI events generally, as an early signal for what the SaaS tools I test in 12 to 24 months will actually be built on. CVPR is not a corporate keynote conference like NVIDIA GTC. It is the biggest peer-reviewed academic venue in computer vision, and its winners tend to show up as shipping features long before they show up in a product demo.

This page is a recap, not an attend-or-skip guide. CVPR 2026 ended over two months ago. Below is what happened, which papers won, what the three keynotes actually covered, and what the submission numbers say about where vision research is heading. Every claim traces to CVPR’s own release, the IEEE Computer Society, or the project pages behind each award-winning paper.

CVPR 2026 dates and venue reflect a Wednesday-to-Sunday run

CVPR 2026 ran June 3-7, 2026 at the Colorado Convention Center in Denver, not June 2-6 as an aggregator listing had it. Workshops and tutorials occupied June 3-4. The main technical program, paper presentations, keynotes, and awards, ran June 5-7 (checked 2026-08-24, cvpr.thecvf.com). More than 100 exhibitors set up in the Expo Hall alongside the main sessions.

Detail	Value
Event	CVPR 2026 (43rd IEEE/CVF Conference on Computer Vision and Pattern Recognition)
Dates	June 3-7, 2026 (workshops Jun 3-4; main conference Jun 5-7)
Venue	Colorado Convention Center, Denver, Colorado, USA
Format	Onsite plus virtual track
Submissions	16,092 (up 24% year over year)
Accepted	4,089 (25.42% acceptance rate)
Program	31 workshop tracks, 11 tutorials
Expo	100+ companies, June 5-7
Keynotes	3 (Latent Labs, IBM, Brown University)
Best Paper	D4RT (Google DeepMind, UCL, Oxford)
Next edition	CVPR 2027, Seattle, June 20-24, 2027

Program co-chair Alexander G. Schwing, an associate professor at UIUC, put the submission trend this way in the official 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.” This is a conference getting harder to get into every year, not easier.

The three keynotes were biology, quantum, and why AI still can’t see like a human

CVPR ran exactly three keynotes across the three main-conference days. No unifying theme, no product launches, no single headline reveal. That is a feature, not a bug, because it lets each speaker be genuinely different.

Simon Kohl, founder and CEO of Latent Labs (June 5). Kohl previously co-led DeepMind’s protein-design team and was a senior researcher on the Nobel Prize-winning AlphaFold2 project. His argument: biology is becoming programmable in the same sense vision became learnable, with generative models designing molecules from scratch rather than screening existing compounds. He cited Latent-X and Latent-X2 producing lab-validated antibody and peptide candidates at hit rates matching traditional screening.

Jerry Chow, IBM Fellow and CTO for Quantum-Centric Supercomputing (June 6). Chow co-led the 2016 launch of IBM Quantum Experience, the first cloud-accessible quantum computer. His talk covered quantum systems integrating directly into supercomputing environments, arguing quantum has moved past the scientific-curiosity phase. This sat furthest from typical CVPR subject matter, which was deliberate.

Thomas Serre, Brown University cognitive science and CS professor (June 7). This was the keynote most relevant to CVPR’s core audience, and the most quietly damning. Serre’s argument: deep networks match or exceed human accuracy on ImageNet, but on simple cognitive-science probes even the largest models drop to near-chance, and the gap widens as models scale up. His proposed fix is not more scale but developmental learning principles borrowed from biological vision. Naturalistic video training and brain-inspired recurrent state-space models close some of the gap. If the vision community accepts that argument, “scale is all you need” gets a lot more expensive to defend.

The award papers all sit outside classic detection and segmentation

CVPR’s paper awards are the closest thing this conference has to a headline moment. The official awards list confirms all five for 2026.

Best Paper: “Efficiently Reconstructing Dynamic Scenes One D4RT at a Time.” A 14-researcher team from Google DeepMind, University College London, and Oxford, including DeepMind names Andrew Zisserman, Zoubin Ghahramani, and Raia Hadsell, won for D4RT (Dynamic 4D Reconstruction and Tracking). Recovering the geometry and motion of a moving scene from video historically meant chaining separate depth, optical-flow, and camera-pose models, then fusing outputs with optimization steps that break exactly on moving objects. D4RT collapses that whole pipeline into one transformer that encodes a video once, then answers a single query: pixel location plus source timestamp, target timestamp, and camera frame. The same model produces depth maps, point tracks, camera parameters, or full point clouds depending on which timestamps you vary. Per the team’s project page, D4RT sets state of the art across every 4D benchmark tested and specifically outperforms VGGT on scenes with moving objects. Model weights are not public yet.

Best Student Paper: TRELLIS.2. An 11-author team led by Jianfeng Xiang won for “Native and Compact Structured Latents for 3D Generation,” which shipped publicly as TRELLIS.2, the successor to Microsoft’s popular open-source TRELLIS model. Its O-Voxel representation generates 3D assets with arbitrary topology, open and non-watertight surfaces as easily as closed meshes, while encoding full PBR materials so outputs are game-ready without a separate texturing pass.

Best Paper Honorable Mentions. Three papers: “ChordEdit,” a fast single-step image-editing approach; “SAM 3D,” a Meta FAIR team extending Segment Anything into 3D object generation from single images, with Piotr Dollár, Georgia Gkioxari, Matthijs Feiszli, and Jitendra Malik on the author list; and “NitroGen,” an NVIDIA-led vision-action foundation model for generalist gaming agents trained on 40,000 hours of gameplay video across 1,000+ games. NitroGen belongs on any watchlist covering the agent-infrastructure stack, next to my MCP servers roundup.

None of the five award papers are classic detection, segmentation, or tracking work. That is a real signal about where the field is going.

Research themes shifted hard toward multimodal and world models

Independent analysis of a highlighted-paper sample from the 4,089 accepted showed multimodal language and vision-language work growing from 4.9% to 10.6% year over year, the largest swing in the dataset. Video generation and world models came second, 3.8% to 8.8%. Embodied AI and robotics grew from 2.9% to 6.2%. Classic detection, segmentation, and tracking, the categories that built CVPR’s 2010s reputation, fell from 3.8% to 1.2%. Worth flagging: the breakdown draws from a 500-paper highlight sample, not the full 4,089. Direction is reliable. Exact percentages are a snapshot.

Program co-chair Chen Change Loy, a professor at Nanyang Technological University, put the shift this way in the official 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 grew substantially this year, in his framing, though that work is still earlier stage than the generative and multimodal categories.

The Expo floor matched the research shift. Tesla, Waymo, NVIDIA (showing the new Nemotron 3 Nano Omni multimodal agent model), Adobe (Firefly workflows), Ultralytics (YOLO26 for edge), and Weights & Biases all set up in Denver.

I could not verify a 2026 attendance figure

CVPR’s boilerplate says the conference is “annually attended by more than 10,000 scientists and engineers.” That is a standing multi-year claim, not a 2026 headcount, and I could not find a post-event number CVPR published for this specific edition. A pre-event third-party report estimated 9,000+ in person, but that is an estimate, not a confirmed count.

You get 16,092 submissions, 4,089 accepted papers, 31 workshop tracks, 11 tutorials, and 100+ exhibitors. Submissions alone are an auditable proxy: on that measure CVPR 2026 was the largest edition in the conference’s history. I would rather say the attendance number is not public than round an estimate into something it is not.

My honest take on CVPR 2026

For the audience CVPR is built for (vision researchers, robotics and AV engineers, applied vision-product teams tracking where research heads before it ships), CVPR 2026 delivered. D4RT is a genuinely clever piece of engineering rather than a benchmark-chasing increment, and the fact that its weights are not yet public tells you this is still frontier research.

The honest caveat applies to every academic conference this size: unless you actively publish or review vision research, the value is diffuse, not concentrated in one video 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.

If your work touches computer vision, multimodal AI, robotics perception, or 3D generation directly, the 2026 paper program is worth digging into now. Both the D4RT and TRELLIS.2 project pages plus arXiv papers are public today, even though D4RT weights are not. If you are tracking where CVPR goes next, CVPR 2027 is confirmed for Seattle, June 20-24, 2027.