The Map of CVPR 2026

CVPR 2026 accepted 5,010 papers — 4,069 in the main conference and 941 more in the new Findings track. That is far more than any one person can read — but it is small enough to see all at once. I embedded every paper from its title and abstract, projected the whole corpus down to two dimensions, and let the topics arrange themselves. What follows is a tour of that map and a few things it reveals about where computer vision actually is this year.

One dot per paper

Below is the entire conference. Each point is one accepted paper, placed so that papers about similar things sit near each other; colors are 24 topic clusters the data fell into on its own. Hover to read titles, click a cluster in the legend to isolate it, and zoom into any neighborhood you care about — or use the toggle at the top right to recolor the map by track, main conference vs. the new Findings papers.

All 5,010 CVPR 2026 papers, embedded with a sentence-transformer and projected to 2D with UMAP. Color = automatically discovered topic cluster.

The first thing the map makes obvious is how generative and language-shaped the field has become. The big continents are diffusion-based image and video generation, vision-language models, and 3D reconstruction — and they are no longer separate islands. They bleed into each other, because the same backbones (diffusion transformers, CLIP-style encoders, large multimodal models) now show up everywhere.

The new Findings track

CVPR 2026 introduced a Findings track — 941 papers (19% of the total) accepted as solid contributions a notch below the main program. Flip the map to Main vs Findings and the two overlap almost everywhere: Findings is mostly a mirror of the main conference, not a different field. But it isn't uniform. Findings is over-represented in the applied, incremental corners — deepfake forensics (31% of that cluster), computational imaging and compression (26%), medical imaging (25%), and VLM efficiency (25%) — and under-represented in the buzzy frontier it leaves to the main track: 3D reconstruction and generation, video generation, and 3D spatial reasoning (each ~12–15%).

What the conference is about

Collapsing the map into cluster sizes gives the topic landscape, and this year it is strikingly flat — no single topic runs away with the conference. The largest clusters each hold 250–350 papers (5–7%): open-vocabulary detection and segmentation, multimodal reasoning, vision-language-action and autonomous driving, diffusion-based generation and super-resolution, video generation, human motion, and image editing. Behind them sits a long tail of 3D reconstruction, continual / federated learning, and pose / depth estimation.

The 24 clusters, sized by paper count and labeled by hand from each cluster's representative titles.

The ideas everyone is building on

Cutting a different way: how many papers even mention a given idea, anywhere in their title or abstract? This is the closest thing to a popularity contest for techniques.

Share of the 5,010 papers mentioning each idea (document frequency over title + abstract). A paper can count toward several.

A few things jump out:

• Multimodal (22%), diffusion (19%), and video (18%) are now the water everyone swims in. Add vision-language models (18%) and LLMs (13%) and the generative-multimodal core is unmistakable.
• Gaussian Splatting has lapped NeRF. 3DGS appears in 5.1% of papers; NeRF and radiance fields in just 1.0% — a ratio that was reversed only two years ago (see the trend below).
• Mamba / state-space models stayed niche at 1.2% — having only just edged past NeRF. The transformer (11% explicit, far more implicit) is still the default, despite the hype cycle around its successors.

What grew and what faded

With the 2024 and 2025 proceedings scraped too, the snapshot becomes a trajectory. The three-year story is the generative-multimodal takeover: multimodal papers more than doubled (10% → 21% of the main track), vision-language models did the same (7% → 17%), and explicit LLM mentions climbed from 7% to 13%. The sharpest newcomer is reinforcement learning — flat near 1% through 2025, then exploding to 6.6% in 2026 as RL-style post-training swept into vision.

Document frequency of each idea across the CVPR 2024, 2025 and 2026 main tracks (Findings excluded for a like-for-like comparison).

The chart also settles the NeRF question. In 2024, NeRF / radiance fields (5.2%) led 3D Gaussian Splatting (2.0%); by 2025 splatting had lapped it (6.5% vs 2.5%), and in 2026 NeRF has all but vanished (1.1%) while splatting holds near its peak. A few ideas cooled too — point clouds, CLIP, and zero-/few-shot framings all drifted down as the field's center of mass moved to generation and reasoning.

How topics connect

The concept words in paper titles co-occur in telling ways. Drawing the strongest 150 co-occurring term pairs as a network shows the field's wiring: a dense vision-language-reasoning hub (language, vision, visual, multimodal, reasoning) sits at the center, wired to a generation-and-diffusion cluster (generation, diffusion, editing, latent, flow). Tighter satellite communities orbit it — detection / segmentation, 3D reconstruction (gaussian, splatting, view, scene), pose / depth estimation, and a small autonomous-driving knot. generation, language, and vision are the most-connected hubs.

Top 150 co-occurring title terms (connected groups only). Node size = how many papers use the term; color = community (greedy modularity); layout = force-directed.

Who ships their code

Roughly a quarter of papers (25%) link a public repo or project page in their abstract — but the rate swings widely by topic. The most reproducible corners are the classic perception tasks: 3D detection and geo-localization (33%), computational imaging and compression (32%), medical imaging (32%), and open-vocabulary detection (30%). The most closed are the frontier generative areas — 3D shape and scene generation (16%), Gaussian Splatting (18%), and the RL / VLA / agent clusters (~19%) — where the work is newer, heavier, and often closer to a product.

Share of each cluster's papers that link code or a project page in the abstract. Dashed line = the 25% conference-wide average.

How the field names its work

Finally, a lighter look — the linguistics of the titles themselves. Just under three out of four titles (72%) now use the Name: Description colon format. The dramatic "X is All You Need" meme has almost died out (just 6 titles), and question-titles stay rare (23). Papers like to open with Learning, Towards, Beyond, and Rethinking — and the most-reused acronyms (MLLM, VLM, CLIP, LLM, LiDAR, GUI) are themselves a snapshot of the year's obsessions.

Title patterns, length distribution, opening words, and the most-reused method names across all 5,010 titles.

And the people

Behind the 5,010 papers are 20,671 distinct authors. The typical paper now carries six authors (mean 6.3); the largest single author list runs to forty. Computer vision has become, decisively, a team sport.

Method & caveats. This is a snapshot of the accepted papers — 4,069 main-conference plus 941 Findings — scraped from CVF Open Access. Embeddings come from the all-MiniLM-L6-v2 sentence-transformer, so proximity on the map means semantic similarity of text — not citations or impact. Clusters and the 2D layout are unsupervised; the 24 cluster labels were written by hand from each cluster's representative titles, so read them as helpful approximations rather than official categories. Idea counts are regex matches over each paper's title and abstract; the concept network is built from title-word co-occurrence; and the open-source rate is a heuristic (does the abstract link a repo or project page?), so it undercounts code released only after publication. Year-over-year trends compare the CVPR 2024–2026 main tracks for consistency. Charts are interactive (Plotly); the code lives on GitHub.