One video. One face. A much wider world.
It turns football figures into animated characters, placing familiar faces inside an imagined World Cup story.
How do we identify recurring visual identities across thousands of these videos? Follow one appearance from this moving scene into the research records.
Video 999359659607952 is 170.96 seconds long: 4,274 frames at 25 frames per second. Playback is optional; the saved images carry the explanation.
A video is a sequence of frames. Different moments can contain very different faces.
These are a few saved frames from our example. We will keep returning to the same selected moment as the view expands.
These stills were chosen for this explanation. The pipeline first tracks people, then samples frames inside retained tracks. This package numbers frames from one; saved times use frame / 25. Frame 3531 has saved time 141.24 s and playback time 141.20 s.
The first model locates people in each frame. A second step links detections across nearby frames, giving each tracked appearance a number.
These numbers organize the moving scene. They are not player names.
YOLO11 detects people; BoT-SORT associates detections. Track 3979, frame 3531: person box [194, 102, 850, 712] in the original 1280 × 720 image. Person-detection confidence 0.936375 is not identity confidence.
A track lets us examine several moments of a recurring appearance instead of treating every frame as a new person.
We sample retained tracks and look for faces with usable detail. Quality and diversity checks keep a small set of observations for comparison.
The paper discards tracks with fewer than five frames, samples up to 15 frames per retained track, and detects faces with SCRFD inside person crops. Size, sharpness and pose inform selection; temporal and perceptual diversity reduce repeated views. At most five faces are kept per track. Rejected candidates and individual rejection reasons were not saved.
The saved crop separates an appearance from the rest of the scene. Its frame and track still connect it to the source video.
The other crops come from different moments in this same video. They show several appearances, including other players; each keeps its own frame and track.
The crop is linked by its saved video, frame and track IDs. Face-localization boxes, alignment landmarks and crop transforms are absent, so the connector shows a metadata relationship, not a pixel-exact extraction. Each image is a separate observation; grouping and identity review come later in this explanation.
Each face becomes a list of numbers: an embedding. This shared representation lets the analysis compare appearances even when images differ.
The numbers describe visual features. They are neither a name nor a probability.
The paper uses five-landmark alignment and ArcFace to produce L2-normalized 512-dimensional vectors. The numeric excerpt is read from row 37 of the saved embedding array, joined to track 3979, frame 3531 and its crop path. No model runs in this page. The saved crop is not the missing aligned model-input image.
Across videos, the analysis gathers visually similar descriptions into small groups, then combines group representatives into final visual clusters.
Our followed face has a recorded route from initial group 1562 to final cluster 0. It still has no name in this explanation.
HDBSCAN groups the embeddings with Euclidean distance (minimum cluster size 3; min_samples 1). Normalized representatives of the most typical 50% of members are combined by complete-link hierarchical clustering with cosine distance. Initially noisy observations can be reassigned at cosine similarity ≥ 0.45. Merge history and global projection coordinates are not supplied. The on-screen arrangement and reveal order do not measure similarity or replay computation.
People reviewed the visual clusters and added a football figure’s name where they could identify one.
Cluster 0 carries the reviewed label Lamine Yamal. Other clusters remained unidentified. A track, a visual group and a named person are different kinds of evidence.
500 cluster rows were reviewed; 114 have a name, spanning 18 player labels. Several clusters can share a name. An individual manual-review verdict for the followed crop is not supplied. Unassigned faces are not automatically low-quality or incorrect.
Take a closer look at two of those outside appearances. They come from different video records, but share final cluster 0 and its reviewed label.
This group contains 330 crop observations across 145 video records. The connection comes from the identity analysis; it does not show that one video copied another.
The outside appearances have recorded initial groups 2058 and -1 and share final cluster 0. Global clustering happened before human naming. Lines show final membership, not pairwise similarity, copying or an audit of reassignment. Their source links open metadata records; outside-video playback is not available here.
Pull back, and our group becomes one of 114 named visual clusters in the saved analysis.
Each of the 18 players has its own color. Several clusters can carry the same player's name and therefore share a color.
One uniform tile represents one named cluster, ordered by recorded ID. Color identifies the reviewed player label. Tile size and position have no similarity or prevalence meaning. You can inspect the recorded counts by choosing a cluster.
The study collected 2,021 Instagram posts and downloaded 1,901 associated video files. Posts and downloaded videos are different records.
The saved identity table contains 19,094 face observations from 1,548 video records. These are the appearances behind the field we just opened.
17,454 observations belong to 900 visual clusters; 1,640 remain unassigned. The Explorer exposes the named subset: 5,540 observations, 114 named clusters and 18 player labels, linked to 728 videos. Clusters are not unique people, and video records are not necessarily unique pieces of content.
From a moving scene to a face. From that face to a recurring group. From one group to appearances across the collection.
Continue with Lamine Yamal, or start your own path through the reviewed images and their source videos.
The player page opens all 809 named Lamine Yamal records across 13 clusters and 219 videos. Its scope is broader than cluster 0. The link does not automatically select this exact crop; the example-video link opens its source record directly.
5,540 named face observations.
18 player labels. Your own path.
Opens all 809 named records across 13 clusters and 219 videos.
Browse the reviewed collection Inspect the example video’s recordVideo 999359659607952
A guided view of saved analysis. Grouping was computed across videos before this story was assembled.
Move between individual face observations, reviewed player identities, and their source videos.
Open the Data ExplorerThe visual identity analysis is one way to examine the supply of this media: which football figures appear, and how often their appearances recur.
Caption analysis adds the anime references and tools named alongside them.
Views, likes and comments describe different forms of engagement. Total attention and average attention per video offer different perspectives on the same collection.
Content-reuse analysis follows shared material across posts and accounts. It is a separate branch of the research from the shared identity relationships in this story.
Observed posting order and shared material do not establish copying or authorship.
The study collected 2,021 Instagram posts. The downloaded visual corpus contains 1,901 video files; the saved identity analysis covers 1,548 video records. These counts describe different scopes of evidence.
The 2026 World Cup, Anime Edition: Production, Engagement, and Circulation of Generative Memetic Sports Content.
Read the research paper (PDF)A computational social science project studying the production, engagement, and circulation of AI-generated football-anime media on Instagram. This story visualizes saved research outputs; the underlying models do not run in your browser.