Article

What WHAM-RT’s 28-hour experiment means for game creators

For game creators, WHAM-RT tests what selected gameplay can teach a world model—and why 28 hours of footage does not establish cheap training.

Editorial illustration for What WHAM-RT’s 28-hour experiment means for game creators: a geometric block represents a model release. Not documentary evidence.

Microsoft’s WHAM-RT paper, submitted to arXiv on November 17, 2026, examines a problem for game creators: recordings can revisit the same starting areas while barely showing less-travelled routes. The authors report training a model on about 28 hours of gameplay for one controlled character on one Bleeding Edge map. That makes the choice of recordings worth examining, but it does not establish an inexpensive recipe for building your own game world.

A world model here predicts the next image from previous images and controller inputs. Instead of playing a finished video, someone can steer what happens next. The research concerns learning that response from recorded experience—not reconstructing an entire game engine or supplying an exportable game.

What went into the 28 hours?

In Bleeding Edge’s Skygarden map, the researchers found that ordinary play heavily favoured spawn points—the places where players enter—and the corridors leading away from them. Their sampling method used recorded positions to choose starting points spread across the playable area, then retrieved the following 100 seconds of each selected sequence. Rarely visited paths were less likely to disappear from the training set.

The total comes from 1,000 sequences × 100 seconds ÷ 3,600 seconds per hour: about 27.8 hours, rounded to 28. That counts sampled sequence duration. It is not a measurement of collection effort or proof that a fresh 28-hour recording session would capture the same variety.

The researchers selected those sequences from a much larger archive. They also reused WHAM’s image tokenizer—an already-trained component that converts pictures into a representation the model can process. The part learning how the scene changes was trained anew; the complete system did not start from nothing.

For a creator, the useful interpretation is to treat a recording collection as a tour of the intended experience. The practical question is whether the training set includes familiar corridors and rarely visited routes, not simply how many hours it contains.

Did carefully selected footage win?

Not on every measure. The paper compared the curated subset with an equally short random subset and with full-data WHAM-RT. It used Fréchet Video Distance, or FVD, which measures statistical differences between generated and reference videos. Lower scores are better; they are not percentages of correctly reproduced game mechanics.

At 900 frames—about 90 seconds of simulated sequence time—the authors report scores of 488.5 for full-data WHAM-RT, 298.5 for curated data and 297.7 for random data. Both small subsets did better than full-data WHAM-RT, but the random subset was numerically slightly ahead of the curated one. Full-data WHAM-RT did better at the shorter 16- and 300-frame checkpoints. The slower WHAM baseline remained better than all three throughout this comparison.

Curation did improve separate measures of video-feature coverage and density on longer sequences. These describe how generated examples relate to the reference-video distribution, not how much physical map area the model visited. They support a narrower case for curation than “selected footage always wins”.

There is another important limit: the small-data models covered one controlled character, while the full model covered 13. The comparison therefore changes the task as well as the amount of data. It cannot isolate how much improvement came from curation, less data or a narrower job.

Less footage still meant a large training cluster

The authors report using 128 A100 GPUs for 14 hours to train the standard low-data model, or nine hours for a smaller version. Multiplying their stated resources gives 1,792 and 1,152 A100 GPU-hours respectively. Those are calculations, not measured bills, minimum hardware requirements or estimates of how long one GPU would take.

These totals exclude the inherited tokenizer’s training, data collection and preparation, staff time, evaluation and serving the finished model. A short run on a large cluster is not evidence of a small-studio budget. No price for training a customer’s own world is established by the paper or Microsoft’s project page.

Real-time generation has its own tradeoffs

Separately, the paper’s full-model evaluation reports around 17 generated frames per second on one H100. That concerns running a trained model, not building it, and is not a separately established speed for every 28-hour variant. The Bleeding Edge images are only 300 × 180 pixels.

The authors acknowledge a quality tradeoff relative to slower WHAM. In the authors’ small blinded comparison of 11-second clips, six experts supplied 180 judgments: 80 favoured WHAM, 36 favoured full-data WHAM-RT and 64 found no discernible difference. This was neither a broad player survey nor a user study of the small-data models. The November preprint’s results remain author-reported, with independent replication unestablished. The study does not demonstrate that the same recipe transfers to another game or art style.

A technical demo is not a custom-training service

WHAM-RT is not appearing for the first time this November. Microsoft’s current project page links a Quake II technical demo in Copilot Labs, following an April 4, 2025 announcement. That was a separate experiment with different data and resolution, not the source of the 28-hour Bleeding Edge result.

The documented demo route does not establish a supported self-service training package, commercial-use terms or a price for this exact experiment. Playing a demonstration and training your own environment are different decisions. The same access-versus-capability distinction matters in Genie 4’s interactive-world preview.

Choose the experience before counting the hours

For a research prototype, start by specifying the map, character and interactions that must survive generation. Then examine whether recordings cover them, with synchronized controller actions and the location information needed for selection. Ordinary screen recordings alone are not what this experiment used.

For a team investigating its own environment, a useful pilot would compare equal-length random and curated sets while keeping character scope and inherited components fixed, then assess the desired interactions on held-out sessions. This is a suggested next step, not a tested recipe for another game. For artists whose actual task is object geometry rather than an interactive environment, custom TRELLIS.2 training with a 3D asset library addresses a different, more specific decision.

WHAM-RT makes a small, deliberately chosen training set worth investigating. It does not make 28 hours a universal target. Before recording another long session, ask which part of the intended experience is still missing from the footage.