Slop TVNewsLatest
Explainers

StrideWorld finds 14 of 15 video world models deform the scene

The authors tested fifteen open interactive world models on 54,600 generated videos and report that almost all of them deform the scene as the camera moves.

Illustration: StrideWorld finds 14 of 15 video world models deform the scene
Illustration: SLOP TV News.

Key takeaways

  • StrideWorld, posted to arXiv on 18 November 2026, tested fifteen open interactive video world models on 54,600 generated videos from 140 start views.
  • The authors report that all models but one deform their scenes as the camera moves, with offsets up to 7.3 times those of a rigid rendering along the same camera path.
  • The benchmark fixes its reference geometry before the motion it judges, and the authors say a reconstruction-based reprojection error detects the deformation in only three of the fifteen models.
  • Fine-tuning two models on 112 rendered clips removed most of their mid-route stalls and more than doubled their Composition scores, the paper says.

A new benchmark finds that the interactive video world models you can drive today mostly deform the room as you move through it. StrideWorld, posted to arXiv on 18 November 2026, ran fifteen open world models through 54,600 generated videos and reports that all but one bend the geometry of the scene as the camera moves, with errors up to 7.3 times those of a rigid rendering along the same path. The measurement is the authors' own, on a benchmark they built, and no independent group has replicated it yet.

What an interactive video world model is, and who drives one

A normal video model turns a prompt into a clip you watch. An interactive video world model turns your commands into an exploration of a scene it generates as you go. The paper's own framing is that systems such as Genie 3 "build controllable generative worlds on demand, aiming to offer a new form of interactive entertainment."

The people reaching for one today are mostly researchers and the small group of creators prototyping for games, previsualization and virtual production. The pitch is a shot you can move inside, rather than a shot you re-render until the camera happens to land where you wanted. For the background on the category and on which models you can actually call today, SLOP TV keeps a running world-models guide.

The hard part is holding the room still

The paper states the two requirements plainly. For generated content to count as a world rather than a string of separately plausible frames, the scene "must first preserve its 3D structure, without deformation, as the viewpoint changes." And usable interaction needs every command to have "a consistent and predictable effect throughout the interaction."

If you make a shot, that second requirement is not abstract. When the walls drift a little as the camera dollies, two passes at the same room will not line up, and a cut between them jumps. A model that looks convincing frame by frame can still hand back a room whose floor plan quietly changed while you crossed it.

Why current benchmarks do not catch it

The older tests miss it because of where they point. The existing metrics rebuild the scene from the video's own frames, using 3D foundation models such as VGGT and Depth Anything 3, or they line the camera trajectory up to the command path with a free scale. Either way the reference is fitted to the video under test, and a reference fitted to the video can absorb a slow deformation, so the error vanishes into the yardstick. In the authors' words, "gradual scene deformation can be absorbed into the reference and go undetected."

The paper puts a number on the blind spot: "a reconstruction-based reprojection error detects scene deformation in only three of the fifteen models." On pacing it reports that when a rendered video's pace grows by half, a scale-fitted trajectory error flags 43% of the affected videos, while the new Pace test flags 94% at the same false-alarm rate.

What StrideWorld measures instead

Three tests, each judged against a reference the authors fix before the motion it measures.

Rigidity asks whether the scene keeps its 3D structure. The benchmark places 3D anchors at the ground-truth depth of the start view and checks whether the visible surfaces stay where they should as the camera moves.

Pace asks whether a held action keeps its effect. The benchmark measures the "stride" of each action, the camera motion of one command, from the model's own first four steps, then checks whether the camera holds that stride across sixteen steps.

Composition asks whether single actions add up. On routes that chain eight to fourteen steps, it checks whether the strides predict where the route ends. The routes come in pairs built to reach the same place, so the same moves in a different order should land together, and each route is compared with its own prediction and with its partner.

Taking the stride from the model's own first steps is what lets one benchmark compare models with different interfaces: it never needs a shared unit of motion. The authors validated the measurements on rendered videos with known camera motion and report that the scores agree with the judgments of ten human raters.

What it found across fifteen models

The benchmark covers 140 start views, 104 rendered views of 24 Unreal Engine 5 scenes and 36 photographs of 6 ETH3D scenes, with 3,640 videos per model and 54,600 in total.

On rigidity, every model deviates between 1.6 and 7.3 times more than a rigid rendering along the same camera path, so tracking noise alone does not explain the excess. One model, Lyra 2.0, is nearly clean once its depth is handled: the authors say its gap comes mostly from a fixed depth error, because it conditions each new view on an explicit 3D cache, and its ratio drops to 1 when depth is refit freely. The other fourteen stay at 1.2 times or above.

On pace, most models behave. Twelve of the fourteen ranked models land inside the steady band the paper defines, a median pace ratio between 0.8 and 1.25 when a single action is held.

Composition is where they break. Along composed routes several models stall partway, and most completed routes end away from the positions their own strides predicted. Only the two models that scale commands by the estimated scene depth bring most routes close to those positions.

Two findings are worth carrying out of the results. How well a model preserves the scene does not predict how well it composes actions; the paper reports Spearman correlations of -0.01 and -0.07 between the two measures. And fine-tuning two models on 112 rendered clips with exact camera paths removed most of their mid-route stalls, more than doubled their Composition scores, and did not lower their image-quality scores.

What it means if you are choosing a world model now

Keep the scope in view. This is the StrideWorld team measuring models on a benchmark the team built, and no third party has replicated it. The leaderboard is open and the code shows how to run the pipeline on a new model, but until others do, treat the ordering as one group's instrument rather than a settled ranking.

What it does change is what to check before committing a shot. If your shot is a single held move, the pace result is the reassuring one; most models keep a steady pace when an action is held. If your shot chains moves, composition is the test to run, because that is where the models in this study stalled and drifted. On a world model, "it looks fine" is a claim about frames. Whether the geometry holds is a separate question, and it is the one StrideWorld was built to ask.

Code, data and the live leaderboard are at github.com/chuangwang123/strideworld.

Sources

  1. arxiv.org - the paper's abstract, authors, submission date and headline results
  2. arxiv.org - the paper's body: the three tests, the per-model numbers, the fine-tuning study
  3. github.com - the maker's own code release and benchmark description
  4. huggingface.co - the maker's own data release, start views and request counts
  5. huggingface.co - Hugging Face Daily Papers listing and submitter note
  6. arxivsignals.io - aggregator page carrying the abstract and citation