Seedance 2.5 leads the published World Models' Last Exam in Physics evaluation, but that result needs context. A convincing ball, reflection, or splash can still behave incorrectly. For a creator, the useful question is whether the event in a particular shot makes sense, rather than whether the model has a strong overall ranking.
This article explains the reported scores and proposes a small editorial review exercise. The numbers come from the researchers; Seeora has not reproduced their evaluation or tested the exercise as a model comparison.
Generated concept illustration, not a Seedance output, benchmark screenshot, or measured trajectory.
What the benchmark result actually says
The October 6, 2026 preprint evaluates eight video models using 40 controlled tasks and 1,280 videos. Each task provides an initial image and prompt. The tests examine measurable physical relationships across nine categories, rather than simply asking whether the output looks realistic.
Its composite score combines continuity and observability with physical measurements. A video needs to pass the consistency screen before its physical score contributes. Passing that screen establishes neither correct physics nor guaranteed recovery of every measurement. The paper also notes that visibility and extraction reliability constrain evaluation.
The authors' published results table distinguishes these figures:
| Reported measure | Score | How to read it |
|---|---|---|
| Seedance 2.5 overall composite | 57.76 / 100 | Highest overall in this tested group |
| MiniMax H3 overall composite | 54.89 / 100 | Another result from the same evaluation |
| Seedance 2.5 free-fall task | 29.50 / 100 | A specific task, not its overall score |
| Free-fall cross-model mean | 26.61 / 100 | An average across models, not Seedance's individual result |
These are benchmark scores, not percentages of successful filmmaking attempts. They do not establish which model will produce your preferred performance, sound, style, or usable clip at the lowest cost.
Decide what your shot must communicate
Before generating, write down the event that the audience must understand. For a dropped object, the important story beat might be the moment a hand releases it and the reaction after impact. For a drink commercial, it might be the liquid entering a glass without the container changing shape. A fantasy transformation may intentionally use impossible motion.
Make that intention explicit. Otherwise, you can spend retries repairing a behavior that the scene never needed, or approve attractive footage that contradicts the central action. A stylized shot and a realistic demonstration deserve different acceptance criteria.
For an ordinary object drop, an original review brief could ask for a visible release, an unobstructed descent, and a clear landing. Keep the camera still for this first exercise and use a contrasting object and background. This is a proposed way to make editorial inspection easier, not a benchmark prompt or a proven method of improving model physics.
Review continuity before judging behavior
Watch once at normal speed to assess the scene's meaning. Then inspect the release, travel, and contact separately. Does the same object remain present? Does its size change unexpectedly? Does it disappear behind an effect at the exact moment you need to judge?
If you cannot identify the object through the event, record the interval as unclear. Do not label the hidden action correct because the surrounding frames look polished. Also distinguish an obscured event from an observed mistake: those lead to different revisions.
Compare the unedited source with the delivered version. A cut, speed ramp, or slow-motion section changes the timing the viewer sees. Keep those edits in your notes so you can tell whether an unusual pause came from generation or an intentional treatment. For external effects, the Seedance postproduction VFX article offers a separate planning brief.
Keep an actionable shot review log
Use short notes tied to visible events instead of a single realism rating. A useful log can contain the source filename, intended action, questionable interval, observation, and next change.
| Observation in your own clip | Next experiment |
|---|---|
| The object changes identity before contact | Simplify the shot and review another take |
| Contact is hidden by a cut | Request or select a take that keeps it visible |
| Motion seems floaty in the original | Inspect the interval and compare another unretimed take |
| A dramatic effect hides the landing | Reduce the effect and review the clean footage |
Treat these as editorial options, not established fixes. Change one element at a time so the next attempt answers a specific question. Keep rejected takes and reasons when comparing settings; selecting only the best output makes a workflow appear more reliable than your actual experience.
When you need a formal physics evaluation
The researchers' repository includes task inputs, generation support, and evaluation tooling. Its existing-video workflow expects the task's supplied first frame and prompt, with task identifiers in video filenames. Scores and evidence are saved separately. This is a route for a technical evaluation, not a scoring service built into Seeora.
For a creative project, start smaller: select one shot, define the event, inspect the source, and document the revision you need. Use the Seedance prompt library for scene planning. A strong model-level result is useful context; your acceptance decision still needs evidence from the clip you intend to publish.