Seedance 2.0 posted the highest result in a new benchmark for video world models, with an OPIS point estimate of 56.01. Across the eight evaluated systems, reported scores ranged from 48.65 to 56.01. The test asks a focused question: after receiving an opening image and an instruction, how well does a system’s generated video preserve the particular objects visible at the start? OPIS breaks that job into Presence, Identity and Structure, measuring whether an object is still there, is treated as the same instance, and remains consistent in its arrangement. The result is about observable output, not a direct readout of a system’s internal memory.
A test anchored to the opening frame
Each case begins with an initial image, a generation instruction and a fixed evaluator-side reference. The system generates a video sequence from the image and instruction, while the reference records the objects visible in the input and auxiliary geometric evidence from that same image. The reference is built once before generation and is never updated with generated content. That makes the comparison independent of a target continuation and means the evaluator does not need access to the model’s internal memory.
Following an object through a generated sequence is not treated as a simple visual lookup. OPIS uses partial, one-to-one association based on appearance embeddings, attributes, category information and geometric evidence. It can leave an observation unmatched and uses explicit visibility reasoning when an object cannot be assessed.
A varied test set
The benchmark covers 500 cases across three scene domains and 10 subcategories, with 12,672 annotated object instances. The inventory includes 8,432 rigid instances, 2,046 articulated instances and 2,194 deformable instances. Of the cases, 200 are real-world, 150 are embodied or robotic and 150 are drawn from game worlds. Across the tracks, 6,128 instances use the static track and 6,544 use the dynamic track.
The construction uses vision-language models to help select and review cases, SAM 3.1 to create masks around individual objects, DINOv3 to encode their appearance, and MoGe-2 to provide geometric scaffolds.
How the systems were compared
Tests targeted video sequences lasting approximately 10 seconds and sampled every eighth frame. Eight systems were evaluated across two interfaces. The analysis calculated 95% confidence intervals by resampling the cases 2,000 times, using a stratified case-bootstrap procedure.
Seedance 2.0’s 56.01 was a point estimate, not a fixed universal score: the supplied analysis gives it a 95% confidence interval from 52.13 to 59.69. H3-Max-Turbo was reported at 55.57. Echo-WM-Flash led the camera-conditioned group at 51.42. The figures describe performance on this benchmark’s object-preservation task, not every capability a video model might have.
What the score can and cannot say
That boundary is central. OPIS measures preservation of observable input objects; it does not identify an internal memory mechanism, establish physical causality or demonstrate complete world-model competence. It also does not show that a larger object inventory causes a decline in per-instance fidelity. The design can reveal patterns in output, but it cannot explain the hidden process that produced them.
The benchmark’s conclusions are tied to the evidence its matching procedure can assess. Association is partial and one-to-one, combining appearance, attribute, category and geometric evidence, with null assignments and visibility reasoning for unmatched or unassessable observations. That makes the rules explicit and keeps the result tied to this protocol.
The work is an arXiv preprint dated 28 Sep 2026. On the evidence supplied, its clearest contribution is a shared, input-grounded way to compare whether video systems preserve specific objects from an opening scene. The scores offer a concrete baseline for that question, while leaving broader claims about memory, causality and deployment performance open.
Paper data and sources
Original title: OPIS: An Input-Grounded Benchmark for Multi-Object Memory in Video World Models
Authors: Hao Wang, Tao Yu, Liuzhou Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text