WROP fine-tunes a 16B video generation model on 1.5M Blender-rendered clips designed around Object Permanence and Object Solidity, producing a continuation model that ranks first among video-continuation systems in a blind human study, competitive with frontier commercial video generators despite outputting at 320×192.
Modern video generators like Sora or Veo make photorealistic clips and are increasingly pitched as world models: systems that can simulate what happens next in a physical scene. But they routinely break basic physics. Objects vanish behind occluders and reappear in wrong places. Balls roll through solid walls. A ball sits mid-air after its support slides away.
Human infants handle these situations by 3-6 months of age. Developmental psychology calls these primitives Core Knowledge: the built-in scaffolding on which all later physical reasoning rests. If a video model can’t preserve an object through occlusion or respect that solids don’t interpenetrate, then any downstream use (planning, simulation, robotics rollouts) inherits those errors.
The practical baseline the paper works against: existing video-reasoning benchmarks are mostly 2D, use image-to-video rather than video-to-video setups, and score outputs with VLM judges. The authors argue VLM judges are unreliable for physics questions because VLMs themselves fail at core-knowledge tasks. So WROP is built to be judged by humans against hand-authored ground truth.
The contribution has two halves: a dataset/benchmark, and a demonstration that fine-tuning on it moves the needle.
The dataset. 150 hand-written Blender scene generators, organized into six task families: three for object permanence (things behind occluders, hidden scenes, containers that move) and three for solidity (barriers, support removal, collisions). Each generator is a parameterized 3D scene where motion is authored by Keyframe Animation rather than by a physics engine. Structural parameters (object count, trajectory, aperture size) vary the physical challenge; surface parameters (color, lighting, camera) randomize appearance. Each generator produces 10,000 samples, for a 1.5M-sample training corpus and a fixed 300-question exam (2 per generator).
The evaluation contract. Every clip is 120 frames. The split is placed right before the key physical event. Models get the first 60 frames plus a text prompt and must generate the next 60 frames, which contain the event and its consequence. This is a strict V2V setup.
The model. PWM-WROP is fine-tuned from Cosmos3-Nano (a 16B open-weight video model from NVIDIA) for one epoch on the corpus. Architecture and tokenizer are unchanged; only the training signal changes. The model conditions on 57 frames and predicts 60 at 320×192. No task labels are provided during training, only the video and the natural-language prompt.
# WROP training/eval contract per sample
clip = render_blender_generator(gen_id, seed) # 120 frames
input_half, target_half = split_at_event(clip, frame=60)
prompt = generator.prompt_text
# training
loss = video_model.train_step(
condition=input_half, target=target_half, text=prompt)
# evaluation (same contract)
pred = video_model.generate(condition=input_half, text=prompt)
human_rate(pred, other_model_pred) # blind pairwise
The scoring. 14 models are compared in three interface classes: true continuation (extends the input), reference-to-video (uses the input as a visual reference but generates on its own timeline), and edit/transfer (repaints the input span frame-by-frame). 20 crowd raters do blind pairwise comparisons, fit with a Bradley-Terry loss model and reported on an Elo scale. Full-reference pixel/perceptual metrics (LPIPS, MS-SSIM, PSNR) serve as secondary target-fit checks.
Overall ranking. Two reference-to-video systems (Wan 3.0 Prime and MiniMax H3) tie at the top with Elo 1723.6. PWM-WROP is third at 1679.5, and first among true-continuation models by a wide margin (next best, Grok Imagine’s video-extend, sits at 1457). The other continuation baselines (LTX-2.3 Extend, MAGI-1 24B) rank near the bottom.
Interface class matters more than scale. Reference-to-video models get freedom to regenerate the whole scene, which seems to help solidity tasks (collisions, drops) where they can plan a physically clean outcome. Edit/transfer models cluster in the middle regardless of provider, likely because frame-wise repainting struggles to depict events that happen after the input ends.
PWM-WROP’s profile is uneven across the six families. It ranks 1st on object-static occlusion, 2nd on support-removal drops, but 8th of 14 on collision. The authors interpret this as evidence that occlusion tracking and contact dynamics may need different internal representations, and that the current fine-tune helps the former more than the latter.
Full-reference metrics. At matched 320×192, PWM-WROP is best on LPIPS (0.081 vs next-best 0.105) and MS-SSIM (0.921 vs 0.877). The authors are careful: these metrics measure similarity to a reference clip, not physics correctness. An edit model that faithfully repaints the pre-event scene can score well without ever generating the required event.
Qualitative failure modes. Two recurring patterns across baselines: representation dropout (a ball fails to re-emerge in its correct lane, or the wrong cup is lifted after a rotation), and causal decoupling (a ball stays suspended after its support slides out, or ball counts change through a collision). PWM-WROP handles most permanence cases cleanly but still misses on some collision scenarios.
•
If you’re evaluating whether a video model can serve as a physics simulator or planner, the WROP exam is a targeted probe. It isolates permanence and solidity from general visual quality, and it’s scored by humans rather than a VLM judge. Note that the exam is only 300 items and the results depend on human pairwise studies you’d need to run or trust.
•
If you’re training a video model and care about physical consistency, the paper is a working example that concept-specific synthetic data can shift human-rated behavior, at least for occlusion tasks. Worth testing: whether the same recipe on your base model helps, and whether specifically augmenting collision data closes PWM-WROP’s own gap there. The authors do not isolate what fraction of the gain comes from the data versus other training choices, so treat the effect as suggestive.
•
If you’re picking a video model for a video-to-video product task today, the interface class finding is directly actionable: for tasks where preserving the input scene matters, true-continuation models are stronger; for tasks where the model needs to produce a fresh physically-plausible event, reference-to-video systems currently win but may drift from the input scene’s identities.
•
Artifacts released: data, exam, model answers, weights, and PWM, the native-PyTorch training stack for AWS AWS Trainium2. See the project site.
•
The top result isn’t a clean win. PWM-WROP is third overall; the two reference-to-video models above it operate under a different contract (they regenerate rather than extend) and at higher resolution. The paper’s headline is “first among true-continuation models,” which is a category ranking, not an absolute one.
•
Confounded comparison. PWM-WROP differs from other continuation baselines in both training data and architecture. The gain cannot be attributed to WROP data alone. There’s no ablation of a same-base-model checkpoint without WROP fine-tuning reported in the main text.
•
Resolution mismatch. PWM-WROP outputs at 320×192 while competitors output at 720p-1080p. Human raters saw all clips normalized to 720p, so PWM-WROP’s outputs were upsampled. Automatic metrics computed at 720p penalize it accordingly.
•
Synthetic-only evaluation. Every scene is Blender-rendered with hand-authored keyframe motion, not physics-simulator output. Transfer to natural video is untested here.
•
Small per-family sample. Human judgments per task family are 36-96 games; bootstrap intervals are wide and adjacent ranks are often not statistically separable. The family-level story (occlusion vs contact) is directional, not conclusive.
•
Judgment criteria are bundled. Raters weighed prompt alignment, motion plausibility, and permanence jointly in a single A/B choice, so the Elo doesn’t cleanly isolate physics reasoning from general quality.