1Why a commercial model sits a physics exam
Viewers notice physics before they notice anything else.
A coffee pour that thickens halfway down, a bottle that drifts on a still table, a ball that rolls uphill: these break an ad in a fraction of a second. Viewers rarely say "the physics is wrong". They just stop trusting the product on screen.
Most video models are judged on how good a frame looks. Our customers need something stricter. When a brand hands us a product shot, the first frame belongs to them. Whatever happens next has to follow from it, with the right weight, timing and stillness.
Physics-IQ measures exactly that. The model receives a real frame from a filmed experiment and has to continue it for five seconds. The benchmark then compares the result with what the camera actually recorded.
2What Physics-IQ Verified measures
Physics-IQ was introduced by researchers at Google DeepMind. It films 66 real experiments covering solid mechanics, fluid dynamics, optics, thermodynamics and magnetism, each from three camera angles, for 198 test clips. Physics-IQ Verified is the maintained edition, with refined ground truth and an official evaluator.
The scoring cares less about pretty frames and more about where and when things move.
| Measure | Question it asks |
|---|---|
| Spatial IoU | Does motion happen in the right place? |
| Spatiotemporal IoU | In the right place, at the right moment? |
| Weighted spatial IoU | How much motion, and where? |
| MSE | How close are the pixels to reality? |
A perfect, physically realistic model would score 100. The four measures share one consequence that shaped all of our work:
Flicker, a shimmering background, or a hand that wanders into frame all count against you, even when a person watching would hardly notice.
3Four problems we found, and what we did about them
Before touching the benchmark we built a private test suite of our own scenes: eight where nothing should move and eight with a single physical event, such as a rolling ball, a falling domino or a swinging pendulum. It scores motion the same way the benchmark does, so we could check every change without ever looking at benchmark answers. These are the problems it surfaced.
Voice-over nobody asked for
26%→12%clips with unrequested voice-over
- Observed
- Modern video models generate sound with the picture. Asked for a silent product shot, the underlying foundation model invented narration or a voice-over in 26% of clips across our internal set of more than 500 commercial briefs. For an ad, that clip is unusable.
- Fix
- We trained a Fabeo adaptation on our own in-house commercial footage, teaching the model that silence in the brief means silence on the track.
- Result
- Unrequested voice-over fell to about 12% of clips, cutting the problem by more than half. Physics-IQ is video-only, so this gain never shows up in the score. It shows up in what customers can ship.
Phantom motion
33.79→+7%Physics-IQ, first configuration → after fix
- Observed
- Our first production configuration scored 33.79. Looking at the motion masks, much of what it lost came from movement that should not exist: faint frame-to-frame shimmer in still backgrounds. In one test scene, a speed-optimised variant marked 89% of the frame as moving, although it looked static to the eye.
- Fix
- We traced the shimmer to speed shortcuts in the rendering path and rebuilt it so that a still scene stays still.
- Result
- The score rose by about 7%, and product shots got cleaner backgrounds. For a brand, that is the difference between a hero shot and a clip that looks slightly unsettled.
Captions burned into the picture
burned-in→cleansubtitles painted into the frame
- Observed
- When spoken lines travelled in the same instruction as the visual description, the model sometimes painted them into the frame as subtitles.
- Fix
- Fabeo's pipeline keeps spoken lines out of the visual instructions, so narration stays in the audio and the picture stays clean.
- Result
- Clean frames that can be captioned, localised and re-voiced afterwards without re-rendering.
The “something always happens” habit
stray hands→base levelunintended motion in still scenes
- Observed
- Commercial footage is full of action, and a model trained on it starts to expect action everywhere. Pushed too hard, our adaptation completed events better but also added stray hands and props to scenes that should stay still.
- Fix
- We calibrated how strongly the adaptation is applied, using the still-scene half of our private suite as the referee.
- Result
- At the calibrated setting used for this benchmark, unintended motion fell back to the level of the unadapted model.
4Where FabMax lands
FabMax scored 36.05 on Physics-IQ Verified. Among creative video generators run the standard way (benchmark prompts, one take per scenario), it ranks fourth of sixteen. It is ahead of Gemini Omni 1.1 Flash, Grok Imagine, Hunyuan Video 1.5, Wan 2.2, Veo 3.1 Lite and Fast, and Sora 2.
FabMax is also one of only two entries in that group run on the benchmark's original prompts. Most entries use best-practice prompt templates written for each model, so our number reflects the model with no prompt help.
FabMaxPublished entry± spread across runs
Scores as published on the Physics-IQ Verified leaderboard (benchmark repository, 8 Oct 2026). FabMax is shown at its verified score from a single run; its leaderboard listing is pending. The full board also includes entries that choose the best of several generations per scenario, use custom or LLM-written prompts, or are physics-specialised world models. Video-to-video entries receive more input frames and are not comparable, so they are left out.
The four sub-scores tell the same story as our field notes. FabMax is strongest at putting motion in the right place and keeping pixels close to reality. Timing, the spatiotemporal measure, is the hardest of the four, and it is where our next round of work is aimed.
- Physics-IQOverall verified score36.05
- Spatial IoUMotion in the right place48.25
- MSE scorePixel fidelity36.02
- Weighted spatialRight amount of motion34.38
- SpatiotemporalRight place, right moment25.58next focus
5A score you can take at face value
Benchmarks reward shortcuts, so we set our own rules before rendering a single test clip.
- No tuning on benchmark data. The adaptation was trained only on Fabeo's commercial footage, and every decision was made on our private scene suite.
- One take per scenario. No generating eight videos and keeping the best.
- The benchmark's own prompts, verbatim. No rewriting and no model-specific templates.
- Official evaluator. All 198 clips scored by the benchmark's verified scorer and checked by its submission validator.
- Paired statistics. Every comparison is scenario by scenario, with bootstrap confidence intervals, so noise does not get reported as progress.
We also tested ideas on our private suite that did not make the cut. Rewriting each scene's description with a language model made the prose more vivid and the physics worse: rewrites added camera cuts, extra dominoes and dramatic lighting that the real experiment never had. Grander prompt templates had the same effect, adding hands that set up the scene. The lesson carries straight into our product.
A good prompt for physical realism describes what happens and stops there.
6A video model tuned for selling, not just for showing
General-purpose generators hand you a clip. Fabeo hands you a finished commercial or explainer, and FabMax is the engine underneath. What we learned on Physics-IQ is built into how it works.
- Your first frame stays yours
- Every scene starts from an approved keyframe of your product, location or character. FabMax continues it with believable physics instead of reinventing it.
- Silence means silence
- FabMax is trained to keep a silent shot silent, cutting unrequested voice-over by more than half. When you want narration, it comes from your approved script.
- Still means still
- We measure background shimmer, stray hands and drifting props on scenes where nothing should move. That is the same discipline that earned FabMax its Physics-IQ points, applied to your product shots.
- From brief to finished film
- Script, scene plan, keyframes, motion, voice and edit in one workflow, with approval points where you want them. You direct the ad and we handle the rendering.
Physics-IQ is the first of several public benchmarks we are running on FabMax. We will publish the results, and the problems they uncover, as we go.