Physics Reasoning
Does an AI image model apply physics it was never asked about?
Research package · published 2026-07-13
The finding
Google Nano Banana 2 was the only tested model to apply the implied optics reliably, passing five of six runs (83.3%); every other model passed at most one of six.
Method: Two implicit-optics tests used neutral prompts that never named refraction, reflection or the expected effect. Seven models generated three runs per test with no retries. Each run was hand-graded pass/fail against a disclosed physical criterion; all 42 source images and exact prompts are published.
“In PromptFrenzy's 42-run implicit-optics benchmark, Nano Banana 2 passed five of six runs; no other model passed more than one.”
Source: PromptFrenzy Physics Reasoning benchmark, published 13 July 2026. CC BY 4.0; link to the canonical research page when citing.
Every prompt here describes a scene neutrally and never names the effect — no “refraction”, no “reflection”, no “reverse”. A model only scores if it renders the optics the scene implies on its own. It’s a different question from prompt-adherence: gpt-image-2, the top model on our adherence benchmark, scores 0 here. Binary scoring, 3 runs per model, no retries — and every run is published raw below so you can re-grade it.
The board
| # | Model | Arrow Refraction | Two-Mirror | Overall |
|---|---|---|---|---|
| 1 | Google Nano Banana 2 | 67 | 100 | 83 |
| 2 | Meta Muse Image | 0 | 33 | 17 |
| 3 | Nano Banana 2 Lite | 33 | 0 | 17 |
| 4 | FLUX Kontext Pro | 0 | 0 | 0 |
| 5 | OpenAI gpt-image-2 | 0 | 0 | 0 |
| 6 | Qwen Image 2.0 Pro | 0 | 0 | 0 |
| 7 | Seedream 5 Lite | 0 | 0 | 0 |
Overall is the mean of the tests a model entered. The pencil-in-water refraction test is intentionally left off the board — its correct appearance depends too much on camera angle, lighting and glass curvature to grade reliably. See the full model leaderboards.
Two Arrows, One Glass of Water
A cylinder of water horizontally reverses the arrow seen through it — so the submerged arrow should point the opposite way to the one above the waterline. The prompt says both point left and never mentions refraction.
| Model | Run 1 | Run 2 | Run 3 | Score | What happened |
|---|---|---|---|---|---|
| Google Nano Banana 2 | ✓ | ✗ | ✓ | 67 | Reverses the submerged arrow in 2 of 3 runs; run 2 fails — reversed head but the shaft seen through the glass stays unreversed. |
| Nano Banana 2 Lite | ✗ | ✓ | ✗ | 33 | One clean reversal (run 2); otherwise leaves the submerged arrow pointing left. |
| FLUX Kontext Pro | ✗ | ✗ | ✗ | 0 | Doesn't stage the test — floats the arrows above the glass, so nothing is seen through the water. |
| Meta Muse Image | ✗ | ✗ | ✗ | 0 | Clean scene, but never reverses the submerged arrow. |
| OpenAI gpt-image-2 | ✗ | ✗ | ✗ | 0 | Renders the scene perfectly but never applies the refraction — both arrows stay left. |
| Qwen Image 2.0 Pro | ✗ | ✗ | ✗ | 0 | Draws arrows wider than the glass or double-headed; never a clean, complete reversal. |
| Seedream 5 Lite | ✗ | ✗ | ✗ | 0 | No reversal; sometimes adds a spurious third arrow. |
Show all 21 runs (raw, unedited)
Prompt (identical for every model, never names the effect): A photorealistic photo taken at eye level, the camera level with the middle of a drinking glass, looking straight on at a horizontal line of sight. A white sheet of paper stands upright on a wooden table, propped vertically against a stack of books so it faces the camera directly. On the paper are exactly two small thick black arrows drawn one above the other, both pointing to the left; each arrow is drawn short and compact, clearly narrower than the width of the glass, so the whole arrow including its arrowhead fits comfortably within the glass. On the table, a few inches in front of the paper, stands a large clear wide cylindrical glass filled with water; it is placed directly in front of the lower arrow so that the entire lower arrow, shaft and arrowhead together, sits behind the body of water and is seen through the glass, while the upper arrow is higher up, above the rim of the glass in open air. A dark blue tablecloth, a marker lying on the table, soft even studio lighting.
Google Nano Banana 2
Nano Banana 2 Lite
FLUX Kontext Pro
Meta Muse Image
OpenAI gpt-image-2
Qwen Image 2.0 Pro
Seedream 5 Lite
Two-Mirror Corner — Chrome Sphere
A mirror facing a person must show their FRONT — never their back. In a right-angle corner the reflections should be coherent with that geometry. The prompt only describes a person seen from behind; it never says what the mirrors should show.
| Model | Run 1 | Run 2 | Run 3 | Score | What happened |
|---|---|---|---|---|---|
| Google Nano Banana 2 | ✓ | ✓ | ✓ | 100 | Correctly reflects the person's front in each mirror, all three runs. |
| Meta Muse Image | ✓ | ✗ | ✗ | 33 | One run reflects the front correctly; the other two show the back / incoherent reflections. |
| FLUX Kontext Pro | ✗ | ✗ | ✗ | 0 | Weak or missing mirror reflection. |
| Nano Banana 2 Lite | ✗ | ✗ | ✗ | 0 | Incoherent reflections that don't fit a right-angle corner. |
| OpenAI gpt-image-2 | ✗ | ✗ | ✗ | 0 | Mirror shows the person's back again — impossible for a mirror facing them. |
| Qwen Image 2.0 Pro | ✗ | ✗ | ✗ | 0 | Chaotic reflections; not staged from behind. |
| Seedream 5 Lite | ✗ | ✗ | ✗ | 0 | Mirror shows the back, not the front. |
Show all 21 runs (raw, unedited)
Prompt (identical for every model, never names the effect): Studio photograph seen from behind a person standing in the corner where two large flat mirrors meet at a right angle. They raise a polished chrome sphere high in their right hand; their left arm hangs at their side. A single mustard-yellow patch is sewn on the left shoulder of their grey jacket. Soft daylight, plain pale room, eye-level camera.
Google Nano Banana 2
Meta Muse Image
FLUX Kontext Pro
Nano Banana 2 Lite
OpenAI gpt-image-2
Qwen Image 2.0 Pro
Seedream 5 Lite
Free to reuse with attribution to the PromptFrenzy Physics Reasoning benchmark (CC BY 4.0).