| Generator | Prompt | Labels | Medical | Concept | Polish | Total |
|---|---|---|---|---|---|---|
| Ghost (Human Baseline) | 5 | 5 | 5 | 5 | 5 | 25 |
| AIMAGE (Ghost fine-tune) | 4 | 5 | 4 | 5 | 4 | 22 |
| Sora | 4 | 5 | 4 | 4 | 4 | 21 |
| Flux 2 | 3 | 5 | 1 | 2 | 4 | 15 |
| Midjourney | 2 | 5 | 1 | 1 | 1 | 10 |
| Nano Banana | 2 | 2 | 2 | 2 | 1 | 9 |
| Grok | 2 | 5 | 0 | 0 | 1 | 8 |
This is the only entry that completely matches all the details of the prompt. It perfectly matches the HLT device because it's based on the same CAD model used to manufacture the device. The valve proportions, leaflet shape, and overall construction look coherent and "designed," not improvised. This is what "production-correct" looks like.
Closest to a "trained junior Ghost artist." Still drifts in subtle engineering details, but it stays in the correct object category and usually maintains plausible proportions. This is likely because it was the only AI that has actually been trained on the exact same and similar transcatheter aortic valve prosthesis that Ghost used to render their image.
Strong renderer and decent coherence, but can still average details into "generic medical device." Often needs tighter constraints to stop it from inventing convenience geometry. Likely, Sora was trained on similar devices, perhaps even Ghost's own images.
Main failure: Overall device shape and proportions are off enough to break physical plausibility. If the implant's profile would not seat correctly at the annulus or is dimensionally nonsensical, it gets a 1 in Medical and Device Accuracy, even if it looks pretty. This is the classic "photoreal but wrong object" trap.
Still behaves like "stylized product concept," not "specific implant." It can be visually pleasing while being structurally useless.
Rendered a fan instead of a valve. Immediate disqualification on correctness. This is what happens when an AI confidently generates the wrong object entirely.