Found a consistent 3D depth-collapse bug in Google’s new Nano Banana 2

{
    "prompt": "<!-- SC_OFF --><div class=\"md\"><p>image 1-<a href=\" image 2-<a href=\" <p>Tested the new automatic AI image generation within Google Search (powered by the 2.1 architecture) and mapped a consistent spatial breakdown when rendering complex optical paths. While the model handles standard 3D assets easily when viewed from the outside, it completely breaks down when forced into a centralized, eye-level mirror perspective—defaulting to flat 2D composition rules over 3D physical constraints. Here is the full breakdown and a proposed structural mitigation strategy. </p> <ol> <li><p>Executive Summary The model demonstrates an inability to correctly calculate three-dimensional physics and perspective when a centralized first-person camera view is requested, simulating the subject&;s exact eye-level perspective looking directly into a mirror in front of them. Instead of applying optical laws of occlusion (where the foreground elements of the vehicle structure should align and frame the perspective along the Z-axis), the model breaks 3D space by applying 2D compositional rules. This results in the driver&;s figure being incorrectly layered as a flat 2D sticker onto the external chassis of the reflected vehicle, collapsing the depth layers of the environment. </p></li> <li><p>Steps to Reproduce The following test scenario is used to intentionally trigger this logic bug:",
    "model": "Claude 3.7 / GPT-4o",
    "aspect_ratio": "1:1",
    "category": "Engineering & Strategy",
    "content_type": "text_prompt",
    "surface": "text_library",
    "quality_score": 87,
    "provenance": "COMMUNITY_PROMPT"
}