Google Search AI image gen breaks on mirrors: driver pasted as a 2D sticker
Expensive-Ad-3569 · reddit · 2026-10-22
A Reddit user found a consistent spatial failure in Google Search's built-in AI image generation (reported as Nano Banana 2.1 architecture): with a standard external view, 3D scenes render fine, but a first-person prompt looking into a truck's mirror makes the model abandon 3D occlusion for 2D composition rules — cutting the female driver out like a sticker and pasting her onto the reflected truck's grille, collapsing depth layers. The post includes a 3-prompt reproduction recipe and proposes a mitigation: a dynamic camera look-at vector system that fixes the observer's 3D coordinates before rendering reflections.
More from Multimodal
- AI film studio Shortical plans 20 hours of AI-made vertical micro-dramas a month — lmoroney · 2026-10-22
- AI music video with authentic 90s rap-VHS vibes, made with MiniMax H3 and Soran't — pwillia7 · 2026-10-22
- Open-source ComfyUI workflow: 4K 600+ frame background replacement with MiniMax H3 tiled diffusion — Emotional_Example_12 · 2026-10-22
- 16 AI tips to make a full music video with just 4 tools — techhalla · 2026-10-22
- HeyGen's video and voice models land on fal: built-in sound, 39-language voice cloning — HeyGen · 2026-10-22
- Filmmaker uses Grok for movie previs: strong art direction, sync still lacking — TinfoilTricorn · 2026-10-22