Even top 3D-gen models keep making the same basic mistakes, researcher notes
teortaxesTex · x · 2026-11-09
A zero-shot Gemini 4 Krypton prompt produced an interactive 3D historical sailing ship in 22m20s with striking realism. But teortaxesTex highlights recurring basic errors across even the strongest models: Z-fighting (likely because models don't see video), backface culling / single-faced geometry, and coordinate axis mismatches between world, camera, and controls.
More from Multimodal
- Low-VRAM showdown: Qedit 2511 skips prompt enhancer but Qwen 2.1 wins on facial drift — Theonewhoknocks_001 · 2026-11-09
- Midjourney sref 3911489532 yields moody blue-orange clock-tower film aesthetics — umesh_ai · 2026-11-09
- WAN 3.0 flaw: can't pin one reference image across video when using multiple refs — Disastrous_Ad_3010 · 2026-11-09
- First AI opera adaptation of Frankenstein built on a coding-driven generative pipeline — hq4ai · 2026-11-09
- Notion-style minimalist B&W illustration prompt template, swap the brackets and go — azed_ai · 2026-11-09
- ByteDance launches Seedream 5.5 image model with 100-round editing and native RGBA — kimmonismus · 2026-11-09