单图生成 3D 实测:背面可信,但浮雕铭文凭空消失了
sweetmiso · reddit · 2026-10-04
A user tested Seed3D 2.0 (via Atlas Cloud) on a single photo of a fire hydrant, exporting the GLB into Blender. Findings:
- The back was surprisingly plausible — overall shape held together from nearly every angle despite only one input photo
- Textures kept, geometry wasn't: rust shows in the texture, but none of it changes the mesh geometry
- Raised cast lettering is simply gone — not mangled, just missing
The takeaway: good enough as a background prop, but a convincing turntable can hide how much object-specific detail was silently stripped. When judging AI 3D demos, do you weigh motion believability or preservation of the original?
「多模态」频道最新
- 16GB 显存想训 LoRA 统一产品摄影风格，该自训还是用 ControlNet？ — Qwesbrz · 2026-10-05
- Seedance 2.5 + GPT Image 2.5 演示生成憨豆先生现身 F1 赛场 — TinfoilTricorn · 2026-10-05
- Claude Code 全自动执导:AI 独立做出三分之三狗侦探短片 — Cheap_Credit_3957 · 2026-10-05
- 为什么 AI 电影人总拍出诡异画面还自认大片？缺的是懂人 — taherdhanera · 2026-10-05
- UGC 广告一次过：Opus 5.5 智能体剪辑直出成片 — alexgoughcooper · 2026-10-05
- 博主预言：十年内电影院将消失，AI 让人人自制影片在家看 — taherdhanera · 2026-10-05