1/ AI 3D demos on X have been wild lately, especially the crazy mechanical assemblies and exploded views.
Gave it a shot myself with MCP + Blender & Three.js, using a Polaroid Now+ as reference.
1
58
2/ ...and here's where I landed 😂
Total miss. Topology is a mesh mess, proportions are way off, and it feels more like a plastic toy block than a camera. (Image-to-3D got me closer, but that's a pretty different pipeline.)
1
37
3/ Thought I just lacked the right Skills/prompts. But having done 3D for years, the closer I look at those viral demos, the more it seems a cool procedural hero asset is a very different thing from nailing a real object from photos.
Anyone actually got reference-accurate results out of MCP + Blender or Three.js? Would love to see your workflows.
7
2/ ...and here's where I landed 😂
Total miss. Topology is a mesh mess, proportions are way off, and it feels more like a plastic toy block than a camera. (Image-to-3D got me closer, but that's a pretty different pipeline.)
6
最近 X 上各种 AI 3D 演示看着是真吓人，机械章鱼、机械蜜蜂一个比一个复杂，标题动不动就“天塌了”。
但我自己试了下 MCP + Blender，还有 Three.js，就一个看着方方正正的宝丽来相机，如图所示。折腾半天也没做出个满意的。
反倒是图生 3D 还靠谱一点。当然，图生 3D 和前面说的也不是一条路子，不能直接拿来比。
所以我在想，是不是让 AI “自由”发挥，反而比照着参考图老老实实还原一个简单东西更容易？或者说那些个通过大量阵列堆叠相似物件相对“容易”一些。
毕竟对照一个参考把形体、比例、结构都做对，感觉完全是两码事。
倒不是说那些演示没技术含量。
当然，也可能是我自己没玩明白，测试的东西也不多。😂
有没有折腾过 MCP、Three.js 。尤其是拿参考图做精确还原的，效果特别好的案例？挺想看看以及学习都是怎么做的。
2
69
