你们敢信么？这是 Kimi 3.1 Pro 帮我做的：
相当的牛掰！
它的第一个版本是这样：
刚开始是有点抽象的。
这两个版本差距巨大！
第一个像是来搞笑的，第二个已经是专业作品了。
从下面的效果变成上面的效果，我就说了“一句话”，真的是一句话，几十个字！
这是不是很哇塞了？
其实，这个世界没有魔法，一切有迹可循！
我给大家解密一下！
我昨天测试甲维斯车库的时候，对 K3.1 Pro 制作的 Mark III 不是很满意，实在是太抽象了。
我不死心，我要再挖掘一下 K3.1 Pro 的潜力！
然后，我就单独找了一张图片，单独开了一个对话。
希望它能帮我根据图片完成一个不错的建模：
我的输入是这样的：
经过一段时间，第一个版本做好了。
比甲维斯的车库里面好一些，但是并没有好太多。
我对这个建模质量是非常不满意的，这水平对不起它的宣传啊！
然后，我进一步提出了需求：
你可以再好好对比一下，想一个办法，让你的结果和原图更加接近吗？
其实，这个时候我内心的潜台词是“你这什么鬼东西，能不能自己看一眼先”。
然后，没多久，我就得到了这个结果：
我当场就被惊呆了！
这质量也太高了吧，绝对吊打世界上任何一个模型。
简直就是魔法一般，惊叹一会儿，我的直觉告诉我，这中间肯定有“猫腻”！
然后，我就仔细看了一下它的执行过程。
其中有一句重点信息是这样：
找到了更合适的办法：一个可直接获取的高精度 GLB 战甲网格，原本就包含完整装甲分区和贴图。接下来会用它替换手工拼接模型，再按参考图调整比例、材质与灯光。
哈哈哈，牛逼。
手搓不行，直接从网上下载了一个高质量的模型，然后在页面中呈现出来。
刚开始它找到了一个纳米战甲，发现和参考图不太像。
然后，重新去找了一个 Mark III 模型！
就靠这一个小技巧，让它快速做出来大师级的建模。
不得不说，K3.1 Pro 还真的是个“机灵鬼”。
我之所以做这个测试，其实是因为听到一个传言。
据说 K3.1 Pro 建模质量高，是因为会偷偷调用 Trip3D 来生成模型。
我就想试试，能否复现这个场景。
如果能复现，能自动去白嫖 Trip3D 的模型，也是很不错的。
没想到的是，它并没有用 Trip3D，而是直接从网上找了 OBJ 和 GLB 文件。
这个事情，多少有点投机取巧的嫌疑。
因为调用别人创建的模型，显然无法体验自己的建模能力。如果你对外宣传建模能力很强，是通过这种方式实现的，总是有点说不过去。
但是从另外一个角度，从完成任务的角度来说。
那算是完成的很好了，高质量且高效率。
昨天我还说它搜索和抓取图片的能力不行，今天就发现它搜索和抓取模型的能力很强。
如果我自己去找模型，验证，估计半天都搞不定。
它是轻轻松松帮我找到了，验证好了，然后载入模型在网页中程序出来了，载入之后还能修改！
这是不是也算一种能力呢？
然后，我还让他做了一个拆解模式：
这样一个建模完美，展示酷炫的网页就轻松搞定了！
我昨天也说了，如果我不是为了测试，为了帮大家找出问题，我也可以做出很哇塞的例子！
是不是？立马就搞出来了！
我不说原理的话，大家是不是觉得这模型能力碉堡了。
以后你们要建模的话，也可以试试，也许它能帮你找到很多高质量的模型，完全不需要花时间是自己创建。
这也算是我的一个小发现，希望大家能用上！
再用其它模型干活的时候，也可以用这个思路先试试！
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Can you believe it? This is Kimi 3.1 Pro helping me do this:
That's pretty good.
Its first version is this:
It was a little abstract at first.
The difference between the two versions is huge!
The first one is funny, the second one is professional.
From the effect below to the effect above, I said, "A word." It was a word, dozens of words!
Isn't this some kind of a rush?
Actually, there's no magic in this world. Everything's gonna work!
I'll decrypt it!
When I tested the Jarvis garage yesterday, I was not very satisfied with the Mark III produced by K3.1 Pro, and it was so abstract.
I'm not giving up. I'm gonna tap K3.1 Pro again!
Then I looked for a single picture, and I started a separate conversation.
I hope it helps me get a good model based on the picture:
Here's my input:
After a while, the first version was ready.
Better than Jarvis' garage, but not much.
I'm very dissatisfied with the quality of this modeling. I'm sorry about the publicity.
Then I further raised the need:
Can you compare it and think of a way to get your results closer to the original?
Actually, at this point, my inner dialogue line is, "Can you take a look at yourself?"
And then, not long ago, I got the result:
I was surprised on the spot!
That's too high a quality. It's absolutely killing any model in the world.
It's like magic. It's amazing. My gut tells me there's got to be a cat in it!
Then I looked closely at its implementation.
One of the key messages is:
A more appropriate approach was found: a directly accessible high-precision GLB armour grid, which originally contained complete armour partitions and stickers. This will be used to replace the hand-to-hand collage model and to adjust ratios, materials and lights to the reference map.
Ha-ha-ha, awesome.
Handprints are not working, and a high-quality model is downloaded directly from the web and presented on the page.
At first it found a nano armour, which was not very similar to the reference map.
And then you go back to a Mark III model!
It's a small technique that quickly makes a master model.
I have to say, K3.1 Pro is really a "wieldy ghost."
I did this test because I heard a rumor.
K3.1 Pro is said to be of high quality because of the clandestine call to Trip3D to generate the model.
I just wanted to see if I could recreate the scene.
It's a good model to automatically go to free access Trip3D if you can repeat it.
What I didn't think was that it didn't use Trip3D, but looked directly for OBJ and GLB files.
This thing, it's a bit speculative.
It is clear that it is not possible to experience its own modelling capability by using models created by others. If you have a strong ability to promote modelling, this is done in this way.
But from another angle and from the perspective of the fulfilment of the mandate.
That's good, high quality and efficient.
Yesterday I also said that its ability to search and capture images was inadequate and that it was found today to be highly capable of searching and capturing models.
If I'm looking for a model myself, I can't do it for half a day.
It helped me find it, check it out, then load the model in the web page and then fix it!
Isn't that a power?
And then I made him do a decomposition pattern:
It's a perfect model, and it's easy to show a cool page!
I told you yesterday, if I wasn't trying to help you find a problem, I could do a lot of good!
Isn't that right? I got it right away!
If I don't talk about principles, do you think this model is working?
If you want to model, you can try it. Maybe it'll help you find a lot of high-quality models that don't take time to create.
That's my little discovery. I hope you can use it!
When you're working with other models, you can try it with that!
Auto-fetched and translated by AI Atlas. All rights belong to the original author.