快问快答，这是钢铁侠还是咸蛋超人？
我们先来个轻松的，后面再给大家分享一些有用的玩法。
Kimi 这次 3.1 Pro 的发布好像没有 K3 热闹，外网还闹出了一点事情。
我估计，主要还是怪“美模”把标准拉太高了，大家已经很难因为简单的效果而兴奋了。
说实话，我也是这样的！
我趁着现在 token 充足，给大家挖掘挖掘，看看有没有好用的场景或者功能。
这次 K3.1 Pro 发布，他们官网多了一个“3D Make”的功能，看宣传视频也是厉害得很。
我一直在找一个比较好的 3D 建模方案，所以对这个功能还是挺感兴趣的。
这个功能是随着 3.1 的发布而新增的！打开官网就可以用：
但是，这个功能本身好像并不是完全由 K3.1 Pro 来完成的，而是通过后端脚本去调用其它第三方 3D 模型来生成的，这个三方模型是 Tripo3D。
这不能体现模型本身的能力，但是能用上 Tripo3D 也是挺好的！
这个功能包含了三个子项，可以做 3D 网页、3D 建模，以及 3D 打印！
下面有很多例子，点一下就能做同款：
这个水月观音的建模和视觉效果看起来完成度很高。
完整提示词是这样的：
用 Blender Python 创建一个可重新生成、可编辑、可导出的三维场景，并完成最终静帧渲染。整体效果严格服从主参考图：高空俯视、右侧浮岛、北缘巨像、左侧云海、冷灰蓝光照。最终成果必须来自真实三维场景，不得用参考图贴片或二维生成图冒充三维渲染。采用展示模式，提供主视角、建筑、雕像和顶部检查相机，可旋转查看与缩放检查，不要求实时交互。参考图是唯一整体效果依据（机位、构图、岛形、雕像位置、建筑范围、云海遮挡与明暗）；其余三张引导图分别只提取中式古建细节、巨像与城市的尺度逻辑、体积云层次与冷灰蓝光照，不得引用其构图或机位。发生冲突时一律以主参考为准。画面硬约束：输出 1920×1080 主画面及 2 倍图。高空近垂直俯视（俯角约 55°–70°）、长焦、全画面深焦。岛屿投影为略倾斜的横向椭圆，位于画面右侧；左侧约一半画面为连续云海，不出现地平线或整片天空。巨型浅色风化石像位于岛屿北缘，以背部或侧背面为主，高冠饰与衣褶清晰，下部被云半掩。暗色密集建筑群集中在岛体下半部，屋脊轮廓清晰可读。主画面不露出完整悬空岛底；白色区域是体积云，全场景禁止积雪与冰雪材质。建筑采用"标准构件 → 参数化模板 → 院落组合 → 群落实例"流程，全部放入独立集合：先实现瓦垄、斗拱、柱梁开间、脊兽、栏杆、城墙段等标准构件，再组合出十类建筑模板（门厅、主殿、厢房、楼阁、凉亭、门楼、高台殿、十字脊抱厦殿、钟鼓楼、连廊院）。岛缘布置随岛形变化的闭合城墙带，中环以 Poisson Disk 布置 12–16 个院落单元，核心为高台对称区与少量楼阁，中央北缘为雕像留空。材质低饱和：青灰瓦、灰黑梁架、暗红柱、灰石墙。记录随机 seed，保证同一次生成可复现。岛体为直径约 200m 的不规则浮岛，偏暗青灰风化石；雕像为浅色风化石灰岩，保留冠饰、肩臂、衣褶与雕刻起伏，不得洁白或塑料化。云海使用真正的三维体积材质与多层 Noise：连续云海填满左侧，贴岛云流吞没岛缘并半掩雕像下部，局部低云打断城市内部；禁止平面云贴图与均匀白地毯。光照为左上方大面积柔和主光加低强度冷环境填充，最终用 Cycles 渲染并降噪；预览阶段可降低采样与体积步进，但不得删减建筑或改变主相机构图。按"环境检查 → 岛体 blockout → 雕像对齐 → 建筑生成 → 云海与光材 → 渲染检查"的顺序推进，每步通过后进入下一步；每轮主要渲染必须生成与主参考的并排比较图、叠加图和差分图，实际查看后继续修正，不得只凭日志或数值宣布完成。
说实话，我是从来不会写这么多提示词的，这个提示词大概率也是 AI 生成的。大家不会写的话，可以参考一下！ （我运行完了，并没有宣传视频的效果）
上面这个例子看起来有点复杂，我测了一个比较独立的：“做一座悬浮在空中的云端花园”。
然后，就会自己去后台跑了。
跑完之后生成了三个文件，一个网页。
三个文件分别是 blend、glb、png。
网页展示如下：
Blend 文件太大了，无法在线预览，GLB 可以打开，效果如下：
你们觉得怎么样？
这种设计和风格，女性可能会比较喜欢，我不知道如何评价，我估计其它模型也是完全能做到这个效果的。
官方例子其实没太大意思，我们直接上强度。
把这个参考图扔给它：
最终得到的结果是这个：
对比一下，还原度怎么样？ 虽然我不想笑，但是总感觉有点喜感！
它第一个版本其实不是这样的，头部很抽象，我就让它再做了一个版本，然后可以看到头和脖子部分是重新生成的，就变成这个圆圆脑袋了。
我想重试一次，发现 API 拒绝，估计已经涉及到版权问题了！ 钢铁侠 Mark 这个看来是有点难搞了。
下面给大家展示一下这个功能的正确用法！
比如我之前做了一个“三国版吃鸡”，里面的人物都很抽象，现在就可以用这个功能来优化一下建模了。
首先让 Codex 帮你生成一张“关羽”的角色图：
然后生成一张三视图：
然后把这个图片扔给 Kimi 3D Make，让它根据参考图生成模型！
然后，我们就有了这个模型了：
最后让它帮我做一个类似参考图的三视图：
根据它生成的GLB文件，就可以导入到游戏里面了。
只要你不太抠细节，这个建模质量完全够用了。
另外这个功能好像不消耗KimiCode的套餐用量：
具体限制不明，生成了几个模型，还没有遇到限制。
这个功能可以玩一玩，Tripo3d本身是需要收费的，现在直接通过Kimi调用 ！ 使用的时候默认可能会用Blend建模，所以最好指定一下用tripo3d来建模。
从整体建模的角度来说，比Claude GPT手搓要好多了，大致上看起来还是像模型一样的！
你要说建模细节和整个渲染效果，还是差很多。
所以，目前的AI模型建模仅限于要求不高的场景，高精度，高质量建模，还是得专业的来。
相比程序员，建模师的还能再嚣张一阵子！
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。
Answer the question. Is this Iron Man or the salted egg man?
Let's have an easy one, and then we'll share some useful games.
Kimi this time 3.1 Pro's publish seems to have no K3 and there's something going on out there.
I think it's mainly because of the fact that the “mode of beauty” raised the standard too high, and it's hard to get excited by the simple effect.
To be honest, I'm the same!
I'll dig for you while I'm here token enough to see if there's a good scene or function.
This time, K3.1 Pro publish, their network added a “3D Make” feature, which is also very good for publicity videos.
I've been looking for a better 3D modeling scheme, so I'm still interested in this function.
This function was added with the 3.1 publish! Open the network to use:
However, the function itself does not appear to have been entirely performed by K3.1 Pro, but is created by using other third-party 3D models, which are Tripo3D, through end scripts.
It doesn't reflect the power of the model itself, but it's good to use Tripo3D!
This feature contains three sub-items to do 3D web pages, 3D modelling, and 3D printing!
There are many examples of how we can do the same:
The modelling and visual effects of this water moon view seem to be quite complete.
Here's the whole prompt:
Creates a three-dimensional regenerated, editable, exportable scene with Blender Python and completes the final static dye. The overall effect is strictly in line with the main reference map: the high-altitude view, the right-side floating islands, the north-side giant, the left-side cloud sea, the cold gray blue light. The final result must come from a real three-dimensional scenario and must not be used as a three-dimensional rendering using a reference image or a two-dimensional image. The presentation model, which provides a master perspective, architecture, statues and top-screening camera, can be rotated and scaled, without requiring real-time interaction. The reference diagrams are the only overall effect basis (space, structure, island shape, statue location, scope of construction, cloud and sea cover and brightness); the remaining three guide maps take only Chinese ancient details, scale logic of mega-images and cities, volume cloud levels and cold gray blue light, and may not be quoted in their structure or position. In the event of conflict, the primary reference shall prevail. Image hard bound: Output 1920 x 1080 Main image and 2 times figure. High-altitude near-vertical overlooking (approx. 55°-70° below), long-focused, full-photo intense. The island is projected as a slightly tilted horizontal ellipse, on the right side of the picture; on the left, about half of the picture is a continuous cloud sea, without a horizon or an entire sky. The giant, light-coloured weather figure is located on the northern edge of the island, mainly on the back or the back side of the side, with a high crown and a clear coat of clothing, and the lower part is covered by clouds. The dark-coloured complex is concentrated in the lower half of the island and the contours of the ridge are clearly readable. The main picture does not reveal the bottom of the entire suspended island; the white area is a mass cloud, and the whole scene prohibits snow and ice material. The construction is based on a standard building module, a parameterized template, a compound combination, a cluster implementation process, all of which are placed in an independent assembly: first, standard building components such as waffles, arches, beams, spinal animals, bars, walls, etc., are to be combined into 10 types of building templates (door hall, main house, compartment, attic, saloon, door building, tower, tower, crestals, bell drum tower, coronary). The island's edge is a closed wall belt that varies with the island's shape, with the centre ring being 12-16 courtyard units, with the core being a symmetrical area and a small number of attics, and the central north edge leaving the statue empty. Low saturation of materials: green ash, gray and black beams, dark red pillars, limestone walls. The records are randomly seen to ensure that the same generation is repeated. The island is an irregular floating island with a diameter of approximately 200 m, with a dim, pyrotechnic stone; the statue is a light-coloured limestone, which preserves corsets, shoulder arms, coats and sculpture, and is not white or plasticized. The cloud sea uses real three-dimensional mass and multilayer Noise: successive clouds fill the left side, adrift clouds engulf the island's edge and cover the lower part of the statue, and local low clouds break into the city; flat clouds are prohibited and flat white carpets. The light is filled with a large area of soft and main light and low intensity cold environment above the left, which is ultimately rendered with Cycles and noise; the preview phase can reduce sampling and volume progress, but may not delete buildings or change the main body chart. Proceeding in the order of "Environmental inspection of the island's blockout statue aligning the building to generate clouds, oceans and light materials rendering inspection" and entering next step after each step; each round of major rendering must generate parallel, stacked and differential maps with the main reference, and continue to be amended after physical inspection, and may not be declared to be done solely on the basis of logs or values.
To be honest, I've never written so many prompts, and this prompt rate is also AI-generated. If you can't write, you can read it!
The example above seems a little complicated, and I measure a more independent one: “to make a cloud garden floating in the air”.
Then he'll run himself backstage.
Three files, one web page, were generated after the run.
The three documents are: blend, glb, png.
The web page is shown below:
Blend files are too big for online previews, and GLB can open them, with the following effects:
What do you think?
This design and style may be more popular with women, and I don't know how to evaluate it, and I think other models can do exactly that.
The official example is not very interesting. We're directly strong.
Throw this reference to it:
The end result is this:
I don't want to laugh, but I feel a little happy!
It's not the first version of it, it's abstract, and I'll make it another version, and then I'll see that the head and neck parts are regenerated, and it's this round head.
I think retry once found that API refused, and probably already had a copyright problem!
Let's show you how it works!
For example, I made a "triple chicken-eating" before, and the characters in it are abstract, and now we can optimize modelling.
First of all, Codex helps you to create a "Kwan Yu" character map:
Then you generate a three-view view:
And then throw this picture to Kimi 3D Make, so it can generate models based on reference maps!
And then we had this model:
Finally, it helps me make a three-view of a similar reference map:
Based on the GLB files it generates, you can import into the game.
As long as you don't know the details, the quality of this model is perfect.
This feature does not seem to consume KimiCode’s package:
Specific restrictions are not known and several models have been produced, and no restrictions have been encountered.
This feature can play, Tripo3d itself is a charge and is now called directly through Kimi! Defaults may be modelled with Blund when used, so it's better to specify tripo3d.
It's much better than Claude GPT's hand in general, in terms of modeling.
You're going to say the details of modelling and the whole rendering, but it's still a lot worse.
Thus, current AI modelling is limited to less demanding scenarios, high-precision, high-quality modelling, or professional.
The role of a modeler can be increased for a while!
Auto-fetched and translated by AI Atlas. All rights belong to the original author.