我们在做 AI 画面或视频时，有些画面脑子里明明想得很清楚，写成提示词就开始打结，有些时候、有些动作或运镜根本无法用语言表达出来，如果采用抽卡的形式，每次的不完美都是给平台“捐款”。
于是，黄啊码昨晚连夜折腾了个三维动作创作工具，叫码上动，现在碰到这种画面，我可以先把人物白模拖进场景，再把动作和运镜摆出来，最后拿着参考图或参考视频继续创作。
而人物的控制方面，该工具更是控制到了手指关节，听上去感觉很抽象，可以看看上边的女性角色，她的手指指着人就是专门用这个功能来控制的。
姿势库由于个人精力有限，没有专门定制太多，只提供了 10 个，但工具是支持每个动作、每个组合专门定制和保存，方便下次使用。
说得再多，还不如讲几个案例，也许就能把这个用法讲得更明白，每个案例的提示词都放在图片下边，注意，别找错地方了。
01 PART
三个人的关系，先看站位
第一组案例，黄码尝试了三人互动的场景，就来个现场抓奸的画面吧，哈哈。
在三维场景里，两个角色靠在一起，第三个角色站在另一侧伸手指向他们，动作还很粗糙，但人物之间的关系已经有了。
提示词
参考图中动作生成一幅男生跟一个女孩约会被女朋友抓到的场景，男生女生形象自由发挥，只参考动作
效果示例：
到了效果图里，场景变成了咖啡馆门口，人物也有了服装和表情看，看这组对照，卧槽，妥妥情绪和动作基本到位了
02 PART
场景物品太多，用几何物品代替
第二组案例，我觉得特别适合拿来解释基础物品的用途，物理世界的东西实在太多了，单纯一个一个手搓真 TM 累，于是我想了一下，大模型不就是从物理数据拿数据去训练的吗？让它自己发挥不香？于是就有了下图:
提示词
参考图片生成玄幻角色在室内的场景，图中的物品和角色都要换成真实场景的东西，比如凳子就要改成椅子，角色改为古装男子
场景中有一个坐着的人、一只蜥蜴和一只展翅的鸟，旁边放着几个简单体块：正方体写着方块 01，长方体写着桌面 A，圆柱体标了数字，人物坐着的圆球写着凳子。
黄啊码不是随便做这些东西的，蜥蜴代表爬行动物，鸟代表飞行动物，人就不用说了，其他的几何物品可以标记任何你想表达的物品。这些一定要记住，记不住就多记几次。
效果示例：
对应的效果是一间古风书房：人物坐在右侧看书，鸟停在左边的架子上，蜥蜴待在地面，桌上摆着书和笔筒。
这组参考提供的是大致的空间安排，场景先从简单体块开始，逐步补充物品含义，等构图差不多了，再交给AI 做后续环节丰富细节。
还有，当前工具中的物品支持尺寸调整，也能标数字或名称，这个功能看着朴素，真正摆起复杂场景来，能帮自己认清每个体块的用途，也方便和别人沟通。
03 PART
多人可以，单人的会不会挂了？
第三组是一个单腿抬起、双手举起的舞蹈姿势。
其实我从一开始就想弄单人的，只是开发的图中开发嗨了，就忘记单人才是重中之重，于是，咱们就来弄个单人的案例吧。
写个比较难点的吧，我们一般形容女孩子活泼可爱，是不是非常简单，只要这四个字，但应该怎么形容？能对应的动作太多了，可具体到腿向哪边弯，双手举到多高，身体偏向哪一侧，还是得把姿势交代清楚才行。
所以我勉为其难地用了一个男性的角色，给 AI 加加难度，哈哈哈，翘腿，手舞足蹈，emmm，确实有点娘，看 AI 的能力了。
提示词
一位亚洲美女微笑模仿图片中的动作跳舞，背景是舞台
效果示例：
AI 没有让我失望，这组对照里，三维小人的大体姿势被用于舞台人物的创作，服装和环境则有了新的表达，总的来说就是：中！！！
我想要的就是这种分工：在工作台里把姿势调整到接近想法，保存成参考，再去尝试画面风格，后续修改时，也能回到同一份姿势继续调整。
04 PART
从一张参考图，走到一段参考视频
单纯用参考图一点难度都没有，还是欠缺点什么，一想到 seedance2.0+支持视频参考，顿时热血沸腾，大胆梭哈了我的 codex 额度，可前面几组解决的是某个瞬间长什么样，视频还要考虑动作怎样发生，以及镜头怎样跟着变化。
于是我在此基础上加上了画面定格和视频录制，从画面到人物运动到运镜全部安排上，人物移动是走动还是跑，速度是否快，都会在参数里边得到完整的定义。
提示词
把视频中的人物改成黄啊码，图中的鸟改成无人机，参照视频的动作和运镜，动漫皮克斯风格，黄啊码在闹市中闯红灯，被警犬和无人机追赶
这一组适合拿来理解动作与运镜参考的用法：先用容易调整的三维角色，把移动方向和镜头过程做出来，再用于后续的视频创作，成片里的角色和场景，可以在外部制作环节继续变化。
为了方便做这种参考，我把定格和录制放进了同一条轨道，可以先摆好关键姿势，保存定格，再录一段拖动操作，把它们排在一起预览。
需要角色从一个位置到另一个位置时，还可以设置走过去或跑过去，每张定格的停留和衔接时间能单独调整，编排完成后导出参考视频。
最后效果如下，seedance 从来不会让我失望，就是参考的图片中啊码是闭着一只眼睛的，它依旧保持着，有点幽默了，但是 seedance 是真的贵，加上我的积分不多了，大家就将就看吧。
在这组视频参考里，有移动的人物、地面的蜥蜴和展翅的鸟，对应的效果示例变成了城市里的动画追逐，画面中出现了新的角色和无人机，再看它的运镜，基本遵循了参考视频中的方式，这个必须得狠狠点赞了。
介绍百遍，不如自己用一遍，代码已完全开源，想用的话可以拿去使用。
项目地址：https://github点 com斜杆huangama666/mashangdong
OK，今天的分享就到此，我是黄啊码，码字的码，如果觉得我说得有道理，欢迎一键三连，如果觉得有异议，欢迎评论区指正，我们都是 AI 时代的共创者。
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。
When we're doing an AI image or video, some of the images are clearly thought out, and prompt starts closing, and sometimes, some actions or mirrors can't be expressed in language, and if they are in the form of a card, each imperfection is a “contribution” to the platform.
So, the yellow code put up a three-dimensional motion-making tool, called code-mart last night, and now that I've got this image, I can drag the character White Moot into the scene, put the action and the mirror out, and then continue with a reference map or a reference video.
In terms of the control of the person, the tool is controlled by the finger joint, which sounds abstract and shows the role of the woman on top, a function that is specifically controlled by her fingers.
Owing to their limited personal energy, the poser is not customised much, and only 10 are available, but the tool is to support each action, each combination is customized and preserved for use next time.
More than a few cases, it might be possible to make this more clear: prompt in each case is placed below the picture, so be careful not to go to the wrong place.
01 PART
The three-person relationship. Let's see the position first.
In the first set of cases, yellow code tried the three-person interaction, so let's get a picture of the rape on the spot, ha ha.
In a three-dimensional scenario, the two players are side by side, the third one is on the other side reaching out to them, and the movement is rough, but the relationship between the characters is already there.
prompt
The action in the reference diagram creates a picture of a boy dating a girl caught by a girlfriend.
Example of effect:
In the results chart, the scene turned into a coffee shop door and the characters looked at their clothes and faces.
02 PART
There are too many scenes to replace them with geometry.
In the second set of cases, I thought it was particularly appropriate to use it to explain the use of the basics, and there was a lot of stuff in the physical world. Does it smell good? So here's the picture:
prompt
The reference picture creates an indoor scene of the phantom role, and the objects and characters in the picture have to be changed to the real scenes, like the chair and the role to the man.
There was a sitting person, a lizard and a winged bird with several simple blocks: squares with square 01, rectangulars with desktop A, cylinders with numbers, and people with a ball with stools.
Yellow codes are not randomly made, lizards are reptiles, birds are flying animals, people don't have to say, other geometrys can mark anything you want to express. It is important to remember these, but not many times.
Example of effect:
The effect is an ancient wind library: people sit on the right side reading books, birds on the left shelf, lizards on the ground, books and pens on the table.
This set of references provides a broad spatial arrangement, starting with simple blocks, gradually supplementing the meaning of the item, and then giving it to AI for further details.
Also, the items in the current tool support size adjustments and can also be numbered or named, a function that looks simple, really presents complex scenes, helps to identify the uses of each block and facilitates communication with others.
03 PART
How many people can do it?
The third group was a single-legged dance position with hands raised.
I've been trying to get a single person from the beginning, but I've just developed a picture that's good enough to forget that it's a top priority, so let's get a single person case.
It's harder to write, isn't it very simple to describe a girl like that? There are too many counter-actions, but the specifics are to which side of the leg is to bend, how high your hands are, which side your body is to turn, and the position is to be clearly explained.
So I was forced to use a male role to make it harder for AI, ha ha ha ha, stomping, hand-dancing, emmm, it's kind of a bitch to see AI's ability.
prompt
An Asian beauty smiled and danced in the image, in the stage.
Example of effect:
AI didn't disappoint me. In the cross-section, the broad position of the three-dimensional people was used for the creation of stage figures, while the costume and the environment had a new expression, in general: "C!"!
What I want is this division of labour: to change positions to close to ideas, to keep them in reference, to try the image style, and then to move back to the same position.
04 PART
From a reference map to a reference video
It's not difficult to use a reference figure alone, or it's not easy to seedance 2.0+ to support a video reference, and it's hot, and it's hot, and it's hot, and it's hot, and it's hot for my codex quota, but the first group of solutions is what it looks like in a moment, and it's about how the action happens and how shot changes.
So I put this in graphics and video recording, from image to character movement, to mirrors, whether the person moves or runs, and whether the person moves fast or not is defined in full by the parameters.
prompt
Turn the characters in the video into yellow codes, the birds in the picture into drones, video action and mirrors, comic pixie style, yellow code in red lights in the city, chased by dogs and drones.
This set is appropriate to understand the use of action and mirror reference: first, using a three-dimensional role that can be easily adapted, making the movement direction and the shot process, and then using it for subsequent video creation, making the characters and scenes in the film, which can continue to change in the external production chain.
For ease of reference, I put stationary and recording in the same orbit, so I could set up key positions, save the grid, then record a drag operation, sort them together and preview them.
When the role is required to move from one position to another, you can also set the walk-in or run-in, and the stop and connect time of each set-up can be adjusted separately and the reference video exported after it has been organized.
The end result is as follows, soedance never lets me down, but in the picture of the reference, it's closed to one eye, and it's still a little funny, but it's really expensive, and I don't have much credit, so let's see it.
In this video reference, there are mobile figures, lizards on the ground, and wingsy birds, and the corresponding example becomes animated chase in the city, with new characters and drones appearing in the picture, looking at its mirrors, and basically following the way it is in the reference video, which has to be very, very good.
The code is completely open and can be used if you want.
Project address: https://github point com slasher huangama666/mashangdong
OK, that's all for today's sharing. I'm yellow, code, if I'm right, welcome three keys, if I have any objections, welcome comment, we're all co-founders of the AI era.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.