即梦AI肢体与动态人物穿搭
#肢体与动态#人物穿搭
中文翻译
第 12 天。用人工智能制作我梦想了 15 年的游戏。
今天游戏里又多了两种角色:手持双手剑的狂野骷髅剑士,以及骷髅弓箭手。我还制作了它们的城堡,之前的一篇帖子里展示过。
我原本很担心弓箭手。让箭和投射物正常运作,听起来会是另一个完全独立的麻烦。结果这一部分相当简单。
组合动画轨道正逐渐成为我工作流程中的重要环节。我一直把 Mixamo 平台的动画片段的部分内容,与动作捕捉以及 GPT 根据我的 Seedance 视频参考制作的动画混合使用。
从骷髅视频中获得动作捕捉数据需要多做一步。我得先把视频里骷髅的动作转移到一个人身上。只有这样,动作捕捉工具才能识别动作,并将其转换成动画数据,供 GPT 用在骷髅身上。
对于某些动作,GPT 只看参考视频就能取得不错的效果。更复杂的动作,比如把手伸到背后,仍然可能让它犯难。这时,我可以让它从某个 Mixamo 平台的动画片段中提取合适的手臂动作,或者从动作捕捉录制中提取腿部动作,同时保留已经正常运作的部分。
狂野骷髅就是一个很好的例子。GPT-6 一直难以摆好双手的位置,让它正确握住双手剑。我为它录制了动作捕捉,获得了所需的手臂动作,但那次捕捉里的腿部动作并不好。
于是,GPT 将我动作捕捉中的手臂动作,与它根据视频参考制作的动画中的腿部站姿结合起来。这终于让它正常运作了。
随着这套工作流程逐渐成形,我在一天内完成了这两种角色,还有时间做游戏的其他工作。我做得越来越快,看到角色阵容不断扩大,感觉很好。
明天我会继续制作其他角色。
原帖
Day 12. Building the game I've dreamed of for 15 years with AI.
Two more characters in the game today: wild skeleton swordsmen with two-handed swords, and skeleton archers. I also worked on their castles, which I showed in an earlier post.
I was worried about the archers. Getting the arrow and projectile working sounded like it would be a whole separate headache. That part turned out to be pretty simple.
Combining animation tracks is becoming a big part of my workflow. I've been mixing pieces of Mixamo clips with mocap, and with animations GPT builds from my Seedance video references.
Getting mocap from the skeleton videos took an extra step. I had to transfer the skeleton's motion onto a human in the video first. Only then could I get the mocap tool to recognize the movement and turn it into animation data for GPT to use on the skeleton.
For some movements, GPT gets a good result just from watching the reference. More complicated ones, like reaching behind the back, can still give it trouble. Then I can ask it to take the right arm movement from a particular Mixamo clip, or the legs from a mocap recording, and keep the parts that already work.
The wild skeleton was a good example. GPT-6 kept struggling to position both hands so it could hold the two-handed sword properly. I recorded mocap for it, which gave me the arm movement I needed, but the legs in that capture weren't good.
So GPT combined the arms from my mocap with the leg stance from the animation it had made using the video reference. That finally got it working.
With this workflow taking shape, I got both characters done in one day and still had time for other work on the game. I'm getting faster at this, and it feels good to see the roster growing.
Tomorrow I'll keep going with the other characters.
