本文来源:https://mp.weixin.qq.com/s/1XMNb9UOPAeNxcjxqHDnOg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

最近的 AI 视频都疯了一样开始做 Video Agent 模式了,这是自 Code Agent、Work Agent 之后第三个爆发的专业 Agent 领域。
但当我们提到 Agent 的时候,一定也会提到 Skill,因为它可以让 Agent 在不同领域、不同模式下发挥出更大的作用。
最近正式上线的 updream,一出手就直接来一个大的,它在自己的平台里搭建了一套 Video Skill 的 Gihub 平台,让所有的创作者都可以在上线分享自己的视频创作 Skill。
先送个福利,下方入口获平台 1000 积分。
www.updream.cn/bsource=xiaocherry
它还将 Skill 从 Agent 的对话框中彻底解放,真正摆到了画布里,让多个 Agent、多套 Skill 可以互相协作工作,让灵感不再受限。

粗粗数了下,现在光是上架的 Skill 已经超过了 200 个,并且涵盖提示词、图片生成、创意策划、AI 工具、视频生成、视频制作、角色设计、分镜、剧本等 70 多种分类,把影像创作的全流程都包圆了。
这是一个非常夸张的数字,而且会对视频创作非常有帮助。
之前做视频,我会感觉自己是一个人的孤身奋斗,我做完了一个非常牛逼的工作流后无法和别人分享,我创作灵感枯竭的时候没办法找人帮助。
但现在在 updream 的技能广场里,我可以发现很多想法志同道合之人,它们的 skill 会给我一种,“咦,这不就我刚才想的吗!” 那种感觉,非常棒。

我挑选了一个当前比较火的 Skill 「第一人称视角萌宠视频提示词生成器」,它支持输出文本、视频内容,直接安排到 Agent 里体验一下子。

进入 updream 画布,侧边栏打开 Agent 对话框之后,输入 / 就可以调用 Skill ,然后填入自己的诉求,比如 生成一个和宠物金刚鹦鹉互动的搞怪 vlog 风格混剪视频,要有字幕台词、气泡贴纸、片段式的内容,体现金刚鹦鹉说话搞怪,和拍摄人物互相怼的场景。

然后 Agent 会直接在画布里创作工作节点,并额外展开一个对话框来进行工作流推进。
是的,Agent 和 Skill 不再局限于对话框里存在,它们已经可以独立的存在画布之上。别的创作平台,每次我只能调用一个 Skill,当遇到问题返工时我还得一条条对话内容看过去,看看到底上次这个节点产出是在啥时候,用了啥 Skill。
而现在,直接在画布里就能找到节点形态的 Skill,这真的太方便了。每一个 Agent/skill 节点,就是独立的对话框,里面不仅有 Agent 思考的全部过程、产出物记录,甚至可以多轮对话不断优化产出。

主题、定位、分镜、剧本、参数、动态进度,每一条都可以在画布节点里打开去看,遇到了问题直接改。

所以在这种模式下,我设置可以同时并发多个 Agent 来工作。
画布上每一个节点的任务,都可以由创作者自主分配到对应的 Skill 上,生图 Skill 就专门盯着生图,提示词优化的就专门走文本创作,生视频的时候不同视频场景根据需要让不同 Skill 去优化。
Updream 把创作 Agent 这件事的主动权彻底交还到创作者手上,想怎么用就怎么用,它只会融入你的工作流,而不会强迫你去适应。

对了,我还发现 updream 的 Agent 已经进化出了 sub-agent 的控制能力,它会像 Code Agent 一样在复杂任务时主动拆解工作,并推送到不同的 sub-agent 身上,各司其职的在画布上协作工作。

在以前,我们生成完视频后还需要把视频一个个的下载到本地,然后通过剪辑工具把它们拼接起来,或加上字幕、片花、贴图等等。
而在 updream 里,直接内部集成了一个剪辑工具,可以导入本地素材、项目素材,甚至可以进行调色、字体、滤镜处理。

仔细去看,会发现它的剪辑功能也非常强大,音轨分离、磁吸、多轨道、蒙版、音频处理、位置变化等功能全部免费使用,这妥妥实现了我梦寐以求的在线剪辑啊。

剪完的时候,我们还可以返回画布去看看成片的时间轴,和画布互相对照一下。而成片也可以直接合并导出到画布,甚至达芬奇里继续精修、优化。

哦对了,除了这个视频届的 Github ,还有一个好消息就是updream 在近期会第一时间接入 Seedance 2.5,原生支持 30s 视频的创作。
而且每一个视频的创作信息,都通过在平台内追溯查询,用的什么模型、用的什么参数都能看得见,绝不惨水降智,让每一分钱都花的货真价实。

对创作者来说,IP 真的非常重要,在很多时候模型会因为版权问题拒绝给我们生成内容,哪怕我们只是简单的做一个 AI 二创。
而在 updream 里,这个问题直接被官方解决。叔叔家的《凡人修仙传》《时光代理人》《天官赐福》等等经典 IP 都已经合法版权的授权给了平台,创作者可以在平台里生成相关 IP 的内容,而不用担心被模型拒绝。

文章写到最后,其实还有一个很重磅的消息。
看看下面,updream 的本地客户端版本。是的,叔叔把平台直接做成了本地客户端,它可以直接让你接入自己的 API key 用来生成图片、视频。而且不限制你使用所有网页端的功能。
全都可以用!如果你走了一些更好的渠道,你完全可以把 updream 的 skill 社区、多 agent 工作流搬到本地!


往期推荐:
|
NAS 教程: 1️⃣无限子弹来了,AI NAS 新标杆,实测性能猛兽 — 铭凡N5 MAX 2️⃣一套免费4K视频会议系统,立省腾讯会议 VIP,NAS 搭建 MiroTalkSFU 3️⃣「肌肉记忆法」学外语,给 NAS 装一个孩子学习的超级外挂 |
|
AI 玩法: 2️⃣分享 WorkBuddy+蚁小二 CLI,全平台自媒体,从创作到运维都打通了 3️⃣实战 WorkBuddy+反爬+爬虫固化 Skill,6大能力搞定浏览器自动化 |
Source: https://mp.weixin.qq.com/s/1XMNb9UOPAeNxcjxqHDnOg If there are copyright issues, please notify us via email: 13951431913#139.com or submit a comment on the right.
Recently, AI video tools have all gone crazy building Video Agent modes. This is the third explosive professional Agent field after Code Agent and Work Agent.
But when we mention Agent, we definitely also mention Skill, because it allows Agent to play a greater role in different fields and different modes.
updream, which recently officially launched, came out swinging right away. It built a GitHub-style platform for Video Skills within its own platform, allowing all creators to share their video creation Skills online.
First, here's a little perk: use the entrance below to get 1,000 platform credits.
It also completely freed Skill from the Agent chat box and truly placed it on the canvas, allowing multiple Agents and multiple Skills to collaborate with each other, so inspiration is no longer limited.
A rough count shows that there are already more than 200 Skills listed, covering more than 70 categories including prompts, image generation, creative planning, AI tools, video generation, video production, character design, storyboarding, and scripts, covering the entire workflow of video creation.
This is an absolutely astonishing number, and it will be very helpful for video creation.
In the past, when making videos, I felt like I was fighting alone. After completing a really awesome workflow, I couldn't share it with others, and when my creative inspiration ran dry, I couldn't find anyone to help.
But now, in updream's Skill Plaza, I can find many people with similar ideas. Their Skills give me that feeling of, "Hey, isn't this exactly what I was just thinking about!" It feels amazing.
I picked a currently popular Skill, "First-Person Perspective Cute Pet Video Prompt Generator." It supports outputting text and video content, so I directly assigned it to an Agent to try it out.
After entering the updream canvas and opening the Agent chat box in the sidebar, type / to invoke the Skill, then fill in your request, such as generating a funny vlog-style mashup video interacting with a pet hyacinth macaw, including subtitles and dialogue, speech bubble stickers, and fragment-style content, showing the macaw saying funny things and bickering with the person filming.
Then the Agent will directly create workflow nodes on the canvas and open an additional chat box to advance the workflow.
That's right, Agent and Skill are no longer limited to existing inside the chat box; they can now independently exist on the canvas. On other creation platforms, I could only call one Skill at a time. When I hit a problem and had to rework things, I had to go through every line of chat to figure out when the last node output was produced and which Skill was used.
But now, being able to find Skills in node form directly on the canvas is incredibly convenient. Every Agent/Skill node is an independent chat box, containing not only the Agent's full reasoning process and output records, but also allowing multi-turn conversations to continuously optimize the output.
Theme, positioning, storyboard, script, parameters, and dynamic progress — every single item can be opened and viewed in the canvas node, and if there's a problem, you can directly edit it.
So in this mode, I can even run multiple Agents concurrently to work at the same time.
The task of every node on the canvas can be independently assigned by the creator to the corresponding Skill. The image generation Skill focuses on image generation, the prompt optimization Skill focuses on text creation, and when generating videos, different video scenes can be optimized by different Skills as needed.
Updream truly hands the initiative of creative Agent work back to creators. You can use it however you want. It only integrates into your workflow; it doesn't force you to adapt.
Oh right, I also discovered that updream's Agent has evolved sub-agent control capabilities. Just like Code Agent, it will proactively break down complex tasks and push them to different sub-agents, each doing their own job and collaborating on the canvas.
In the past, after generating videos, we had to download them one by one to our local devices, then use editing tools to stitch them together, or add subtitles, bumpers, stickers, and so on.
But in updream, an editing tool is directly integrated. You can import local materials and project materials, and even perform color grading, font, and filter processing.
Looking closely, its editing features are also very powerful. Audio track separation, magnetic snapping, multi-track, masks, audio processing, and position changes are all free to use. This truly realizes the online editing I've always dreamed of.
When editing is done, we can also return to the canvas to check the timeline of the final video and compare it with the canvas. The final video can also be merged and exported directly back to the canvas, or even continue being refined and optimized in DaVinci Resolve.
Oh, and besides this GitHub of the video world, another piece of good news is that updream will soon integrate Seedance 2.5 as one of the first, natively supporting the creation of 30-second videos.
Moreover, the creation information for every video can be traced and queried within the platform. You can see which model was used and which parameters were used, with absolutely no watered-down quality, ensuring every penny you spend is worth it.
For creators, IP is extremely important. In many cases, models refuse to generate content due to copyright issues, even if we're simply doing an AI remix or fan creation.
But in updream, this problem has been directly solved by the official team. Classic IPs from "Uncle" like "A Record of a Mortal's Journey to Immortality," "Link Click," and "Heaven Official's Blessing" have all been legally licensed to the platform. Creators can generate content related to these IPs on the platform without worrying about being rejected by the model.
As I reach the end of this article, there's actually one more huge piece of news.
Look below — it's updream's local client version. Yes, Uncle has turned the platform directly into a local client. It lets you connect your own API key to generate images and videos, and you're not restricted in using any of the web version's features.
Everything is available! If you have access to better channels, you can completely bring updream's Skill community and multi-agent workflow to your local machine!
Previous recommendations:
NAS Tutorials:
1️⃣ Infinite ammo is here — a new benchmark for AI NAS, hands-on with this performance beast: MINISFORUM N5 MAX
2️⃣ A free 4K video conferencing system that saves you Tencent Meeting VIP — build MiroTalkSFU on NAS
3️⃣ Use the "muscle memory method" to learn foreign languages — install a super learning aid for kids on NAS
4️⃣ Xiaomi NAS hardware teardown: the hardware is praiseworthy, but the software is... well...
5️⃣ Deploy AI home asset management on NAS — the god-tier HomeBox plugin has arrived
AI Fun:
1️⃣ Rehearse first, then shoot. I got a cinematic short-drama workflow running
2️⃣ Sharing WorkBuddy + Yi Xiao Er CLI — cross-platform self-media, fully connected from creation to operations
3️⃣ Hands-on with WorkBuddy + anti-crawling + crawler hardening Skills — 6 major capabilities for browser automation
4️⃣ Another masterpiece from Tencent! Marvis is genuinely impressive, and it can even be used for NAS!
5️⃣ Hands-on tests of these 5 major tech companies' Claw models — one article explains which one ordinary people should install