3 周前发现了 Agnes AI。
那时候它刚把文本、图片、视频三个核心模型 API 全部免费开放,文本模型一周被调用了 1 万亿 Token。
3 周过去,最新数据出来了。
单周调用量 5 万亿 Token,全网白嫖党一周烧掉的 Token,相当于每个中国网民每天调用几千 Token。
图片模型一周生成了 567 万张图。视频模型一周生成了 237 万秒,差不多 650 多个小时的成片。
很明显,免费策略真的把全网创作者和开发者撬动了。
我看 Agnes 这周没闲着,整出了个新东西:Pavo。
01
Pavo 是个啥
这是一个 PC 端的 AI 创作平台。
Pavo 把图片生成、视频生成、短剧创作塞进了一个工作流,再加一个 Agent 当总指挥。
你可以理解成:不用再到处切工具,从想法到成片都在一个浏览器里搞定。
四个核心模块:
- Agent:一句话下需求
- 图片生成:海报、商品图、写真
- 视频生成:文生视频、图生视频
- 剧情短片:从剧本到成片的短剧工作流
下面具体说。
02
剧情短片:从一句话到一条短剧
这是 Pavo 区别于其他 AI 工具的地方。
市面上 AI 视频工具大都能生成单个镜头。但你想做一条完整的短剧,得自己拼剧本、做人设、分镜、配音、剪辑。
而且把他们自研的免费多模态模型放到产品里面了,可以直接免费使用。
Pavo 的剧情短片模块可以把剧情短片的生产变成一个流水线,从一个故事创意开始,最后出一条能看的短片:
剧本创作 → 人物和场景设定 → 分镜设计 → 视频生成 → 成片渲染
适合做热点短剧、职场轻喜剧、情侣剧情、萌宠拟人和品牌剧情广告
我做了一个简单测试,输入:
"晴朗的天空突然变暗,然后又大大小小的飞碟,宇宙飞船、战舰在天空中飞过,很远但是看起来很震撼。最后其中一个最大的母舰突破云层和远处的大山相撞了。"
当我输入了这个命令后,Pavo 会先确认你的需求。
逐渐确认这个创作大概的视觉风格、时长等等。
你确认之后,它会自己撰写剧本大概。你只需要确认就好了。
然后它会设计角色、场景和相关的道具。整的很专业,虽然我写的内容很简单,但是它会一步步的细化丰富。
第四步是生成角色、场景相关配图。让你确认一下是不是符合你预期的。如果觉得 OK 才继续往下走。
接下来就是写分镜脚本了。它会把你的需求拆解成好几幕,每一幕是什么运镜、画面、台词都写的清清楚楚。
然后会为每一个分镜设计关键帧。到这里其实大体视频是什么样子,你已经心里有数了。
然后就开始生成分镜视频了。你整体看看没啥问题,就可以一键合成了。
在过程中,如果有某一个阶段不符合预期,还是要提前调整。
人工介入一下会好很多。
但是要追求电影级质感的,目前还差点意思。
03
Agent:一句话跑完一条广告片
Pavo Agent 是基于自研的 Harness 框架做的。用一句话描述需求,它自己规划步骤、调模型、出素材。
举个例子,我输入:
帮我为一款香水做一张商业海报和一条 15 秒广告视频
Agent 收到后会自己拆任务,过程你不用管细节,等结果就行。
更关键的是上下文记忆。第一次跑完之后,你可以接着说提需求,多轮对话调整,不用每次重新描述整个项目背景。
04
视频模型升级:Agnes-Video-2.5-preview
而且 Agnes-Video-2.5-preview 模型升级了。
即将上线,生成速度提升一倍。
运镜更流畅,少点僵硬和突兀。物理规律还原增强,少点漂浮、穿模、不自然碰撞。
新增多参考图能力,人物和商品能保持一致。
字幕乱码改善,我刚刚生成的视频是 2.0 视频模型,2.5 会好一些。
这些升级都是冲着批量生产去的。
如果你做的是广告投放、短视频矩阵这种需要不停试错的活,2.5-preview 的提速对你最有感。
05
社会反馈
刚刚提到了这个 Agnes AI 在无限量免费提供文本、视频、生图模型调用。
但是社区里也有反馈相关问题:
比如首 Token 响应时间有时不太理想,图片模型偶发 503:在高峰期偶尔抽风。有时候 Codex 这类客户端接入时配置不当会保存。但是 Agnes 把 GitHub Issues 和 Projects 看板全部公开了,bug 跟踪、功能排期、迭代进度都能看到。
想看进度直接去看板:
https://github.com/AgnesAI-Labs/Agnes-AI/issueshttps://github.com/users/AgnesAI-Labs/projects/1
其它相关地址:
官网:agnes-ai.comPavo:https://app.pavo-ai.work/API Platform:platform.agnes-ai.com开发者文档:agnes-ai.com/doc
06
点击下方卡片,关注逛逛 GitHub
这个公众号历史发布过很多有趣的开源项目,如果你懒得翻文章一个个找,你直接关注微信公众号:逛逛 GitHub ,后台对话聊天就行了:
3 weeks ago, I found Agnes AI.
At that time, it had just fully opened its core text, image, and video model APIs for free, with the text model being called 1 trillion tokens in just one week.
3 weeks later, the latest data is out.
The weekly call volume has reached 5 trillion tokens. The free usage from the community in one week is equivalent to every Chinese netizen calling the model thousands of tokens per day.
The image model generated 5.67 million pictures in one week, and the video model generated 2.37 million seconds of footage, which is about 650 hours of finished video.
Clearly, the free strategy has successfully mobilized creators and developers worldwide.
I see that Agnes hasn't been idle this week and has launched something new: Pavo.
01
What is Pavo?
This is a PC-based AI creation platform.
Pavo integrates image generation, video generation, and short drama creation into a single workflow, with an agent serving as the overall commander.
You can think of it as a way to complete everything from ideation to finished product in one web browser, without needing to switch between different tools.
The four core modules are: - Agent: Enter your requirements in one sentence - Image generation: Posters, product images, portraits - Video generation: Text-to-video, image-to-video - Drama short film: A short drama workflow from script to finished product
Below, I'll explain them in detail.
02
Drama Short Film: From a single line to a complete short video
This is what sets Pavo apart from other AI tools.
Existing AI video tools are capable of generating single shots, but for a complete short drama, you need to stitch together the script, set up characters, storyboard, voice over, and edit.
Pavo's drama short film module turns this production into an assembly line, starting from a story idea and ending with a viewable short film: - Script writing → Character and scene setup → Storyboarding → Video generation → Final rendering
This is suitable for creating trending short dramas, workplace comedies, couple stories, anthropomorphic pets, and brand story advertisements.
I ran a simple test, inputting: "The clear sky suddenly darkened, and then UFOs, spacecraft, and battleships of various sizes flew across the sky. They were distant but looked very impressive. Finally, the largest mothership broke through the clouds and collided with a distant mountain."
When I entered this command, Pavo first confirmed my requirements.
It gradually confirmed the approximate visual style, duration, and so on.
If any stage doesn't meet expectations, adjustments should be made early on.
Manual intervention can significantly improve the result.
However, for those seeking cinematic quality, the current output is still somewhat lacking.
03
Agent: Completing an ad in a single sentence
The Pavo Agent is built on the Harness framework. You can describe your requirements in one sentence, and it will plan the steps, call the models, and produce the material itself.
For example, I input: "Create a commercial poster and a 15-second advertisement video for a perfume."
The Agent will deconstruct the task immediately. You don't need to worry about the details; just wait for the result.
The key feature here is contextual memory. After the first run, you can continue to provide feedback and adjust requirements. There's no need to re-describe the entire project background from scratch every time.
04
Video model upgrade: Agnes-Video-2.5-preview
The Agnes-Video-2.5-preview model has also been upgraded.
Once launched, the generation speed will increase by 50%.
The camera movement will be smoother, with fewer awkward cuts and jerks. The physical fidelity will improve, reducing floating objects, clipping, and unnatural collisions.
A new feature allows for multiple reference images, helping to maintain consistent character and product appearances.
Subtitle errors will be corrected. The video I generated just now used the 2.0 model, but the 2.5 version will be an improvement.
These upgrades are specifically aimed at streamlining the production process.
If your work involves frequent trial and error, such as ad targeting or short video matrices, the speed boost from the 2.5-preview will be particularly beneficial.
05
Community Feedback
I mentioned earlier that Agnes AI is offering unlimited free access to text, video, and image generation models.
However, the community has raised some concerns: - The video quality and speed are currently insufficient for commercial production, especially for higher-quality content. - There have been errors in processing prompts, such as incorrect aspect ratios. - The user interface is still in an early alpha stage.
But Agnes has made its GitHub Issues and Projects boards fully public, so you can track bugs, feature roadmaps, and update progress in real time.
You can view the progress board directly at: https://github.com/AgnesAI/agnes/projects
Other relevant links: - GitHub repository: https://github.com/AgnesAI/agnes - Discord community: https://discord.com/invite/agnes - WeChat official account: Search for "Agnes AI"
06
Click the card below to follow "Explore GitHub"
This official account has previously shared many interesting open-source projects. If you don't want to scroll through past articles, just follow the WeChat official account "Explore GitHub" and send a message for a chat.
I just mentioned that Agnes AI offers unlimited free text, video, and image model API calls.
But the community has also raised related issues:
However, Agnes has made the GitHub Issues and Projects boards public, so bug tracking, feature scheduling, and iteration progress are all visible.
To check progress, just look at the board:
其它相关地址:
06
Click the card below to follow "Guan Guan GitHub".
This public account has previously published many interesting open-source projects. If you don't want to browse articles one by one, just follow "Guan Guan GitHub" on WeChat and chat in the background dialogue.
