Describe the video you want in a chat and the agent writes the script, records the voice, renders the shots and composes the music, picking the models from the catalog for each stage. You review every piece and steer in the same conversation, instead of moving files between four tools.
Making a narrated video with music by hand means chaining four kinds of tool: a language model for the script, a voice model for the narration, a video model for the shots and a music model for the track, each with its own prompt and its own export. Supercomputer, the agent in Picasso IA, does the chaining. You write one brief in a chat, it plans the job, runs the models and hands you the pieces.
A narrated video with a soundtrack is not one model's job. It is a script from a language model, a voice from a text-to-speech model, shots from a video model and a track from a music model, and between every pair sits a small chore: copy the script, split it into scenes, write a video prompt for each scene, name the files so you can find them again. None of it is hard. All of it is where an afternoon quietly disappears.
Supercomputer takes that chaining over. It is a chat, not a workspace: you describe the result you want in plain words, it plans the steps, picks models from the catalog, renders them and shows you what came back. You steer in the same conversation. The catalog behind it is the one the toolkit runs, 502 models today, of which 121 are text-to-video, 75 are large language models, 24 are text-to-speech and 10 generate music. The full catalog is browsable if you want to see what the agent has to choose from.
It is a separate product from the toolkit, not a replacement for it. The toolkit is the manual workspace where you pick the model yourself and every setting is in your hands. The agent is for the days you would rather state the outcome and review it. Same models, same credits, a different seat at the table.
Behind a single sentence from you there is a short pipeline. The table shows the stages the agent walks through, the kind of model that does the work (the names are examples from the catalog, not a fixed list) and what each stage hands to the next.
| Stage | Kind of model | What comes out | What you can change afterward |
|---|---|---|---|
| Script | A large language model, for example claude-sonnet-5 | Narration text split into scenes | Tone, length, the wording of any line |
| Voice | A text-to-speech model, for example speech-2.8-hd | The narration as an audio file | The voice, the pace, a re-read of one scene |
| Shots | A text-to-video model, for example veo-3.1 or seedance-2.0 | One clip per scene | The prompt for a single shot, or another model for it |
| Music | A music model, for example lyria-3 | A background track | Mood, instruments, how loud it sits under the voice |
| Hand-off | The agent itself | The pieces gathered in the chat, ready to download | Any stage, rerun on its own |
The order matters because each stage feeds the next. The script decides how many scenes exist, the scenes decide how many shots are needed, and the narration length tells you roughly how long the picture has to run. That is why a change early in the chain, a shorter script for instance, ripples into every later stage, and why it is worth settling the script first.
The agent can only plan what you tell it, so the brief is the part worth your effort. A good one answers six questions in a few sentences: what the video is about, who it is for, how long it should feel, what tone the narration takes, what the pictures should look like, and what the music should do. Skip the technical vocabulary. You do not name models or settings, the agent chooses those.
Here is the difference in practice. A thin brief says make a video about coffee. A workable one says a calm, thirty second explainer for people who have only ever used pods, narrated in a warm and unhurried voice, with slow close shots of a pour-over and a quiet acoustic track. The second version gives the language model a tone, the voice model a character, the video model a visual direction and the music model a mood, all from one paragraph.
Two details save credits later. State the language of the narration, because the voice and the script both depend on it. And say what must not appear: no text on screen, no people, no brand marks. Video models invent freely, and a sentence of exclusions is cheaper than rerunning a shot. If your idea involves a person, use a fictional character or your own face and voice, never a real identifiable person, and never a voice you do not have the right to use.
This is the whole journey in plain words. If you want to hear a voice or watch a shot before committing a full run, any single stage can be tested by hand in the toolkit first, since it runs the same catalog.
The first pass is a draft, and the chat is where it becomes a video you are happy with. You talk to the agent the way you would talk to an editor: make the narration shorter, use a lower voice, redo the third shot with slower movement, give the music less percussion. It reruns the stage you pointed at and leaves the rest alone, so a fix to one shot does not cost you a new script.
Change one thing per message. If you ask for a new voice, a warmer script and a different visual style in a single request, you will not know which change fixed the video or broke it, and every rerun is another generation that spends credits.
Because the toolkit sits on the same catalog, the two products work well together. A good way to use them is to let the agent produce a first version, then take a single weak shot into the toolkit and try it against two or three video models by hand. Pages such as AI voiceover for videos and background music for videos go deeper on the individual stages, and faceless YouTube videos shows the kind of finished project this chain is built for. Credit use depends on the models and settings involved, and the pricing page is where plans live.
An agent that chains models inherits the limits of every model in the chain, and it adds one or two of its own. Five places to plan around:
The agent is a first-draft machine with a short feedback loop, not a director. If your video depends on a precise performance, a specific real location or a tightly timed edit, the toolkit and a video editor give you more control.
It is Supercomputer, the agent in Picasso IA, used for video. You describe the video in a chat, and it plans the job, chooses models from the catalog for the script, the voice, the shots and the music, renders each stage and shows you the result. You then steer in the same conversation. It is a chat inside the web app, so there is nothing to install.
The toolkit is a manual workspace: you pick the model, write the prompt, set the options and run one generation at a time. The agent works from the outcome instead. You say what you want, and it decides which models to chain and in what order. Both use the same catalog and the same credits, so the choice is about how much control you want. Many people use the agent for a first version and the toolkit for fine-tuning one piece.
No. Supercomputer is a chat in the web app, so you sign in to your Picasso IA account and write your brief. Keys belong to the API, the MCP server and the connectors, where you create one in your account and treat it like a password. For the agent there is nothing to paste or store.
It picks from the catalog, which holds 502 models today, including 121 for text-to-video, 75 large language models, 24 for text-to-speech and 10 for music. Names such as claude-sonnet-5, speech-2.8-hd, veo-3.1, seedance-2.0 and lyria-3 are examples of what sits there, not a promise that a given run uses them. You can ask for a particular model or style in the conversation and see what changes.
A generation costs credits, and the amount depends on the models and the settings, so a video made of several stages costs more than one image. Any number written here would be wrong soon, so there is none. New accounts get free credits to try things, and the plans are listed on the pricing page. Testing a single stage in the toolkit first is a cheap way to see what a model costs.
Yes, by saying so in the conversation. You can ask for a lower voice, a particular kind of music or a different video model for one shot, and the agent reruns that part. If you want to compare models side by side with every setting visible, the toolkit is the better place, because it lets you run the same prompt through several models and judge the results yourself.
That depends on the video models in the chain, because each shot is generated as its own clip and clips are short compared with a full film. Longer videos are built from many shots, which is more generations and more credits. State the length you are aiming for in the brief and check the clips that come back.
Your own, yes. Fictional characters, yes. A real identifiable person who has not agreed, no, and the same goes for a voice you have no right to use. This matters most for anything you plan to publish or sell. Say in the brief that the character is fictional or that the face and voice are yours, and keep proof of your own permission if the video will be used commercially.
Usually a little. The agent gives you the script, the narration, the clips and the music as matched pieces, and joining them into one timeline with the right timing is a normal final step in a video editor you like. Some people find the first version is close enough for a draft and some rework a few shots. Plan for review, because models differ and a quick listen and watch catches most problems.
Often yes, but it depends on the terms of the models behind each piece and on the rules of the place you publish, which change over time. Check the current requirements of the platform before you post, and avoid real people, brand marks and protected characters in the brief. If you are unsure about a specific case, ask support before you build a campaign on it.
Want to hear a script read aloud or watch one shot before you write the whole brief? Open the toolkit and test any model from the catalog now, and send questions to support.
Choose your language
