Build an n8n flow that writes a script with a language model, voices it, generates the clips and prepares shorts, all through your Picasso IA account. The generation steps run inside the flow. Joining the clips into one finished video is a step you add with another tool.
Most people try AI video by hand: a script in one tab, a voice tool in another, a video model in a third, then a pile of downloads that someone has to sort again tomorrow. n8n is a workflow automation tool, and it turns that routine into a flow that runs on its own. This page shows how one flow can write the script, voice it and generate the clips through Picasso IA, and which step you still finish elsewhere.
A single AI video is a craft task. Ten videos a week is a production line, and production lines are boring in a very specific way: the same five actions in the same order, with a small change of topic each time. That is exactly the shape of work an automation tool is built for. In n8n you describe the sequence once, as steps connected by lines, and the flow repeats it every time something starts it, a new row in a spreadsheet, a form submission, a schedule you set.
The Picasso IA connector for n8n is what lets the generation steps live inside that flow. It links your account to n8n with an API key you create in your account (treat it like a password), and each generation is one request that comes back as a file you can download or hand to the next step. It runs the same catalog as the toolkit, 502 models, so the flow can call a language model for the script, a text-to-speech model for the voice and a text-to-video model for the clips without you opening three different services.
What you gain is repetition without retyping. What you do not gain is judgment: the flow cannot tell whether the third clip looks wrong.
Think of the flow as four stages that hand work to each other. The table shows what each stage does, which kind of Picasso IA model it uses and who owns the result.
| Stage | What happens in the flow | Model type on Picasso IA | Who finishes it |
|---|---|---|---|
| Script | A language model turns a topic into narration and a list of scenes | Large language models (75 in the catalog) | You review the text |
| Voice | The narration is read aloud by a speech model | Text-to-speech (24 in the catalog) | You listen once |
| Clips | One short clip is generated per scene description | Text-to-video (121 in the catalog) | You discard the weak ones |
| Assembly | Clips, voice and captions are joined in order | Another tool you add to the flow | The other tool, then you |
The last row matters most. The connector covers the first three stages, the generation steps. Putting the clips in order on a timeline, laying the voice under them and exporting one finished file is a step you add with another tool, such as a video editing service you already use or a render step n8n can call. Plan for it from the start, because a flow with great clips and no assembly step ends as a folder of loose files.
Start in the toolkit, not in n8n. Every stage below can be tested by hand first, and a prompt that fails by hand will also fail inside a flow, only faster and more often.
Until the flow is stable, run it on a short topic with the cheapest model that gives acceptable results. Generations cost credits, the amount depends on the model and the settings, and the pricing page is where plans live.
A prompt typed by a person can lean on context in that person's head. A prompt inside a flow gets a topic and a template and nothing else, so it has to carry everything. Three habits make the difference.
First, make the script step return a fixed structure: the narration as plain paragraphs, then one scene description per line, each describing something a camera could actually film. Models are far better at following a rigid output shape than a vague request for something cinematic, and your loop step needs lines it can split without guessing.
Second, keep every clip prompt self-contained. Each clip is generated alone, so a line that says the same character as before means nothing to the video model. Repeat the subject, the setting and the style in every line. If a recurring character matters, the approach in consistent AI characters explains how reference images hold a face steady across clips.
Third, ask for scenes at the length the models handle well. Most video models produce short clips, so a narration of any length should be cut into many small scenes rather than one long request. AI explainer videos goes deeper on pacing a voice against pictures.
Tip: put the review gate right after the script step, not at the end. Reading a script takes a minute, while generating clips from a bad script costs credits and time you cannot get back.
A flow shines when the input is a list. Put twenty topics in a spreadsheet and the same flow produces twenty scripts, twenty voice tracks and the clips for each. A mistake in the template is now repeated twenty times, so run one row first and look at every output.
Shorts need a few extra steps after generation. The picture shape matters: vertical clips suit most short-video feeds, so choose that format when you generate rather than cropping later, or use the approach in resizing videos for every platform. Captions are worth adding because many people watch without sound, and automatic video subtitles covers that step. If your clips are cutaways for a longer piece, AI b-roll shows how to prompt for footage that supports narration instead of competing with it.
For faceless channels the same flow becomes a weekly routine, and faceless YouTube videos walks through the format. Whatever you publish, check the current requirements of each platform before you post.
Faces and voices need one more rule. Use characters you invented, or your own face and voice, never a real identifiable person you have no permission to depict.
A flow is a multiplier, and it multiplies weak spots as faithfully as strong ones. These are the places where it stops helping:
None of this is a reason to avoid automation. It is the checklist that decides where the human steps go.
Yes. It links your Picasso IA account to n8n so a flow can start a generation and receive the result as a file. It authenticates with an API key you create in your account, and it uses the same models and the same credit system as the toolkit. It covers the generation steps. Other parts of your flow, such as the trigger, the file storage or the video assembly, come from n8n and from the tools you connect to it.
Not for the shape of the flow. n8n is built around connecting steps visually, and each Picasso IA generation is one step with a model and a prompt. Code becomes useful for small jobs such as splitting a script into lines, and n8n lets you add it where you need it. If you can describe your process as a numbered list, you can build most of it.
You create an API key in your account and give it to the connector inside n8n. The key works like a password, so keep it in n8n's credential storage, do not paste it into a shared document and replace it if you think it leaked. While you wait to wire anything up, you can test every model and prompt by hand in the toolkit, and questions go to support.
The same catalog the toolkit runs, 502 models at the last count. For this kind of flow the useful categories are large language models (75) for the script, text-to-speech (24) for the voice and text-to-video (121) for the clips. Names worth testing include claude-sonnet-5 and gpt-5 for writing, speech-2.8-hd for narration, and veo-3.1, kling-v3-video, seedance-2.0 and wan-2.6-t2v for video. The all models page lists the rest.
Not from the generation steps alone. The connector produces the script, the voice and the individual clips. Joining them in order, laying the voice under the picture, adding captions and exporting one file is a separate step that you add to the flow with another tool. Many people start by letting the flow deliver a folder of clips and a voice file, then add the assembly step once the first part runs reliably.
Every generation in the flow costs credits, and the amount depends on the model and the settings you choose, so a run with a short clip on a light model costs less than one with several clips on a heavy model. A run is the sum of its steps: the script, the voice and every clip in the loop. Plans and credit sizes live on the pricing page. Testing a topic in the toolkit first shows you roughly what a scene costs before you loop it.
Yes, and it is the best habit in this whole workflow. The toolkit runs the same catalog, so you can try the script prompt, the voice settings and the video prompt by hand and keep the versions that work. A flow only repeats your decisions faster. If the prompt is weak, the flow produces weak clips at scale, so every stage should pass a manual test before it becomes a node.
It depends on the models and on how many clips the loop generates, and video is the slowest stage. Generation takes time, so the flow has to wait for each result before it moves on, and a script with many scenes takes longer than one with a few. Schedule the flow ahead of when you need the output.
Only your own, or fictional characters. Do not clone the voice or recreate the face of a real identifiable person without their permission, even if the models are capable of it. If you want a consistent presenter, create a fictional character and reuse the same reference image and voice settings across runs.
Better not, at least at first. Models can return clips with odd hands, faces or motion, and a script can contain a claim that needs checking, so a human review step before publishing protects your channel and your clients. Each platform also has its own requirements for AI content, which you should check on the platform itself because they change. Let the flow prepare the files and leave the final publish decision to a person.
Your first flow starts with one topic run by hand. Open the Picasso IA toolkit, test the script, the voice and a clip, and bring the settings that worked to n8n.
Choose your language
