Paste the script you already wrote and get a narrated module back: an avatar presenter reading it on screen, or a voice over slides you illustrate with image models. You still own the teaching and the facts. The models only handle delivery, so you can review, fix and republish a lesson in an afternoon.
You already wrote the lesson. What stands between that script and a finished course module is usually the part nobody enjoys: finding a presenter, recording clean audio, and building slides that match what you say. With Picasso IA you can hand the script to an avatar presenter, or to a narrator voice over illustrated slides, and get a module you can review, fix and publish.
Most course projects stall in production, not in teaching. The outline exists, the lesson notes are good, and then the recording sessions slip, the microphone picks up the neighbor's dog, and one wrong sentence in minute six means re-recording the whole take. When the script is the source of truth, a correction is a text edit and a new generation, not a new afternoon in front of a camera.
That changes how you can work. You can generate a rough version of every module first, watch the whole course in order, and spot the lesson that repeats an earlier one before you have invested in polish. You can also keep one master script per lesson and produce a second language later from the same text. Instructors who dislike being on camera get a presenter without becoming the presenter.
Picasso IA puts the pieces in one workspace, the toolkit: avatar video models, text-to-speech models and image models, all of them taking the same script as raw material. The catalog holds 502 models in total, with 24 in text-to-speech and 121 in text-to-video, and you can browse them on the all models page. New accounts get free credits, so a single test module costs you nothing to try.
There are two honest ways to turn a script into video, and they suit different lessons. An avatar presenter gives the learner a face to follow, which helps with welcome videos, module intros and recaps. Illustrated slides with narration carry concepts better, because the picture can show the thing being explained while the voice stays out of its way. Many courses use both: an avatar to open and close each module, slides in the middle.
| Format | Models to start with | Suits | Watch for |
|---|---|---|---|
| Avatar presenter, full frame | heygen/avatar-v | Welcome videos, module intros, recaps | Long stretches of talking head tire viewers, so keep these short |
| Avatar presenter, circle overlay | heygen/avatar-iv | Lessons where slides carry the content | The circle covers a corner of the slide, so keep text away from it |
| Illustrated slides with narration | google/nano-banana-pro or bytedance/seedream-5-pro plus elevenlabs/v3 | Concepts, diagrams, step-by-step lessons | You assemble pictures and audio in an editor you already use |
| Narration only over a screen recording | google/gemini-3.1-flash-tts | Software walkthroughs you record yourself | The voice must match the pace of your clicks, so time the script to the recording |
| Another language for the same lesson | elevenlabs/v3 or google/gemini-3.1-flash-tts with a language code | Reaching a second audience from one script | A fluent reviewer has to listen before release |
The avatar models read your text with a voice from HeyGen and take an avatar ID and a voice ID, which you pick from HeyGen's own lists. They do not accept a finished audio file, so a narration made with a separate voice model belongs to the slides route, not the avatar route.
This walks one lesson from script to a file you can upload. Run it once on your shortest lesson before you commit to the whole course.
A script written for reading and a script written for listening are different documents. Sentences that look fine on a page, with three clauses and a parenthesis, come out of a synthetic voice as a breathless run. Write shorter, say a number or term once and repeat it later instead of pointing back with "as mentioned above", and read the block aloud yourself before you generate anything.
Pronunciation is the first thing learners notice. If your lesson contains a product name, a surname or a technical term, spell it phonetically in the script, run a ten-second test, and only then generate the full block.
A few habits save regenerations. Put the learning goal in the first sentence of each block, end on the single thing to remember, and write numbers and abbreviations the way they should be spoken. For Gemini 3.1 Flash TTS, a plain-language style instruction such as a calm, unhurried teaching tone steers the pace, and for elevenlabs/v3 the speed, stability and style settings do the same job.
One good module is easy. Twenty modules that feel like the same course is the real work, and AI makes drift easy because every generation is a fresh roll. Fix the variables once and write them down: the avatar ID and voice ID, the narrator voice and its settings, the style sentence for slides, the aspect ratio and the resolution. Treat that list as part of the course template.
For slide art, the same trick used for consistent AI characters applies: keep a reference image of the first slide you like and edit from it, rather than describing the style from scratch every time. Naming your files by module and lesson number saves more time than any model setting.
Neighboring pages help here: AI training videos for teams covers internal onboarding, and Picasso IA vs HeyGen compares this setup with a dedicated avatar platform.
An avatar delivers a lesson. It does not understand the lesson, and it will read a wrong sentence as confidently as a right one. These are the places the format stops working:
None of this rules the method out. It tells you where your own time still belongs: designing the lesson, checking the facts and deciding what a learner should be able to do afterwards.
Yes, in the sense that it can deliver the script. An avatar model such as heygen/avatar-v reads your text on screen with a presenter, and a voice model such as elevenlabs/v3 can narrate the same text over slides you make with an image model. What it cannot do is write a good lesson or check that it is correct. Bring a finished script and treat the generated video as a draft you review, the way you would review a recording from a human presenter.
heygen/avatar-v and heygen/avatar-iv are the two avatar models in the catalog for this. Avatar V offers 720p, 1080p and 4K output in 16:9 or 9:16, plus burned-in captions. Avatar IV offers voice emotion options and three display styles, including a circle overlay. Both need an avatar ID and a voice ID from HeyGen, and both take a script of under 5,000 characters per run.
It depends on the voice, the language and the lesson, so listen before you commit. elevenlabs/v3 has a range of voices and settings for speed, stability and style, which helps keep a long course sounding like one narrator. google/gemini-3.1-flash-tts has 30 voices, many language codes and a plain-language style prompt. Generate the same two paragraphs with both, and pick the one a learner could listen to for twenty minutes without noticing the voice.
Your own likeness is fine to use if you have the right setup with the avatar provider, and a fictional or stock avatar is the simplest option. What you should not do is build a presenter from someone else without their clear permission, including a colleague, a famous teacher or a former employee. The same goes for voices. If you are unsure, use a stock avatar and voice.
Describe each slide to an image model like google/nano-banana-pro or bytedance/seedream-5-pro, using one fixed style sentence so every slide matches. Ask for the picture only, with space left for your own text, and add titles, labels and formulas in your slide or video editor. That keeps the words correct and editable.
Each model has its own limit per run. The avatar models take a script of under 5,000 characters, and Gemini 3.1 Flash TTS takes about 4,000 bytes of text. In practice you will generate a module as several short blocks anyway, which is easier to review and cheaper to fix when one sentence is wrong. Join the finished blocks in your video editor.
You can generate a second version from the same script, using a translated script with a voice model that supports the language, or an avatar voice for it. Gemini 3.1 Flash TTS accepts many language codes, and elevenlabs/v3 takes a language code too. The translation itself still needs a fluent human to check terminology and tone.
Check the current requirements of the platform where you publish. Course marketplaces, video sites and employers each have their own rules about synthetic presenters and voices, and those rules change. Beyond the rules, a short line in the course introduction saying the presenter is generated is simple honesty.
Only if you fix the variables. Use one avatar ID and voice ID, one narrator voice with the same settings, one style sentence for slides, and one resolution and aspect ratio across the course, and keep them in a short template document. Even then, every generation varies slightly, so watch the modules in sequence before release and regenerate any block that sticks out.
Generations run on credits, and the cost depends on the model and the settings you choose, with video costing more than audio or images. New accounts get free credits, enough to test one module. Because costs change, the pricing page is the only place to trust for numbers, so open the pricing page before you plan a full course, and test your shortest lesson first to see what one block uses.
You have the script already. Open the Picasso IA toolkit, run your shortest lesson through a presenter or a voice, and watch what one finished module actually looks like.
Choose your language
