Create a key in your account, name a video model in a request and download the clip when it is ready. One key reaches the text-to-video catalog, so comparing Veo, Kling, Seedance and Wan means changing a model name, not opening four accounts.
Video models arrive faster than any team can wire them up one by one: a separate account, a separate billing page and a separate way of asking for a clip each time. The Picasso IA API is built to collapse that into one key from your account, a request that names a model, and a file you download when the clip is ready. This page covers what that looks like for text to video, and where it stops.
Text to video is the category where choosing a model hurts most. Veo, Kling, Seedance and Wan all turn a written description into a moving clip, yet they interpret the same sentence differently, and none of them wins every shot. If your product generates video, the useful question is rarely which model is best. It is which model is best for this prompt, this budget and this deadline.
The Picasso IA catalog holds 502 models, and 121 of them sit in the text-to-video category. Five are named on this page: veo-3.1 from Google, kling-v3-video from kwaivgi, seedance-2.5 from bytedance, wan-3 from alibaba and ltx-2.5-fast from lightricks. With the API, one key from your account is the only credential involved. The request names the model, so trying a second one means editing a single value instead of integrating a second vendor.
The same models run in the toolkit today, which matters for a practical reason: you can settle the prompt and the model by hand first, then ask your code to repeat it. The API uses the same credit system as the toolkit, so there is no second balance to watch and nothing to reconcile between the two.
Worth knowing: a video request is a job, not an instant answer. Your code starts it, waits, and collects a file, so build the waiting into the design from day one instead of bolting it on after a timeout.
In plain words, a text to video request carries three things: which model should do the work, the prompt describing the clip, and the settings that model offers. Those settings are not identical across models. Some let you choose resolution and duration, some accept a seed so a good result can be repeated, some can produce audio with the picture, and some let you list what must not appear. wan-3 offers a negative prompt and a seed, for example, while ltx-2.5-fast can generate audio alongside the video. Read each model page in the catalog before you rely on a setting.
Video is asynchronous, and that shapes everything about your code. You start a job, the model works for as long as it needs, and when the clip is ready you get the file back. Short clips and heavy models do not take the same time, and the wait is not something your code can shorten, only plan around. A queue, a worker or a simple status check on a timer all work, as long as nothing in your product sits frozen waiting on a single clip.
Cost follows the same logic. Every generation costs credits, and how many depends on the model, the duration and the resolution you ask for. A longer, sharper clip from a heavier model costs more than a short draft from a lighter one. The pricing page lists the plans, and new accounts start with free credits, enough to run a few real prompts before you decide anything.
You will not read every model page before building. This table is a shortcut for deciding where to start, based on the job rather than on a ranking that would be stale in a month. Treat the last column as the thing to look at when you run the same prompt through two models side by side.
| Job | Models to try first | What to compare |
|---|---|---|
| Quick draft to test a prompt | ltx-2.5-fast | How close the clip gets to your idea against the time and credits spent |
| Polished hero clip | veo-3.1, kling-v3-video | Motion, detail and how faithfully each follows the prompt |
| A second opinion on the same prompt | seedance-2.5 | How its framing and movement differ from the first two |
| Longer clip or repeatable variations | wan-3 | The duration it allows and whether a fixed seed repeats the result |
| Many clips in a row | Any of them | Credits per clip at the duration and resolution you chose |
Run the same prompt through two or three of these before you commit a product to one. The differences show up in the clips, not in the descriptions. If you want a closer look at one family first, the comparison pages for Google Veo and Kling AI cover them in depth.
The journey is short on purpose. The craft sits in choosing the prompt and judging the result, not in the plumbing, so plan to loop through the last three steps more than once.
If your clips end up in ads or product pages, the video ads for social and product demo videos pages show what that finished work looks like. For a starting point that is a still image instead of a sentence, see turn photos into videos.
Most trouble with generated video comes from treating a job like a function call. These five habits are the difference between a pipeline you trust and one you babysit.
An API gives you repetition, not judgment. Everything a model gets wrong in the toolkit, it also gets wrong from code, only faster and in larger numbers, so the first limit is review: someone still has to watch the clips before they reach a customer.
Models differ, and so do their settings. A request that works on one model can need changes on another, because duration, resolution, audio and seed options are not the same everywhere. Switching is a small edit, but it is an edit you should test, not assume.
Generation takes time and costs credits. A clip is a job that can run for a while, a heavy model on a long, sharp clip costs noticeably more than a light one on a short draft, and a failed or discarded clip still used the attempt. Hands, text on screen, crowds and physics are the usual weak spots in any model, and prompting well reduces them without removing them.
Faces and voices need care. Use your own likeness, people who agreed, or fictional characters, never a real, identifiable person who did not consent. And this page describes what the API does, not more: it is not a video editor, a timeline or a way to guarantee a specific result. When you publish clips on a platform, check that platform's current requirements for format and AI-generated content.
It is a way to generate video from your own code instead of from the web app. You create a key in your account, send a request that names a model and gives a prompt, and the generated clip comes back as a file you can download. It uses the same models and the same credit system as the toolkit, so anything you can test there is the kind of thing a request can ask for.
The text-to-video category of the catalog has 121 models, including veo-3.1, kling-v3-video, seedance-2.5, wan-3 and ltx-2.5-fast. The intended coverage goes beyond video to the image, voice, music, upscale, background removal, 3D, lipsync and chat models of the same catalog. You pick the model by name in each request, and the full list lives on the all models page.
Yes, and it is the best first step. The toolkit runs the same catalog today: pick a model, write a prompt, generate and watch the clip. Settle which model and which wording work for your idea there, then ask your code to repeat it. New accounts get free credits, so a handful of test clips costs nothing to start with.
Because generating video takes real time, and the length of the wait depends on the model, the duration and the resolution. So the request starts a job, and the file arrives when the job is done. Design for that: start the job, keep your app responsive while it runs, and fetch the result once it is ready. Never block a user screen on a single clip.
It costs credits, and the amount depends on the model and the settings. A longer clip, a higher resolution or a heavier model costs more than a short draft on a light one. This page does not quote numbers because they change over time. The pricing page lists the current plans, and running a few clips in the toolkit shows you what your own settings cost.
Treat it like a password. Keep it on your server as a secret, never inside front-end code, a public repository or a screenshot, and never share it in a chat. Anyone holding the key can spend your credits. If you think it leaked, create a new one in your account and stop using the old one, then check your credit balance for activity you do not recognize.
The request names the model, so switching starts with changing that name. The catch is that settings differ between models: duration, resolution, audio and seed options are not the same on every one. Keep your model-specific settings in one place, change the name, run the clip and compare. A short test per model is faster than finding a mismatch in production.
Only with your own likeness or with people who clearly agreed, and fictional characters are always fine. Do not generate real, identifiable people who did not consent, and do not present a generated clip as a recording of something that happened. If a clip could mislead viewers about what is real, label it as AI-generated and check the platform's current rules before posting.
Often yes, but check two things first. Look at the rights to anything you feed in, such as an image or a brand asset, and at the terms attached to your plan for generated outputs. Then have a human watch every clip before it goes out, since text on screen, hands and physical details are where video models slip most often.
Start by rerunning the same prompt once, since results vary between runs. If the clip is close but off, narrow the prompt to one subject, one action and one camera move, or try another model, which is just a change of name. If a job fails repeatedly or something looks broken on our side, contact support with the model, the prompt and the time it happened.
Ready to see the clips before you write a line of code? Open the toolkit, run one prompt through two models and decide which one your request should name.
Choose your language
