Cast your characters once as reference images, then generate vertical episodes where the same faces and outfits return scene after scene. Give them dialogue with a video model that makes its own audio, or add voices afterward with text to speech and lipsync. Fictional characters only, and short scenes work best.
A micro-drama is a serialized story cut into episodes short enough to watch in a spare minute, shot vertically for the phone and held together by cliffhangers and recurring faces. With AI video the hard part is never one good shot. It is getting the same woman in the same coat to show up in episode seven looking like herself. This page covers how to cast, voice and chain episodes on Picasso IA, and where the method stops.
A live-action micro-drama needs actors, locations, a sound person and a lot of scheduling, which is why most people with a good premise never film it. A text-to-video model removes the set. A reference-to-video model goes further: it takes pictures of your characters and uses them as the identity of the people on screen, so the cast is something you design once instead of something you hope the model reinvents correctly every time.
That makes the format a good fit for AI. Episodes are short, scenes are simple (two people, one room, one problem), and the audience forgives a lot when the plot moves. Picasso IA keeps the models you need in one place, with 502 models in the catalog, so the picture, the motion and the voice can all come from the same workspace. New accounts get free credits to test a first scene.
Continuity starts before any video exists. Use an image model such as nano-banana-pro or seedream-5-pro to design each character on a plain background: one clear face, the signature outfit, a neutral expression. Make a front view and a three-quarter view, and keep the outfit simple enough to describe in six words. A red wool coat survives from scene to scene. A jacket with a pattern of tiny birds does not.
Those images become your references. A model like wan-2.7-r2v takes reference images or short clips and anchors the subject to them, with vertical 9:16 output and a shot type setting for one subject or several. grok-imagine-r2v takes one to seven reference images and writes a clip from a short prompt. seedance-2.5 accepts up to 30 references and lets you point at them in the prompt by number. For a deeper look at the technique, see consistent AI characters, and for planning scenes before you spend credits, AI storyboards.
Dialogue is the decision that shapes the rest of your workflow. Each route below works, and each trades control for convenience differently.
| Route | How it works | Models to try | Where it fits |
|---|---|---|---|
| Native audio | The video model renders picture, speech and sound together from a prompt with the lines in quotes | seedance-2.5 | Short scenes where speed matters more than a perfect voice |
| Silent video plus a voice track | Generate the clip with references, then record or synthesize the lines and lay them over it | wan-2.7-r2v with speech-2.8-hd | Scenes where you want to choose the exact voice and timing |
| Talking portrait | One character photo plus a script becomes a speaking close-up | p-video-avatar | Confession shots, phone calls, direct-to-camera moments |
| Re-sync an existing clip | Keep the footage, swap in new audio, and re-animate the mouth to match | lipsync-precision | Fixing a line after the picture is already approved |
| Off-screen voices | Characters speak while the camera stays on a hand, a door or a window | grok-imagine-r2v with a text to speech model | Voice-over drama that avoids lip movement altogether |
Native audio is the quickest, because dialogue, effects and music arrive in one render, and seedance-2.5 can also lip-sync a character to a short audio clip you supply. The separate-voice routes cost more steps but give you a voice you chose on purpose, which matters when the same character has to sound identical in every episode. AI video dubbing covers the re-sync route in more detail.
This is the loop for one episode, from a blank page to a clip you can post. Repeat it for each scene and join the clips in any editor.
Video models do best with one clear action per clip, so write beats, not paragraphs. A scene that says she opens the letter, reads it, and looks up at the door is three beats and will probably need three clips. Give each clip a camera note (close-up, over the shoulder, slow push in) and a mood in a few words, and leave out anything the reference images already say, like hair color or coat.
The mistakes that break a series are almost always the same. People change the outfit between scenes, describe the face again in words and fight the reference, cram a whole conversation into one generation, or move the location every episode so nothing looks familiar. Keep a one-page cast and location sheet, reuse the same reference images every time, and set a seed where the model offers one, so a good result can be repeated.
Tip: Write the next episode's first scene before you generate the current one. A cliffhanger only works if you already know what it is hanging over, and a consistent opening shot gives the next episode a stable starting point to reuse.
Dialogue also lands better when it is short. A line a character can say in one breath is easier for any model, and for any lipsync pass, than a speech. Let the picture carry the rest, and use the pauses.
A reference image gets a character close, not identical, and a series gives drift many chances to show. Here are the limits that matter most when you plan a season:
None of this stops a short, well-planned series. It decides what you should not attempt: a feature-length plot with a dozen speaking roles will take far more regeneration than the format rewards.
Reasonably well, not perfectly. Reference-to-video models such as wan-2.7-r2v, grok-imagine-r2v and seedance-2.5 take your character images as the identity of the person on screen, which is far more stable than describing a face in words. Small changes in hair, clothing detail or age can still creep in, especially across episodes, so keep one reference sheet, reuse it every time, and regenerate any shot that drifts.
It depends on how you plan to handle dialogue. seedance-2.5 is the shortest path because it can render speech, sound effects and music with the picture and accepts up to 30 reference images. wan-2.7-r2v is a good fit when you want vertical output and a silent clip you voice yourself. grok-imagine-r2v suits quick scenes from one to seven reference images. Try the same scene in two of them.
There are two routes. Native audio means the video model creates the voice with the picture: in seedance-2.5 you put the line in quotes in the prompt. The other route is to generate a silent clip, then add a voice with a text to speech model and sync the mouth with a lipsync model such as p-video-avatar for a talking portrait or lipsync-precision for existing footage.
Choose 9:16, the standard for full-screen phone video. wan-2.7-r2v, grok-imagine-r2v and seedance-2.5 all offer it in their settings. Set it before you generate rather than cropping a widescreen clip afterward, because cropping throws away part of the frame and can cut off a character you framed carefully. Check the platform you publish on for its current recommendations.
It depends on the model and its settings. grok-imagine-r2v lets you set a clip length from one to ten seconds, and seedance-2.5 can render up to thirty seconds in one generation, while wan-2.7-r2v lets you set a duration at generation time. Either way, a micro-drama is built from several short clips cut together, so plan scenes in beats instead of long takes.
No, and you should not try. Use fictional characters you designed, or your own likeness. Putting a real identifiable person into a generated drama without their consent raises serious ethical and legal problems, and it is the fastest way to get a series removed. Invent your cast, give them original faces, and avoid prompts that name or imitate a real person.
A short outline is enough, but write the lines first. Models handle one clear action per clip, so a list of beats with a single line of dialogue each is the most useful form. Knowing the end of the episode before you start also helps you place the cliffhanger. A page of notes and a cast sheet will save you more credits than any prompt trick.
Generate each scene as its own clip, download them and join them in any video editor you like. Cutting between short clips also hides small differences between shots, which is one reason the format suits AI video. If you need music across the episode, you can generate a track separately and lay it under the cuts, and check that your dialogue levels match from clip to clip.
New accounts get free credits, which is enough to test a scene or two and see whether a model keeps your character stable. After that, generation costs credits, and the cost depends on the model and the settings you pick, such as resolution and duration. Video costs more than images, so build the cast with image models first. Current plans are on the pricing page.
Usually, but check each platform's current rules before you post. Several platforms ask you to label AI-generated or synthetic video, and their requirements change. Make sure you own or have rights to everything you upload as a reference, avoid real people and copyrighted characters, and read the terms of the platform where you plan to publish. If you are unsure about a rule, the support page can help with product questions.
Your cast is one prompt away from its first scene. Open the Picasso IA toolkit, design your lead character, and shoot episode one today.
Choose your language
