Write a fictional scary story, have it read in a low, slow voice, then generate the shots and the sound bed that make it feel like a real production. Picasso IA puts the voice models, the video models and the sound design models in one toolkit, so the whole thing happens in a single place.
A horror story works on the ear before it works on the eye. A low voice, a pause that lasts one second too long, a floorboard creaking somewhere in the house: that is most of the fear, and none of it needs a film crew. With AI you can write a fictional tale, narrate it in a slow whispered voice, generate shots that match every beat and lay a sound bed underneath, all from one workspace.
Scary story channels have always been a production problem disguised as a writing problem. The story is the easy part. The hard parts are finding a narrator who sounds unsettling at three in the morning, sourcing footage that fits a haunted hallway you invented, and building a soundscape that does not sound like a stock library from 2009. Picasso IA runs 502 models behind one interface, including 24 text-to-speech models and 121 text-to-video models, so the voice, the picture and the sound can all come from the same place; the full catalog lists every one. New accounts get free credits, enough to find out whether a story holds up on screen before you spend anything on it. Horror also forgives the weak spots of AI video better than most genres do. Dim light, slow movement, shadows and fog hide the small glitches that would ruin a cheerful product ad, and a story told mostly through voice only needs the image to suggest, never to show everything.
One rule shapes everything on this page: keep it fictional and keep it atmospheric. The fear comes from what the listener imagines, not from graphic content.
Start with the narration, not the footage. In a horror video the voice carries the pacing, and every shot you generate later exists to serve a sentence you have already written. Short sentences work best. A line that ends in a pause gives you a natural cut point, and a pause is where dread lives.
For the voice, two models in the catalog suit this job in different ways. elevenlabs/v3 takes your text and exposes sliders for speed, style exaggeration and stability, so you can pull the delivery down to a slow, controlled read and dial the theatrics up or down. inworld/tts-1.5-max reads emotion tags and break tags written straight into the script, which lets you mark a sigh, a whispered line or a full second of silence exactly where you want it.
Write for the ear, then slow it down: set the speed a notch below normal, add a pause before every reveal, and cut the adjectives. A calm voice describing something wrong in a plain sentence is scarier than a voice that is acting scared.
Keep each generation to a scene or a paragraph rather than the whole story. Long scripts drift in tone, and splitting at the scene breaks gives you clean places to change the shot, the room or the sound. Pick a preset voice or a recording you made yourself. Never imitate a real person's voice, and keep the narrator, like the story, fictional.
There is more than one way to put a horror video together, and the right one depends on how long the story is and how much control you want over the sound. This table compares the five that work in practice.
| Approach | Best for | Models to start with | Watch out for |
|---|---|---|---|
| Video with sound generated in the same render | Short scares and cold opens, a few seconds to half a minute | google/veo-3.1, bytedance/seedance-2.5 | The sound is the model's choice, so a precise whisper is hard to direct |
| Separate narration over silent visuals | Full spoken stories of several minutes | elevenlabs/v3 or inworld/tts-1.5-max plus a text-to-video model | You assemble and time the pieces in an editor yourself |
| Sound bed added after the clip | Rescuing a silent clip that looks good | mirelo/video-to-sfx-v1.5, zsxkib/mmaudio | The SFX model trims input to ten seconds, so work clip by clip |
| Still images with slow push-ins | Creepypasta narration where the voice is the star | A text-to-image model, then a video model for gentle motion | Less movement, so the pacing must come from the script |
| Locked reference character | A recurring figure across scenes | bytedance/seedance-2.5 or google/veo-3.1 with reference images | Fictional characters only, never a real person's likeness |
Most finished horror videos mix two of these: a narrated story built from separate voice and picture, plus a few native-audio clips for moments where a sound should land with the image, such as a door opening.
This is the workflow from a blank page to a video with a voice, shots and a sound bed.
If you want the voice layer explained on its own, AI voiceover for videos covers it in more depth, and adding sound effects to silent AI videos does the same for the sound bed.
The difference between a creepy video and a cheap one is mostly restraint. These habits come up again and again when the result works.
The weak point is continuity. A house you describe in scene one will look different in scene three unless you anchor it with reference images, and even then a hallway can change length between cuts. Plan the story so the audience sees fragments, a hand on a doorknob, a window, a corner of a room, rather than a floor plan they might compare from shot to shot.
Faces and hands remain the other tell. Close shots of a human face in motion are where generated video looks most artificial, and that is a problem in a genre that relies on realism for its effect. Keep people small in the frame, in silhouette or turned away, which is also the classic way to frighten.
Sound is the third limit. The native audio in a video model is chosen for the scene, not for your story, and a text prompt only steers it loosely. For a whisper that must hit a particular word, generate the voice separately and sync it yourself. Finally, a video model does not know what a platform will accept. Rules about scary or graphic content and about labeling AI-generated media differ between services and change over time, so check each platform's current requirements before you upload, and keep the content fictional and non-graphic.
New accounts on Picasso IA get free credits, which is enough to try a narration and a few short clips and see whether your story works on screen. Generation costs credits, and the amount depends on the model and the settings you pick, so check the pricing page for the current plans instead of trusting a number written here. Start with one scene before you commit to the whole story.
There is no single answer, which is why trying two helps. elevenlabs/v3 offers speed, style and stability controls, so it suits a slow, controlled read, while inworld/tts-1.5-max accepts emotion tags and break tags inside the text, so it suits scripts where you mark every pause by hand. Run the same paragraph through both, slow the speed slightly and keep the voice that sounds calm rather than dramatic.
Lower the speed setting a little below normal, write short sentences and put a pause before each reveal. With elevenlabs/v3 you can also lower the style exaggeration so the voice sounds composed instead of theatrical, and with inworld/tts-1.5-max you can add a sigh or a break tag at the point you want silence.
Some do. google/veo-3.1 and bytedance/seedance-2.5 can generate audio in the same render, so a clip arrives with ambient sound already in it. The sound is the model's interpretation of the scene, though, so for a narrated story most people generate the voice separately and use the built-in audio only for short moments. For a clip that came back silent, a sound model can add a matching track afterwards.
Upload the silent clip to mirelo/video-to-sfx-v1.5 or zsxkib/mmaudio and describe what you want to hear in a few words. The Mirelo model returns several variations to choose from, and MMAudio accepts a negative prompt so you can keep music out. The Mirelo model trims video to ten seconds, so a long video is handled clip by clip, and you can mix the results in your editor.
Yes, with reference images. Seedance 2.5 accepts many reference images and Veo 3.1 accepts up to three, which helps a character or a room stay recognizable from one clip to the next. It will not be perfect, so favor shots that show a silhouette, a hand or the back of a head. Use fictional characters only, never the face of a real person you have not got permission to depict.
Each generation is short. Veo 3.1 clips run four to eight seconds, and Seedance 2.5 can produce clips of up to thirty seconds in one render. A full story is built by generating clips scene by scene and laying a narration track over them in an editor.
Fictional, non-graphic horror is common on both, but each platform has its own rules about scary content, graphic content and how AI-generated media must be disclosed, and those rules change. Read the current requirements of the platform you post on before uploading. Keeping the story atmospheric and implied rather than gory is the safest choice, and it usually makes the video better too.
It is better not to. Faces, voices and names of real people should not be generated without their permission, and turning a real tragedy into entertainment can hurt the people involved. Stick to invented characters, invented places and original plots. Folklore that no one owns is a fine source if you retell it in your own words.
Regenerate it with a gentler prompt. Describe the setting and the mood instead of the threat, add words such as slow, dim and empty, and use a negative prompt where the model allows one. If a face or hand still looks off, change the shot to a silhouette or a wide frame. If you are unsure how something in the product works, the support page can answer it.
Your story is already half a video once the voice is slow enough. Open the Picasso IA toolkit, narrate the first scene and see what the hallway looks like.
Choose your language
