AI video
Which AI for video by task, what each one can do and how to write instructions so exactly what you want moves. Video models change fast, which is why the overview has a date.
🧭 Which AI for what
Based on what you want to do. The information is from the makers' help pages and announcements. Where it's only from articles, it says so.
For the images and photos you'll make a video from, see the AI images page. For texts and the script, Which AI for what.
🔢 Model overview
As of 24 October 2026. Length is the longest clip in one step; some can go longer by extending.
| Model | Company | Clip length | Sound | Where to find it |
|---|---|---|---|---|
| Gemini Omni 1.1 Flash | 10 s | Yes | The Gemini app, with a Google AI subscription | |
| Seedance 2.5 | ByteDance | Up to 30 s | Yes | CapCut, for subscribers |
| Kling 4.0 | Kuaishou | 3–30 s | Yes, stereo, dialogue too | kling.ai |
| Wan 3.0 | Alibaba | Up to 30 s (according to articles) | Yes | Alibaba Cloud, wan.video for members |
| MiniMax H3 | MiniMax | Up to 15 s | Yes, stereo | The Hailuo app, open weights |
| Utopai X | Utopai Studios | Not stated by the maker | Yes | The PAI platform |
| HeyGen Video | HeyGen | 5 s to try, longer in the app | Yes, speech too | HeyGen |
| Grok Imagine Video 1.5 | xAI | Up to 15 s | Yes (according to articles) | Grok, SuperGrok only |
| FLUX 3 | Black Forest Labs | Up to 20 s (according to articles) | Yes | API, preview |
| Runway Gen-4.5 | Runway | 2–10 s | Not stated by the maker | Runway |
| LTX-2.5 | Lightricks | Depends on the settings | Yes | At home on your own graphics card |
| Wan 2.2 | Alibaba | 5 s | — | At home, a 24 GB graphics card |
The Artificial Analysis leaderboard (video with sound, blind voting): 1. Wan 3.0, 2. Utopai X, 3. HeyGen Video, 4. MiniMax H3, 5. Seedance 2.5. Other leaderboards rank differently; on llm-stats, for example, Kling 3.0 was still in the lead. The new Kling 4.0 came out on 15 October 2026 and isn't in the leaderboards yet. Prices aren't here, they change the fastest.
⛔ What has ended or changed
- OpenAI's Sora has shut down. The app on 26 April 2026, the API on 24 September 2026. Videos you downloaded stay yours.
- Veo in the Gemini app has been replaced by Gemini Omni.
- Runway Gen-3 has ended; its successor is Gen-4.5.
- Kling has a new version 4.0 since 15 October 2026 (longer clips, 4K, more references).
- Luma Dream Machine keeps running with the Ray3 series models, but it isn't in the Artificial Analysis top fifteen in October 2026.
🎬 How to write a video prompt
- What's in the shot and how close (wide, medium, close-up).
- What moves and how fast. Write the character's movement and the camera's movement separately.
- Light and setting.
- What can be heard: sound effects, music, or explicitly “no music”.
- For a video from an image, describe only the movement. The look is already in the image; say that it should stay.
- Write several shots in order and with timings (0–4 s, 4–8 s…).
- Dialogue: for every line, write who says it and how, and in what language.
- Set the length and aspect ratio in the app when you can. Otherwise write them into the instructions.
📋 Prompts to copy
Replace the yellow parts in square brackets with your own. The labels show where the prompt works best.
Video [8 s], [16:9]. Shot: [a medium shot of an old man in the rain under a street lamp]. Action: [he slowly lights a cigarette, raindrops run down his face]. Camera: [a slow push-in towards the face]. Setting and light: [a street at night, neon reflections in puddles, light mist]. Sound: [rain, a distant car, no music]. Style: real video, natural movement.
Bring this photo to life. Keep the look, the face and the clothes exactly as in the photo. Character movement: [slowly opens her eyes and looks into the camera, blinks naturally]. Movement around: [the wind gently ruffles her hair, dust floats in the background]. Camera: [static / a very slow push-in]. Sound: [a quiet wind].
A [12 s] video from three shots, the same character in all of them: [a girl in a red jacket with a bike]. Shot 1 (0–4 s): [wide, the girl rides out of the garage, morning, fog]. The camera is static. Shot 2 (4–8 s): [a close-up of the pedals and the wheel on the wet road]. The camera moves alongside the bike. Shot 3 (8–12 s): [a medium-wide shot from behind, the girl disappears over the hill]. The camera slowly rises. Sound: [the clicking chain, drops, birds, no music].
A [10 s] video, [medium shot], two people at a kitchen table, evening, warm lamp light. 0–4 s: Mum (quietly, surprised, [in English]): “[You did this?]” 4–8 s: Boy (proudly, smiling, [in English]): “[All by myself.]” 8–10 s: the camera slowly pulls back, silence, only a ticking clock. Sound: a quiet kitchen, no music.
Kling speaks only some languages. Write the line in quotation marks in the language the character should speak. With other models, test your language on a short clip first.
Character: images 1, 2 and 3 are the same character [from the front, in profile and full-length]. Keep the face, hairstyle and clothes exactly. Clip: [the character walks into a café and looks around]. Camera: [follows her from behind, over the shoulder]. Light: [morning, warm]. Sound: [the ring of the door bell, the hum of the café].
Edit this video. Change only [the light and colours: turn the day into early evening with a sunset].
Keep the people's movement, the camera, the cuts, the length and the sound.
Don't add new objects, text or logos.
A product video [8 s], vertical 9:16, for [Reels]. Product: image 1, [a perfume bottle]. Keep the shape, colour and label exactly. Scene: [the bottle stands on a wet black stone, steam slowly rises around it]. Camera: [a slow arc around the bottle from left to right]. Light: [a dark studio, one soft source from behind, reflections on the glass]. Sound: [a quiet deep tone, a drop of water at the end]. No text or logos except the label on the product.
Image 1 is my character: [a girl in a red jacket]. Video 1 shows the movement I want: [a dance step / walking up stairs]. Transfer the movement from video 1 onto the character from image 1. From the character, keep the face, hairstyle and clothes exactly. From the video, take only the movement and the rhythm, not the look of the person in the video. Setting: [a rooftop at sunset]. Camera: [the same as in video 1 / static]. Sound: [no music, just footsteps].
Film the movement video yourself or use one you have the rights to. Don't use a character from someone else's photo without consent.
🧩 For advanced users: ComfyUI
When you want every step under control, or want to run models at home on your own graphics card.
- How to start: download the Comfy Desktop app for Windows and macOS at comfy.org. Without a strong graphics card, try Comfy Cloud; it's paid and has a free trial. Start from a ready-made template, not from an empty canvas.
- Graphics card: according to the authors, even big models run from 4 GB VRAM and 8 GB RAM, just more slowly. The official Wan 2.2 model card counts on 24 GB.
- Safety: community nodes (custom nodes) are someone else's code running on your computer. Install only ones with public source code and many users, not from links in discussions.
In ComfyUI I want to [make a short video from a photo with a slow camera push-in]. My computer: [Windows, an RTX 4070 graphics card with 12 GB, 32 GB RAM]. Design a workflow: which models to download (with a link to the official page), which nodes to connect and in what order, and what settings to start with. If it won't fit my graphics card, say so straight away and suggest a smaller model or Comfy Cloud. For every community node, say where it's from, so I download it only from a verified source.
My workflow in ComfyUI ends with an error. Here's the message from the console: <error> [paste the whole error message] </error> My computer: [Windows, RTX 4070 12 GB, 32 GB RAM]. Workflow: [a video from an image with the Wan 2.2 model]. Explain in one sentence what went wrong. Then give the steps to fix it, starting with the simplest option. If the cause can't be told from the message, say what else I should find out.
🎥 Camera movement
Terms that video models know. Use them in your prompt as they are.
- Static camera: static shot, locked-off camera - A pan left or right: slow pan from left to right - A push-in towards the character: slow dolly in on the face - A pull-back that reveals the surroundings: camera pulls back to reveal the city - Following the character: tracking shot, camera follows the subject from behind - An arc around the character: camera slowly orbits around the subject - A drone over the landscape: aerial drone shot over the forest, top-down view - Handheld, documentary style: handheld camera, slight shake, documentary style
🛡️ Real people and other people's content
- Someone else's face only with consent. Don't animate photos of real people without their consent. Seedance explicitly forbids unauthorised use of someone else's likeness, and authorities are investigating Grok over deepfakes.
- Age: Gemini Omni is 18+, Seedance in CapCut 16+.
- Don't put other people's music, logos and film characters into a video when you don't have the rights to them.