How the AI video generator works

Pick a trend, upload a photo, press generate — and about a minute later you have a short vertical clip ready to post. Here is what actually happens between those two clicks.

One photo is the material

Every template on this site starts from a face. The generator keeps the person in your photo — the hair, the expression, the outfit — and animates them inside a preset motion script: a kiss-cam turn, a sand-storm walk, a neon-city swagger.

The video never copies the example clip's costume onto you. The example only shows the motion; your photo supplies the look.

The mechanism is documented in the research these models build on. Diffusion models synthesize an image by reversing a noise process inside a compressed latent space (Rombach et al., 2022), and adding a control signal — a pose or an edge map — lets the model keep the subject's structure while changing what it does (Zhang et al., 2023). Extending that idea from stills to motion adds a temporal layer so consecutive frames stay consistent (Blattmann et al., 2023).

"…turn the image generator into a video generator by introducing a temporal dimension."

— Blattmann et al., 2023, Align your Latents

What happens when you press generate

1. The price is fixed before you start. Each template lists its credit price next to the button — video, GIF, or image. The number you see is the number charged; lookups happen on the server, so the label cannot drift from the price.
2. Your photo is uploaded over an encrypted connection and stored with your account, so you can download the result later from your settings.
3. The render runs on production AI models. The engine picks the model that fits the mode — video for motion, GIF loops for reactions, image models for style pieces — then streams its output back as MP4 or GIF. Current commercial video models such as Veo run on the same diffusion-plus-temporal-layers approach described above.
4. Credits are consumed exactly once. If a render fails on the provider's side, the credits come back automatically.

Why the outputs are silent

Generated clips ship without audio on purpose. They are built for TikTok, Reels and Shorts, where almost everyone posts over a licensed sound. You pick the music in the app when you publish — the app's sound libraries are there for exactly this.

What you can change

Before generating you can pick the aspect ratio (9:16 for shorts, 16:9 for horizontal), the resolution, and — on video templates — the clip length. Everything else is preset so the trend still looks like the trend.

That is the whole loop: photo in, clip out, credits spent, done.

Sources

• Rombach, R. et al. (2022). High-Resolution Image Synthesis with Latent Diffusion Models. CVPR.
• Zhang, L., Rao, A., Agrawal, M. (2023). Adding Conditional Control to Text-to-Image Diffusion Models.
• Blattmann, A. et al. (2023). Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. CVPR.
• Google DeepMind. Veo — video generation model.