How Does AI Video Generation Work? Plain Steps
AI video generation works by starting from random noise and cleaning it step by step until frames appear that match your text or photo, guided by a model trained on many videos. Wan2.1's GitHub page names the parts: a text encoder reads the prompt, a video autoencoder packs frames into a compact form, and a diffusion transformer does the step-by-step cleaning.
- 100
- free credits at sign-up
- $0
- subscription needed
The four steps, using Wan2.1 as the example
1. Read the prompt: a text encoder (T5 in Wan2.1) turns your words into numbers the model understands. 2. Work in a small space: a video autoencoder (Wan-VAE, a 3D causal VAE) packs the frames into a compact form, so the model handles far less data. 3. Clean the noise: a diffusion transformer, trained with Flow Matching in Wan2.1, removes noise in many small steps, checking against the prompt each time. 4. Decode: the autoencoder turns the result back into full video frames.
Read from each site's own page on 21 October 2026: Wan2.1 on GitHub.
Why photos and audio change the result
A reference photo gives the model a fixed starting point, so the face or product stays close to the picture. Reference video gives it motion to follow, and reference audio gives it timing. Reviral's Cinema studio takes all three, depending on the model.
Trying it on Reviral
Cinema sends your prompt to the model word for word. A 5-second Seedance 2.5 clip costs 115 credits (~$1.43), and the studio says clips are ready in minutes.
Live price
What it costs on Reviral, live price
Dollar prices are at the $4.99 Mini Pack rate ($4.99 for 400 credits).
Reviral · live price
Read from Reviral's live prices each time this page loads.
| Clip | Credits | Dollars |
|---|---|---|
| Seedance 2.5 clip, 5 s at 720p | 115 credits | $1.43 |
| Seedance 2.5 clip, 30 s at 720p | 690 credits | $8.61 |
New accounts get 100 free credits, no card needed. The exact price is on the button before you render.
Frequently asked questions
How long does AI video generation take?+
It depends on the model and the hardware. Wan2.1's GitHub page says its 1.3B model makes 5 seconds of 480P video in about 4 minutes on an RTX 4090; on Reviral's hosted models, clips are ready in minutes.
Does AI video generation copy existing videos?+
No. The model learns patterns of motion and appearance from training videos, then builds new frames from noise for your prompt.
Why do AI videos come out different each time?+
Each render starts from different random noise, so the same prompt can give different shots. Reference photos narrow the result.
Price shown before you render
Generate a clip
New accounts get 100 free credits, no card needed. Credit packs never expire.
Generate a clipBy Wilson Sik Founder, Reviral Last updated