Anyone arriving at video from image generation gets a shock. The same card that returns an image in two seconds takes six to ten minutes to produce five seconds of video. This is not the same problem at a larger scale: it is a different problem.
📚 Series "Build Your Own ChatGPT" — 9 parts
1. What's inside · 2. The brain · 3. The eyes (OCR) · 4. The memory (documents) · 5. Images · 6. Video ← you are here · 7. Voice · 8. Tools and agents · 9. Putting it together
Why video is a different league
Five seconds at 24 frames per second is 120 images. If that were all, it would be 120 times the cost of an image — a lot, but manageable.
Except it is not enough to generate 120 images: they have to be the same world. The jacket cannot change colour at frame 40, the camera has to move coherently, the hand that started rising must keep rising. To do that the model processes frames jointly, with temporal attention between them — and that part consumes memory and time disproportionately.
| Target | Typical generation time | Minimum card |
|---|---|---|
| 5 s, 480p, 24 fps | ~2 to 4 min | 24 GB (R$ 2.78/h) |
| 5 s, 720p, 24 fps | ~6 to 10 min | 48 GB (R$ 7.08/h) |
| 10 s, 720p, 24 fps | ~15 to 25 min | 48 GB (R$ 7.08/h) |
| 5 s, 1080p directly | rarely worth it | generate at 720p and upscale |
August 2026 prices, single GPU, no scheduled interruption. Prices are in Brazilian reais (BRL), the billing currency. The catalogue moves — check the live one before committing to a number.
The number that puts everything in perspective
A 5-second clip at 720p, eight minutes on a card at R$ 7.08 per hour, costs R$ 0.94 — about R$ 0.19 per second of video.
Against the other pieces of this series, using the same numbers:
| Piece | One unit costs | Equivalent to |
|---|---|---|
| A text message | R$ 0.0025 | 1 message |
| A generated image | R$ 0.0020 | ~1 message |
| A page of OCR + structuring | R$ 0.00087 | less than 1 message |
| 5 seconds of video | R$ 0.94 | ~376 messages |
This is why no large platform offers unlimited video, and why they all count how many you generated. The piece does not survive free usage.
The pipeline that works: image first
The naive approach asks for video straight from text. It works, and it is the most expensive and least controllable path: you wait eight minutes to discover the framing was wrong, adjust one word, and wait another eight.
What studios do instead:
- The brain writes the script and describes each scene. Costs cents, takes seconds.
- Generate the key frame for each scene (part 5). Two seconds and R$ 0.002 per attempt — you iterate until the composition is right, essentially for free.
- Animate from the approved frame (image-to-video, with Wan 2.2). Far more predictable than text-to-video, because half the visual decisions are already made.
- Generate the narration (part 7) and assemble with ordinary video tooling.
💡 Why the order matters so much
Iterating on an image costs R$ 0.002 and 2 seconds. Iterating on video costs R$ 0.94 and 8 minutes — nearly 500 times more expensive and 240 times slower.
Every creative decision you can settle at the image stage is a decision you will not pay for at the video stage. That is the entire rule.
The use case that actually shows up
Generic video is a showcase. What clients genuinely ask for, repeatedly, is a talking head: an avatar, a training presenter, a spokesperson for an internal announcement.
For that you do not generate video from scratch — you use LivePortrait, which animates a still photo driven by audio or a reference video. It is orders of magnitude cheaper, runs on a 12 GB card, and the result is more consistent because the identity comes from a real photograph rather than the model's imagination.
⚠️ This is where the legal risk of the whole series lives
Animating a real person's face without explicit, documented consent is the shortest route to a lawsuit — and, depending on the content, to a criminal offence. This is not a terms-of-service question: it is the law.
Product rule: only animate a face with recorded authorisation from the person. Store the consent alongside the file. And label generated material as synthetic.
Where the brain fits in
In everything that is not pixels — and that part costs almost nothing:
- Script and scene breakdown from a two-line brief.
- The visual prompt for each scene, in the form the image model understands (part 5).
- Narration written to length — 150 words per minute is the yardstick.
- Captions, titles, descriptions and social cuts.
A 30-second video consumes maybe R$ 0.02 of text and R$ 5.60 of generation. The brain is 0.4% of the bill and drives a large share of the perceived quality.
The honest advice: do not build it yet
If you are building your platform now, video is last in the queue, for three reasons:
- Almost nobody uses it. In the typical distribution from part 1, video is ~0.2% of interactions.
- It is the piece that changes fastest. Open video models move quickly; what you build today will be outdated in three months — and without an installed user base you rebuilt for nothing.
- It is the piece that disappoints most. User expectations are calibrated by the big vendors' launch reels; five open-source seconds at 720p can underwhelm someone expecting a commercial.
When demand appears, it arrives specific ("I need product videos for the catalogue", "I want the training avatar"). Build it then — for that case, on a large card rented by the hour, running only while the queue has work.
Start with the piece everyone uses
Video is a rare burst on an expensive card. The brain is what runs on every interaction, every day — and it is billed per token, with no machine to administer.
See the token API →Next: voice — listening and speaking are two models, not one.
Keep reading: the state of AI video · a cheap video pipeline · part 7: voice