How AI video models work, from your words to a file
By Christophe Paka · Facts checked on 23 October 2026
You ask for a white mug printed MONDAY. The steam rises, the light is warm, and the mug reads MODNAY. You rewrite the prompt. Then you rewrite it again. One reader put the feeling in a line: it is "basically rolling the dice every time".
Each rewrite is a paid take, and most of them cannot fix the fault. The short answer: lay printed words over the clip in the edit, and start from a photo when a face or a product must match. The figure below shows how AI video models work and why those two fixes work, following one clip through six stages from your words to an MP4 file.
How to read this figure
Start with your words, top left. Stage 1 is your text prompt. Stage 2, the text encoder, turns the words into numbers about meaning, not spelling. That is why the model never copies your letters, and why MONDAY can come out as MODNAY.
Watch the static drop into stage 3. Random noise, like the static on an old TV with no signal, falls into stage 3 from above. The model starts from static because it learned by cleaning static off real clips. The label "latent cubes" only means the static comes in a small, packed form of the video, cut into tiny blocks. Packing it small saves work, but fine detail such as small letters is left out.
Follow the loop inside stage 3. Stage 3 is the cleaner. Its technical name is diffusion transformer. It looks at the whole clip at once, removes a little static, checks your words, and repeats that many times. The "spacetime patches" in the box are the same tiny blocks. The diffusion figure shows that loop on one frame.
Read the lower row: a picture you hand in. The figure calls the packer a 3D VAE: the part that makes video small and unpacks it again, across time as well as space. In the lower row it packs an optional picture, such as a first frame. That packed picture takes the place of part of the starting static, so the clip starts closer to your mug or face.
Exit right: stage 4 unpacks the clip. Stage 4 is the same packer run the other way: it turns the cleaned blocks back into pictures. Small letters were left out when the video was packed small, so they do not come back here. That is why the fix is real text added in the edit.
Finish with sound, filter and file. The dashed audio decoder makes sound in the same run as the picture. The figure names Omni Flash, Veo and Kling as models that have it. Each clip then gets a new voice, so a character can sound different every time. Stage 6 checks the clip, adds a hidden watermark, and writes the MP4 file.
Key takeaways
- Add printed words and logos when you edit; more prompt words cannot bring back detail the model throws away.
- Hand the model a picture when a face or a product must match, because the picture replaces part of the noise.
- Budget three or more takes for every clip you plan to keep.
Worked example: the MONDAY mug, three takes
Run the same prompt three times on Google's Gemini Omni Flash, 8 seconds at 720p. Each take is a fresh trip through all six stages, from new noise.
- Take 1: MONDAY is spelled right, but the steam does not move. Not usable.
- Take 2: the steam rises, but the mug says MODNAY. The letters were fine detail that the compressed copy dropped. Not usable.
- Take 3: right word, moving steam. You keep it.
- The bill: three takes at about $0.80 each is about $2.40 for one usable 8 second clip.
Result: Take 2 failed inside the machine, not in your prompt. The cheaper fix is a clean still of the mug at frame one, or the real word laid over the clip in the edit.
Questions readers ask
Why does AI video mess up text and logos?
The model works on a latent, a compressed copy of the video, and turns it into pixels only at the end. Small letters and logos are the fine detail that compression throws away. Google's own report says its Veo 3 model "is still poor at generating text". Add the real text when you edit.
Am I charged if my video gets blocked?
Not on the major services. Stage 6 is a safety filter, and Google's Veo guide says "You will not be charged if your video is blocked from generating". A clip you simply dislike is different: it passed every stage, so it was billed.
Will a longer, more detailed prompt fix my clip?
Up to a point. Models learned from long, detailed captions, so a paragraph helps. But the model has never met your mug or your actor, and no description fixes that. When a face or a product must match, give the model a photo.
Why does the voice change in every AI video clip?
Models that make sound themselves build a new voice for each clip, so the same character can sound different every time. The fix lives outside the model: make the voice once in a separate voice tool such as ElevenLabs, switch the model's own audio off, and lay that one voice over every clip in the edit.
Sources: Veo 3 Tech Report (Google DeepMind, PDF); Gemini API: Veo guide; Video generation models as world simulators (OpenAI).
This figure is from AI Video Made Crystal Clear, chapter 2.
The chapter builds the idea step by step, with recipes you can follow, a Try it to check yourself, and every source.
Get the book
