# Generating images, music and video

Chat is not only for code. Images, music and video are produced by dedicated models, you ask for them in plain language, and the results land under the message ready to download. This page covers what each of the three can do, how to revise a version you already like, and where the assets should end up.

## Three pipelines, in one line each

| You want | Just say so |
|---|---|
| An image | Posters, article art, illustrations, logos, covers, product shots — plus editing, background swaps, outpainting and erasing things from a picture |
| To read an image | What is in this picture, or pull the text out of it (OCR) |
| Music | Background music, a theme, an intro sting — instrumental or with vocals |
| Video | Teasers, intros, product demos, or bringing a still image to life |

There is no parameter syntax to learn and no need to write a long prompt up front — **the default is to make a first version and show it to you**. Retries are cheap. The only things asked in advance are the ones where guessing wrong wastes the whole attempt: the exact copy that goes on a poster, the lyrics to be sung, which direction to extend an image.

## What the image pipeline does

| What you want | Example |
|---|---|
| Text to image | "Make an image for this post — warm tones, a cup of coffee on a desk" |
| Local edit | "Same image, change the headline to *What should I drink today*" |
| Cut out and replace the background | "Cut the product out and put it on plain white" |
| Outpaint | "This is too narrow — extend it sideways to 16:9" |
| Erase an element | "Remove that cup in the bottom right" |
| Read an image / OCR | "What error is this screenshot showing?" "Pull the text off this invoice" |

Reading images saves the most time in practice: paste an error screenshot, a photo of a contract or a handwritten note straight into the composer instead of retyping it.

## Music

You get an actual piece of music, not stitched-together samples. Genre, mood, what it is for, vocals or not, and which voice — those are things you have an opinion about and they change the result, so the agent will ask in a sentence or a card. BPM, instrumentation and mixing jargon will not be put in front of you.

If the lyrics have to say something specific, say so up front; otherwise they can be written for you.

## Video

- **Two to thirty seconds** per generation, with an audio track
- **Text to video** — describe the shot
- **Image to video** — hand over a still and let it move; this is the safest route when you already like the frame
- **Multi-shot pieces** — anything past five to eight seconds, or with several actions or locations, gets a script and a look locked first, then each shot is generated against it, then the shots are cut together and given one continuous music bed

Video is the one pipeline worth a short conversation before starting, because a run takes minutes. Expect three questions: **what is in the shot, where will this be used, and do you already have footage or a still**. Length, aspect ratio and whether it needs sound are derived from where it will be used — you will not be asked to pick parameters.

## Revising a version: the part that goes wrong

"Make this one brighter" is the most common follow-up, and it can mean two very different things:

| | What you mean | What happens |
|---|---|---|
| **Local edit** | This exact version, change one thing (brighter, drop the cup, new background, different headline) | **The asset itself** goes in as the input, so the composition and details survive |
| **Start over** | Wrong direction, different style, try again | The previous prompt is the baseline; only the part you disliked changes, then it regenerates |

When you say "this one" or "that shot", you almost always mean a local edit. Running text-to-image again produces **a different picture** — composition, subject and details all change, and nothing you approved survives. **Being specific protects you**: "same image, make the background off-white" is much safer than "try another one".

Revisions always get a new filename (`poster_v2.png`). The earlier version is never overwritten, so you can compare and go back.

## Where the assets should live

- **Just for you to look at or download** — leave them in the artifact list under the message
- **For the app to use long term** (site imagery, page music, a demo video on the homepage) — have the agent upload them to that app's object storage and reference them from the code. They then load over an accelerated address instead of your app server's bandwidth, and they do not bloat the container image
- The original URLs returned by the generation models **all expire**. Never write one into your code or hand it to someone as a permanent link

## Common questions

**Q: Do I need to write a long prompt?**
A: No. Say what you want and where it will be used, in plain language.

**Q: How long can a video be?**
A: Two to thirty seconds per generation. Longer pieces are several shots generated separately and cut together, so there is no hard ceiling — just more time.

**Q: Will the same person look consistent across shots?**
A: Multi-shot pieces lock the look up front before generating each shot, which is far more consistent than generating clips one by one, but it is not perfect. When a character has to stay identical, driving image-to-video from one reference still is the more reliable route.

**Q: Can a generated image go straight onto my site?**
A: Yes. Say "put this in the homepage banner" and the agent uploads it to the app's object storage and wires it up.

**Q: Can I edit a photo I uploaded?**
A: Yes. Uploads work exactly like generated assets as inputs for editing, outpainting, erasing and image-to-video.

**Q: Do I have to ask for images separately when writing marketing content?**
A: No — finished marketing copy comes with its artwork and cover already generated. See `07-marketing.md`.

Read next: `11-documents.md`