How to write image prompts

Every picture in Generate Media starts from your character's portrait. This guide explains how that works, then shows three ways to prompt, from the quickest to the most control: templates, your own prompt, and Edit Image.

How your images are made

The model doesn't draw your character from scratch. It takes their portrait as a reference and redraws the pose, outfit and setting around them. That's why you never need to describe their face, build or hair color: the portrait already shows them to the image model.

You can choose between two models under Settings → Model:

• Muse (the default) is built on a Krea variant and tuned for explicit, photo real detail. It's less reliable at keeping a character's likeness but provides the best image quality.

• Morph (Premium) is built on Qwen architecture, a large general-purpose model made for editing pictures. It stays closer to your character's portrait, so their face and body change less. In a group chat it can put up to three characters in one picture. It can struggle with fully explicit scenes.

Both cost 6 Spunk per image.

Step 1: Start with the templates

The quickest way to a good image is the three pickers: Pose, Outfit and Scene. Every card is a hand-written, tested prompt, and each has several variations, so the same card gives you a slightly different shot each time.

1. Open Generate Media and choose Create Image.

2. Select a character.

3. Pick a Pose.

4. Pick an Outfit.

5. Pick a Scene.

6. If you like, add Extra detail, such as a specific tattoo or piercing, a mood or a camera angle (Premium).

7. Press Generate.

Custom pose, outfit and scene

If nothing in a picker fits, choose Custom and type your own. Keep it short, like the cards:

• Pose: kneeling on the bed, looking back over her shoulder

• Outfit: red silk kimono, open (you don't need to write "wearing"; it's added for you)

• Scene: rooftop pool at sunset

Keep the prompt enhancer on

The prompt enhancer (Settings → Optimize prompt) is on by default. Before your request reaches the model, it rewrites your picks and extra detail into one full description of the shot. It adds the pose and expression, what the hands are doing, the state of the clothes, a few props that suit the setting, and lighting that fits the scene. It keeps what you asked for, and it doesn't tone anything down.

With templates, leave it on. It's what turns a list of picks into a scene.

Tip: generate the same set of picks a few times before changing anything in edit mode. Each card has several variations, and the enhancer adds its own.

Step 2: Write your own prompt (Premium)

Once you know what the templates give you, you can take full control. Select your character, skip the pickers and write the whole prompt yourself.

For the rawest result, turn the enhancer off. The model then gets your words exactly as you typed them, with nothing added.

What to include

Write one paragraph in the present tense that describes the finished picture, in roughly this order:

1. The shot: a selfie, a mirror shot, full body, a close-up or POV.

2. Pose and expression: what their body is doing, where they're looking and what their hands are doing.

3. Clothing: name each item with its colour and material, and say how it sits. For example, "a white shirt, half unbuttoned and falling off one shoulder".

4. Setting: the place plus two to four props, such as "rumpled sheets, a bedside lamp, a stack of books".

5. Lighting: where the light comes from, such as "soft daylight from a side window".

Here's a quick prompt rewritten by hand, with the enhancer off:

• You type: she is taking selfie, on her knees sticking her tongue out seductively, wearing a slutty college outfit

• Better: A high-angle phone selfie, her arm stretched toward the camera. She kneels on a rumpled dorm bed, tongue out playfully, eyes locked on the lens. She wears a cropped grey university hoodie pulled off one shoulder and tiny low-rise denim shorts. Behind her are string lights on the wall, a stack of textbooks and a band poster. Warm light comes from a desk lamp beside the bed.

It names the shot, spells out the pose and expression, and swaps "slutty college outfit" for specific items with colour and fit. It also adds a setting with props and says where the light comes from. At about 380 characters, it's well under the 700 limit.

Here's what the enhancer makes of a short request, so you can see the level of detail that works:

• You type: show me you in the kitchen

• The model gets: A young woman standing in a modern kitchen, leaning against the marble counter with a slight smile, wearing a cropped white tank top and denim shorts, soft daylight from a side window, kettle and mugs on the counter behind her.

Keep your prompt under about 700 characters. Morph cuts anything past 800.

What to leave out

• How your character looks. Hair colour, age, ethnicity and height come from the portrait, and describing them differently fights it. To add a permanent detail such as freckles, a tattoo or a piercing, put it in Custom Physical Details or Custom Face Details when you create the character, and every image will include it.

• Quality words like "masterpiece, 8k, ultra-detailed". The practice originates from early stable diffusion tags to access higher quality images in data sets which is no longer used with modern models.

• "No…" phrases. The models tend to ignore instructions about what not to show. Describe what is in the picture instead, and put anything you want to avoid in Settings → Negative prompt (up to 200 characters).

Step 3: Refine with Edit Image (Premium)

When an image is close but not quite right, you don't have to start again. Edit Image takes a picture you've already made and changes only what you ask for.

1. Choose Edit Image.

2. Select the image you want to change.

3. Describe the change, or pick a new outfit or scene.

4. Press Generate.

On Morph, edits run on Qwen's editing model, which is made to change one part of a picture while keeping the rest. It looks at your original image, so you can point to things in it ("the same bed", "her necklace") and tell it what to change and what to keep.

How to word an edit

• Start with the change, as an instruction: Change her hair to a glossy vivid blue, keeping its length and style.

• Then say what stays the same: Same pose, same bed, same lighting.

• Make one change at a time. A few small edits in a row work better than one big one.

Tip: for precise changes, turn off the prompt enhancer (Settings → Optimize prompt) before you edit. The model then gets your instruction exactly as written, so it changes only what you asked for and leaves the rest of the picture alone.

If your character's face or body changes

If your pictures start to drift from your character, such as a different face or a slimmer build, switch the model to Morph. Muse draws more of each picture fresh, which gives it more explicit detail but lets features wander. Morph keeps going back to your character's portrait and rebuilds the face and body from it, so they change far less. This works the same way in Edit Image.

Quick fixes

• The image looks generic: add a setting, a few props and the lighting, or turn the enhancer back on.

• The enhancer changed something you wanted: try being more specific, or turn it off and write the whole prompt yourself.

• Something unwanted keeps appearing: add it to the Negative prompt.

• The face or body doesn't match your character: switch to Morph.

• It's almost right: use Edit Image instead of generating again.

• An edit changed more than you asked for: turn the enhancer off and run the edit again.

• The generation failed: see Why did my generation fail?