AI image prompts
How to write AI image prompts that work
A good image prompt says what is in the picture, where, in what light, and how it is framed, in plain sentences. Here is how to write one, the real limits of each model on Root, and the prompt helper built into Root that drafts one for you.
Updated
First month free on the Beginner plan, with 650 credits to try any model.
Images turned into shots in Root
Write a prompt in five parts
Say each part in plain words. Leave out what you don't care about and the model fills it in.
- Subject. Who or what, doing what: an old fisherman mending a net.
- Setting. Where and when: on a harbour wall at dawn, fog over the water.
- Light. Where the light comes from and what it is like: low sun from the left, soft and warm.
- Framing. How close and from where: close-up, eye level, shallow focus, 35mm lens.
- Look. The style and finish: a 1970s colour photograph with fine grain, or a flat-colour illustration.
Root's prompt generator
The prompt box in Root has three buttons. Draft turns a short idea such as "rainy neon alley, detective" into a full prompt. Improve rewrites the prompt you have, keeping your subjects and adding light, framing and texture. Vary gives you two or three different versions to pick from. Pick a house style for all three: Cinematic, Documentary, Anime, Product shot or Noir.
In the node graph, a Prompt Enhancer node rewrites a prompt before it reaches the image node, and a Describe node turns an image into a prompt. Both follow an instruction you can edit. The helper uses credits, and is free when you add your own Anthropic key.
Each model's limits
What the main image models on Root take. The prompt limits are in the questions below.
| Model | Reference images | Shapes |
|---|---|---|
| Nano Banana Pro | Yes | 1:1, 16:9, 9:16, 4:3, 3:4 |
| Nano Banana | Yes | 1:1, 16:9, 9:16, 4:3, 3:4 |
| GPT Image 2 | Up to 8 | 1:1, 3:2, 2:3, 16:9, 9:16, 4:3, 3:4 |
| GPT Image 2.5 Flare | Up to 8 | 1:1, 3:2, 2:3, 16:9, 9:16, 4:3, 3:4 |
| GPT Image 2.5 Sunburst | Up to 8 | 1:1, 3:2, 2:3, 16:9, 9:16, 4:3, 3:4 |
| Seedream 5.0 Pro | Up to 10 | 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 |
| Seedream 5.0 Lite | Up to 14 | 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 |
| Seedream 4.5 | Up to 14 | 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 |
| Seedream 4.0 | Up to 14 | 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 |
| FLUX.2 [max] | Yes | 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3, 21:9 |
| FLUX.2 [pro] | Yes | 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3, 21:9 |
| Ideogram v3 | Yes | 1:1, 16:9, 9:16, 4:3, 3:4 |
| Krea 2 Large | Style images | 1:1, 4:3, 3:2, 16:9, 2.35:1, 4:5, 2:3, 9:16 |
| Krea 2 Turbo | No | 1:1, 16:9, 9:16, 4:3, 3:4 |
Tips that change the result
- Pick the shape in the settings, not in the prompt. Writing “wide shot” does not make the image wide.
- Put words that must appear in the image in quotes, and keep them short.
- Only Ideogram v3 reads a negative prompt, and only one sent by API or from Claude and ChatGPT; the Root app has no negative prompt box. Elsewhere, describe what you want instead.
- For a character or product that has to match, add a reference image rather than a longer description.
- When you edit an image, say what changes and what must stay: “make it night, keep her face and the car”.
- For a masked fill, describe only what goes inside the mask, and how it matches the light around it.
- Change one thing at a time between takes, so you can see what made the difference.
Related guides
AI image prompt questions
Is there an AI image prompt generator in Root?
Yes. In the prompt box, Draft turns a few words into a full prompt, Improve rewrites the prompt you have, and Vary gives you two or three different takes on it. You can steer it with a house style: Cinematic, Documentary, Anime, Product shot or Noir. It uses credits, and is free when you add your own Anthropic key.
How long can an image prompt be?
It depends on the model. Krea 2 Large and Krea 2 Turbo: up to 5,000 characters; GPT Image 2, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst: up to 32,000 characters. The other models in the table above have no limit of Root's own. In practice a few clear sentences work better than a page.
Can I turn an image into a prompt?
Yes. In Root's node graph, wire an image into a Describe node and it writes a prompt that an image or video model could use to make a picture like it. You can edit the instruction it follows.
Do negative prompts work?
Only on Ideogram v3 among these image models, and only when you send one by API or from Claude and ChatGPT: the Root app has no negative prompt box. Otherwise, say what you want rather than what you don't: "a clear blue sky" rather than "no clouds".
How do I get the words in an image spelled right?
Put the exact words in quotes, keep them short, and say where they go and what the letters look like. Use a model made for text, such as Ideogram v3 or GPT Image.
How do I keep the same character in every image?
Use reference images of the character instead of describing them again each time. Seedream 5.0 Lite, Seedream 4.5 and Seedream 4.0 take the most, up to 14 in one run. Keep the name and the clothes the same in every prompt.
Can I try Root's image models for free?
For a month, yes. The first month of Root's Beginner plan is free and comes with 650 credits, enough for 325 Krea 2 Turbo images. After that the plan is $9 a month, and you can cancel whenever you like.
Try Root free for a month
The first month of the Beginner plan is free and comes with 650 credits to spend on any model. After that it's $9 a month, and you can cancel whenever you like.