Most disappointing AI images come from one of two mistakes: the wrong kind of generation for the job, or a prompt that describes a mood instead of a picture. Both are quick to fix.
Three different jobs
Image models on TurboMax are grouped by what you give them:
- Text to image. You write a description; the model invents everything. Best when you have no source material and want to explore.
- Image to image. You upload a picture as a reference, and the model makes a new image guided by it. Best for variations, restyling, and keeping a character or product recognisable.
- Image editing. You upload a picture and ask for a specific change: swap the background, remove an object, change the colour of the jacket. Everything you do not mention should stay as it was.
If you want "this photo, but the sky is at sunset", that is an edit. Writing it as a fresh text-to-image prompt throws away the photo you already have.
The models at a glance
All of these are on one prepaid balance, so you can try the same prompt on several and keep the one that works.
| Model | Output size | Notable for |
|---|---|---|
| GPT Image 2.5 | 1K, 2K, 4K | Precise prompt adherence, stills and edits |
| GPT Image 2.6 Economy | 1K | The newest GPT Image at a fixed low price |
| Nano Banana 2.1 | 1K, 2K, 4K | Google's latest image model, strong edits |
| Nano Banana 2 Lite | 1K | Quick, inexpensive images and edits |
| Nano Banana | HD | Fast, consistent generation and edits |
| Seedream 5 Pro | 1K, 2K | ByteDance's top-quality Seedream 5 |
| Seedream 5 Flash | 1K, 1.5K, 2K | Fast Seedream 5 at one low price |
| FLUX 3 Klein | 1K, 2K, 4K | Fast, with multi-image edits |
| FLUX.1 Kontext | HD | Instruction-based editing |
| Qwen-Image 2.1 | 1K, 2K | Sharp text rendering, transparent backgrounds |
| Grok Imagine Image 2.0 | HD | xAI's image model |
| GPT-4o Image | HD | Prompt understanding and text rendering |
Per-image prices differ a lot between these, and some models charge more at higher resolutions. The models page lists the current rate for each.
Picking one
- You need print or large-screen quality: GPT Image 2.5, Nano Banana 2.1 or FLUX 3 Klein at 4K.
- You are exploring ideas and want many attempts: a low-cost model at 1K, like GPT Image 2.6 Economy, Nano Banana 2 Lite or Seedream 5 Flash. Move the winning prompt to a premium model afterwards.
- There are words in the image (a poster, a label, a thumbnail): Qwen-Image 2.1 or GPT-4o Image.
- You are changing an existing picture: any model that lists image editing. FLUX.1 Kontext and Nano Banana 2.1 are built around editing.
Aspect ratio first, not last
Decide where the image is going before you generate it. Cropping a square image to 9:16 afterwards throws away most of the frame, and the model composed the picture for a shape you did not use.
| Destination | Ratio |
|---|---|
| Phone wallpaper, Stories, Reels, Shorts | 9:16 |
| Website hero, YouTube thumbnail, slides | 16:9 |
| Feed posts, profile images, product tiles | 1:1 |
| Portrait feed posts | 4:5 |
| Cinematic banner | 21:9 |
Most models offer 1:1, 16:9, 9:16, 4:3, 3:4, 3:2 and 2:3. 4:5 is offered by Nano Banana 2.1 and Nano Banana 2 Lite. 9:21 is on FLUX 3 Klein and Qwen-Image 2.1. Several models also offer auto, which lets the model choose; set it explicitly when the destination matters.
If the image will become a video, generate it at the video's aspect ratio. See the image-to-video guide for why that still decides the shot.
A prompt structure that works
Image models respond to concrete, visible things. Words like "stunning", "masterpiece" or "8k" describe how you hope to feel, not what is in the frame, and they rarely change the result.
Write the prompt in this order:
- Subject. Who or what, specifically. "A ceramic coffee mug, matte sage green, with a chipped rim", not "a mug".
- Action or state. What it is doing or how it is placed. "Steam rising, sitting on a stack of two hardback books."
- Setting. Where. "On a wooden windowsill, rain on the glass behind."
- Composition. Framing and viewpoint. "Close-up, eye level, subject on the left third, shallow depth of field."
- Light. Direction and quality. "Soft overcast daylight from the window, no harsh shadows."
- Style or medium. Photo, illustration, 3D render, watercolour. Name the medium instead of piling up adjectives.
Put together:
A matte sage-green ceramic coffee mug with a chipped rim, steam rising, sitting on two hardback books on a wooden windowsill, rain on the glass behind. Close-up at eye level, mug on the left third, shallow depth of field. Soft overcast window light. Product photograph.
Every phrase in that prompt is something you could point to in the finished image. That is the test.
Text in images
Put the exact words in quotation marks and say where they go: a poster with the headline "OPEN LATE" in bold white capitals across the top. Keep it to a few words. Long sentences are where even the best text-rendering models start to misspell.
Editing prompts
For an edit, describe only the change, and say what must stay:
Replace the background with a plain light-grey studio backdrop. Keep the product, its shadow and the lighting exactly as they are.
Re-describing the whole picture in an edit invites the model to redraw parts you wanted to keep.
Getting consistency across a set
A brand shoot, a storyboard or a product range needs images that clearly belong together. Separate text-to-image generations will drift in face, colour and style, even with the same prompt.
The reliable method:
- Generate, or upload, one image you are happy with.
- Use it as the reference for image-to-image or image editing for every variation.
- Change one thing per generation: the pose, the background, the angle.
Debugging a prompt
| Problem | Likely cause | Fix |
|---|---|---|
| Result ignores half the prompt | Too many competing ideas | Cut to one subject and one setting |
| Generic, stock-photo look | Mood words instead of details | Replace adjectives with visible specifics |
| Wrong framing | No composition stated | Add viewpoint, distance and placement |
| Misspelled text | Too much text, or not quoted | Quote it, shorten it, try Qwen-Image 2.1 |
| Edit changed more than asked | Prompt re-described the whole image | Describe only the change; say what stays |
| Awkward crop later | Generated at the wrong ratio | Set the destination's aspect ratio up front |
Spend less per keeper
The cheapest way to reach a good image is to separate exploring from finishing:
- Explore composition and wording on a low-cost 1K model.
- When a prompt reliably gives the image you want, run it once on a premium model at the size you need.
- Make the variations with image editing, not new text prompts.
That pattern carries straight into video, where each attempt costs much more. Fix the still first, then animate it.
Generate and edit images on TurboMax AI with GPT Image, Nano Banana, Seedream, FLUX, Qwen-Image and Grok Imagine, all from one prepaid balance. No monthly subscription.