Photopurr DocsGenerating images
Choosing parameters and writing prompts
This page covers a single generation in the Workspace. Image Sets, Workflow and Personas work differently on the surface, but they share the same generation pipeline and the same credit ledger, so every parameter means the same thing everywhere.
How one generation runs
- Upload reference images (optional) and describe the shot you want
- Pick a purpose, model, ratio and image count. The UI shows how many credits this run will cost
- Click generate. Credits are reserved at this moment
- Images land in the current conversation. If the run fails or you cancel it, credits are refunded automatically
To iterate, just keep talking in the same conversation. The previous round's result is carried into the next round as a reference image.
Purpose
The purpose decides which set of e-commerce photography rules is added on top of your own prompt.
| Purpose | Best for |
|---|---|
| Free | Default. Follows your prompt only, no extra rules |
| Hero image | Marketplace hero shots: product centered, clean background, selling point up front |
| Detail | Fabric, stitching, close-ups |
| Buyer show | Real-life feel in an everyday setting |
| White background, no person | Pure white background, product only |
| Edit | Don't paint from scratch, change the image you have. Selecting it reveals an "edit mode" picker |
Models
| Model | Vendor | Strengths | Resolution | Reference limit | Credits per image |
|---|---|---|---|---|---|
| Nano Banana 2 | Strong multi-image consistency, outfit and face swaps. Auto's default pick | 1K / 2K / 4K, default 2K | 14 | 12 / 16 / 23 | |
| GPT Image 2 | OpenAI | When the image must contain text. Auto switches to it when it detects a text request | 1K / 2K / 4K, plus quality low / medium / high | 16 | By quality 5 / 15 / 50, +1 per reference image |
| GPT Image 2.5 | OpenAI | Next generation, in testing | Same as GPT Image 2 | 16 | Same as GPT Image 2 |
| Seedream 5.0 Pro | ByteDance | Layout and structure understanding, in testing | 1K / 2K, no 4:5 | 10 | 6 / 12 |
| Qwen Image 3.0 Pro | Alibaba | More accurate Chinese text, in testing | 1K / 2K, no 4:5 | 3 | 6 / 11 |
| Flux Kontext | Black Forest Labs | Edit-only: changes just the part you name, keeps everything else | 1K | 1 | 5 |
Auto is the default: with no special requirement it uses Nano Banana 2; if the prompt asks for text in the image (brand name, slogan, price) it switches to GPT Image 2; if the number of reference images exceeds the current model's limit, it picks another model that can hold them.
Models marked "in testing" are usable, but output quality and limits are still being measured. Please report anything odd.
Only want a few models in the dropdown? Turn off the ones you don't use under Account → Settings → Models.
Ratio
1:1, 3:4, 4:5, 16:9, 9:16. Default 3:4.
Marketplace hero images are usually 1:1, detail-page portraits 3:4, short-video covers 9:16.
Resolution and quality
Resolution is 1K / 2K / 4K. 4K is noticeably more expensive and slower: test your prompt at 1K or 2K, then render the final at 4K.
Quality low / medium / high exists only on the GPT models, default medium. high costs 50 credits per image, so save it for finals.
Image count
Generate 1, 2 or 4 images per run, charged per image. If you ask for 4 and only 2 succeed, the 2 that failed are refunded.
Reference images
Reference images make the new image follow their product, composition, model or style. Drag them anywhere on the page, paste them, or pick from the Library. The limit depends on the model (see the table above); going over is blocked before generation.
Images you use all the time (store logo, fixed model, fixed background) can be marked as favorite references. They then sit at the top of the input box, one click away.
Editing
With purpose set to Edit you can choose from five edit modes: AI recolor, product repaint, face repaint, hands and feet fix, white-background cutout. The last three touch one area and must leave the rest untouched, so the system suggests switching to Flux Kontext. The first two repaint the whole image and follow whichever model you picked.
Editing requires exactly one source image. Without it the run is blocked.
Cancelling and partial success
You can cancel while generating. Cancel before the request reaches the model and you get a full refund; once it has been sent, the vendor is already charging us, so that run is not refunded.
Asking for 4 and getting 2 does happen: you are charged for the actual count and the difference is refunded automatically. If the system could not honor part of your request (say you asked for 4K and that model can't deliver it), the result card shows a notice.
Writing prompts
Describe concrete, photographable things, not abstract judgments:
- "White-background T-shirt hero shot, model standing front-on, soft top light, slight turn of the body"
- Not "premium feel, good texture, make it nicer"
The model can't parse "premium feel", but it understands "soft top light". If you need text in the image, write the exact text in quotes, for example the words "Free shipping" across the top.