ENHANCE PROMPT
PROMPT UPSAMPLING - GETTY IMAGES/iSTOCK GEN AI
2024-2025
SUMMARY
The AI Generator lets customers describe an image in their own words and generate it. The quality of what comes back is almost entirely determined by the quality of what goes in — a detailed prompt produces a usable photograph, a vague one produces a guess.
When the text-to-image feature launched, roughly 70% of users entered very short prompts. Those prompts generally led to underwhelming image results, and customer satisfaction was negatively impacted. 'Enhance prompt' addresses this by expanding a short prompt into a detailed one before it reaches the image model — and by showing users the expansion, so they can edit it or take their own words back.
PROBLEM + OPPORTUNITY
Problem
Around 70% of users typed a few words and then judged the tool by the weak images that came back. Prompt-writing is a learned skill, and we were asking every customer to acquire it before the product could show its value. The cost fell hardest on the people least equipped to absorb it — newcomers to AI generation, and anyone writing in a language other than English.
Opportunity
The system could write the detailed prompt on the customer's behalf. But we realised users won't fully grasp the benefit of the tool unless we made the feature customer-facing — an invisible improvement produces better images and no understanding of why. Making the rewrite visible, and supporting review and edit, would give users control over something the product was doing for them and would likely increase customer satisfaction.
DISCOVERY: TESTING A HYPOTHESIS
Before designing an interface, we needed to know whether image quality was actually the cause of low uptake. Prompt upsampling was installed behind the scenes for 50% of users across Getty Images and iStock. This allowed us to see whether enhancing prompts invisibly would change the behaviour of that 50%, and to determine whether image quality and gen-to-download rate were correlated at all.
Findings
-
iStock: a 23% increase in gen-to-download rate amongst the 50%.
-
Getty Images: a 2.3× higher download rate than users without it (14% vs. 6%).
The hypothesis held. Better prompts produced images people were willing to take away — which meant the feature was worth building properly, and worth showing.
SOLUTION
1
Done-for-you prompts
After generation, the prompt field splits in two: the previous prompt the user typed, held read-only above, and the enhanced prompt below it, editable, with the cursor landing at the end. The screen is always a comparison rather than a substitution, so the customer can see exactly what the product added to their words.
Enhancement itself runs server-side at the moment of generation. The toggle doesn't transform anything when you press it — it declares an intention, which is what makes it safe to leave on by default.
2
Deciding when to enhance
Upsampling originally applied only to prompts under 20 words. That threshold came from model evaluation: past roughly twenty words, the team stopped seeing meaningful improvement in the generated images. It was a defensible engineering heuristic and a poor user-facing rule, because it made the product's behaviour depend on a word count nobody could see.
It broke down at the point of editing. A user who trimmed an over-cooked prompt back to something simpler would find it inflated again on the next generation — because they had shortened it. We walked all three iteration scenarios under both rules and aligned on removing the threshold: when the setting is on, the prompt is enhanced. One rule, one visual structure, one thing to learn.
The Setting Contract
We considered exceptions — if the user has edited the prompt down, assume they don't want enhancement. Each one added a hidden rule for the customer to carry, and meant the product quietly flipping a setting they control. The principle we worked to instead: whatever the rule is, the product doesn't break it and doesn't toggle it on the user's behalf. Every way out is one the customer chooses.
3
Control and recovery
Use previous prompt reverts the field to exactly what the customer typed, and hides itself once reverted — at that point the prompt is the previous prompt, so the link would be a lie. Turning the setting off collapses the stack entirely and returns the plain prompt bar.
Three moments needed explaining and none deserved permanent UI, so each fires a tooltip once, beside the toggle, dismissed with a close button: hovering the control, reverting to the previous prompt with the setting still on, and beginning to edit an enhanced prompt. That last one matters most — someone who has just spent a minute tuning their wording should never have it rewritten by surprise.
Enhancement sits in the critical path of the primary action, so it also had to fail well. When the model can't be reached, the system generates from the customer's own prompt and says so plainly: "Your prompt couldn't be enhanced, but we generated your images anyway."
ENHANCE PROMPT | PROMPT BAR DESIGN
