Recovering from a failed image generation
Designing recovery for AI image generation when a blocked result leaves the user unsure what to change.
Prototype recording · 12 s
Role
UX/UI Design
Project type
Self-initiated AI product concept
Research
Usability Testing
Deliverable
Interactive Prototype
Project constraints
Solo, self-initiated project
→ Scoped the work to one recovery flow, from a blocked generation to a successful result.
No access to real usage data
→ Gathered evidence through a Tally survey (24 submissions) and online discussions.
Prototype, not a shipped product
→ Tested the interactive prototype in Lyssna instead of measuring live usage.
Research limitations: unmoderated Lyssna tests with 5 and 3 participants, so findings are directional. No live business metrics were measured after testing.
The problem
A failed generation ends the result, not the task.
For people using an AI image generator whose prompt gets blocked. The user still wants the image, but a generic block leaves them guessing what caused the failure and what to change next. Initial research: Survey · Tally · 24 submissions. The recovery flow was then tested in two Lyssna rounds.
Why I explored this: I kept running into failed AI image generations that used up credits without making it clear whether the prompt, a reference image or another input was responsible. After seeing similar frustrations in a survey and in online discussions, I explored how a recovery experience could explain failures and help people act on them.
Usability test · Lyssna · Round 1 · 5 participants
4 of 5 participants recovered successfully
Participants were asked to recover from a blocked generation and reach a successful result.
01
80%
Task success
4 / 5 participants completed the recovery task.
02
4.2 / 5
Average ease of recovery
Participants generally found the recovery flow easy to use.
03
5
Participants
Unmoderated Lyssna usability test.
Key finding: participants expected a clearer way to recover or regenerate after failure.
The solution
Turn failure into recovery.
Instead of treating a blocked generation as an endpoint, the flow helps users identify what to change, review the suggestion, and continue editing.
Key decisions
Four decisions that shaped the flow
The prompt survives the failure
Failed attempts use up credits and give no clue what to change. The request stays on screen, so recovery is a targeted edit.
Point at a fragment, not the whole request
Users had to guess whether the prompt, a reference image or something else caused the block. The suspected input is highlighted.
The system suggests, the user commits
The prompt is the user’s own work. The system proposes a change and nothing is applied until they approve it.
Editing stays open the whole way through
In testing, a participant ignored the suggested fix and edited the prompt themselves. Editing stays open throughout.
User flow
From a blocked generation to a fixed prompt
The user finds out why a generation was blocked, reviews a suggested prompt edit and generates again without starting over. Screens: Generation blocked → What needs changing → Review the change → New generation.
What testing changed
2 of 3 participants recovered in Round 2
The recovery flow was tested again in a second, smaller round. Round 1 is shown alongside for reference.
Round 1
Usability test · Lyssna · Round 1 · 5 participants
80% (4/5)
Task success
4.2 / 5
Average ease of recovery
Round 2
Usability test · Lyssna · Round 2 · 3 participants
67% (2/3)
Task success
4.33 / 5
Average ease of recovery
Round 2 had only 3 participants, so the two rounds are not directly comparable. Unmoderated tests in Lyssna; small samples, so these are observations rather than statistically reliable results.
Final experience
Recovery without starting over.
Reflection
What I learned
Make failure actionable, not merely understandable.
Suggestions should accelerate decisions, not make them for the user.
Recovery paths need to preserve user control.
Potential business impact
↓ Generation abandonment
↑ Successful recovery
↑ Continued use
Potential outcomes based on the concept and prototype; no live business metrics were measured.
