AI Image Generation
Image Generation Productization Checklist
In Alice, connecting the image generation API took just one day. But between that day and users actually being able to use it, there were three months of productization work. This checklist shows what that 90% looks like.
Connecting the API ≠ Feature Launch
Check items below to see how completion changes
Productization Checklist (expand by dimension)
Experience Layer
Generation Progress Feedback
Alice shows a "drawing..." status animation during image generation so users never face a blank screen. Estimated time and progress indicators make waiting predictable.
Generation Progress Feedback
Progress animation + estimated time — no blank-screen waiting
Failure Retry Mechanism
One-click retry after model failure, no need to re-describe the request
Result Preview and Selection
Multiple images per generation for selection, with zoom preview support
History
Previous images can be reviewed, reused, and compared
Quality Layer
Reference Image + Prompt Optimization = Stable High Quality
Character reference images ensure consistency; LLM translation optimization ensures every Prompt is in the format the image model understands best. A dual guarantee.
Prompt Optimization (LLM Translation)
User speaks naturally; LLM translates into the model's optimal input format
Character Consistency
Reference image anchoring keeps character appearance stable across scenes
Size Adaptation
Avatar / chat image / social wide image — auto-adapts to different sizes
Quality Fallback
Detect and filter obviously low-quality results (blurry, distorted)
Engineering Layer
Models can go down, slow down, or get expensive — all need handling
Fallback chains ensure availability, timeout policies protect experience, and cost controls prevent financial loss. All three are essential.
Model Fallback Chain
Automatically switches to backup when primary model goes down
Timeout and Retry
Reasonable timeout threshold + auto-retry N times before giving up
Cost Control
Per-user/time-slot quotas, monitor abnormal calls, prevent cost explosion
Result Persistence
Images stored on CDN — no loss, traceable, replayable
Safety Layer
Image generation involves content safety and privacy
Reference images uploaded by users must not leak; generated results must comply with rules; permission boundaries must be clear and transparent.
Input Content Moderation
Filter non-compliant Prompts, intercept sensitive/illegal descriptions
Output Content Moderation
Even if the Prompt is fine, generated results must also pass review
Copyright Risk Control
Avoid generating images highly similar to existing works
Privacy Protection
User reference images are not leaked or used for training; storage policy is transparent
Connecting the API is only 10%; the other 90% is productization. Alice's image generation feature took three months from working API to production launch. This checklist isn't only for image generation — any AI feature moving from Demo to Production must cover all four dimensions: experience, quality, engineering, and safety.