Design an AI Image Generation Service — System Design Practice | AmanAI Lab

Sheet›Design an AI Image Generation Service

OpenAIGoogleMetaStability AI

45:00

Hard

The Problem

Problem

Design a text-to-image generation service (like Midjourney / DALL·E) serving 2 million users who submit prompts and receive generated images.

The service should:

GPU-bound async job processing, queueing & autoscaling, result delivery, cost under expensive hardware, and content safety.

Scale & Constraints

Must Cover0/10

Hints (if stuck)

💡 This is an async, GPU-bound job system: the API enqueues, GPU workers pull from a queue, results land in object storage + CDN.

💡 Batch multiple generations per GPU forward pass where possible to lift utilisation and cut cost.

💡 Use priority queues (paid users first) and autoscale workers on queue depth; spot GPUs cut cost with checkpoint/retry.

💡 Deliver results via polling or webhooks/SSE with a job id — never hold an HTTP request open for 30 seconds.

💡 Run NSFW/safety classifiers on both the prompt and the final image; cache by (prompt, params, seed) for identical requests.

0 words · auto-saved

Problem

Design a text-to-image generation service (like Midjourney / DALL·E) serving 2 million users who submit prompts and receive generated images.

The service should:

•Accept a prompt (+ optional params: size, style, seed) and return images

•Handle generations that take seconds to tens of seconds on GPUs

•Show progress / queue position and deliver results when ready

•Enforce safety (NSFW / disallowed content) on both prompt and output

What you'll be assessed on