Design Image Generation - System Design | WinJob

Problem Statement: Text-to-Image Generation Platform

Midjourney, Stability AI, and DALL·E turned diffusion into a multi-tenant GPU product: users submit natural-language prompts (plus optional reference images) and receive watermarked PNG/WebP assets within seconds to minutes. Interviewers want control plane vs data plane separation, not a Jupyter demo.

Phase context (understanding)

Section 1 locks decisions for text-to-image generation so later sections never contradict earlier numbers. State upfront that generation is never synchronous at 1024² for SDXL-class models at viral scale—only tiny previews may stream.

Mechanisms to verbalize

Async job model with poll/WebSocket completion
GPU diffusion workers with VRAM-aware scheduling
Safety stack: prompt filter + output classifier
SLO tiers: interactive preview vs HD final

Operational notes

Treat GPU seconds as the scarce currency—API QPS is only the envelope. When interviewers challenge cost, connect queue depth to required g5/A10 fleet size. When they challenge safety, separate pre-generation policy from post-generation pixel classifiers.

Edge cases

Viral drop overloads free tier while Pro SLA must hold
Duplicate prompts should hit content-hash cache, not re-run UNet
Partial S3 upload must not mark job succeeded
Registry publishes new LoRA—workers must pin manifest hash

Interview checkpoints

State average vs peak enqueue QPS with arithmetic
Draw control plane vs GPU data plane in under 90 seconds
Name failure playbook for stuck running jobs

javaOne Dark Pro

1 public record GenerationJob(String id, String userId, JobStatus status, Instant createdAt) {}

pythonOne Dark Pro

1 @dataclass(frozen=True)
2 class GenerationJob:
3     id: str
4     user_id: str
5     status: str
6     created_at: datetime

typescriptOne Dark Pro

1 export interface GenerationJob { id: string; userId: string; status: JobStatus; createdAt: string; }

Why interviewers care

Image Generation interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Problem Statement: Text-to-Image Generation Platform that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Async job model with poll/WebSocket completion
•GPU diffusion workers with VRAM-aware scheduling
•Safety stack: prompt filter + output classifier
•SLO tiers: interactive preview vs HD final

1	@dataclass(frozen=True)
2	class GenerationJob:
3	id: str
4	user_id: str
5	status: str
6	created_at: datetime