
Realistic AI stock photos of Asian people without fine-tuning: Flux, Seedream and good prompt context
Summary
Stocklibrary.ai generates photorealistic AI stock photos — including realistic images of Asian people — without fine-tuning a custom model, using Flux, Seedream and enough prompt context to make region and style choices believable by default.
Problem
Paid stock photography charges per photo and rarely has the exact combination of subject, region, style and framing a buyer actually wants. Fine-tuning a dedicated model per look is slow and expensive to maintain. The alternative is prompt engineering: turning a handful of simple choices into prompt context precise enough that an off-the-shelf model produces a believable, on-brief result without any fine-tuning step.
Architecture
A parameter injection layer turns a visitor’s style, vibe, colour and aspect-ratio choices into prompt text, rather than exposing raw prompting to the user. Generation runs across Flux and Seedream, selected per request. Region-aware prompt context is what makes generated people read as believable locals — including realistic images of Asian subjects — rather than the generic look off-the-shelf models default to. Where a visitor uploads a reference photo, image-to-image generation anchors the output’s composition to it instead of generating from a blank prompt. Cloudflare stores the generated library, Airtable holds the underlying schema, Make.com wires the generation workflow together, and Webflow runs the front end. Grafana and Sentry cover observability across the pipeline.
The Hard Parts
Getting a general-purpose, non-fine-tuned model to produce believable regional representation consistently is a prompt-context problem, not a model problem: the parameter injection layer has to encode enough region- and style-specific detail in every prompt that the model doesn’t fall back to its generic default. Image-to-image anchoring has a similar failure mode in reverse — anchor too loosely and the output drifts from the reference photo; anchor too tightly and it just returns the input.
Results
Stocklibrary.ai has 600+ users generating stock photography without a photoshoot or a fine-tuned model behind it.
Stack
Flux and Seedream handle generation, chosen per request. Cloudflare stores the library, Airtable holds the schema, Make.com runs the workflow, and Webflow serves the front end. Grafana and Sentry cover observability.