Three Nights on an 8GB GPU: Building an Image Site With an AI Coding Agent
Local image generation is practical on a modest consumer GPU, and an AI coding agent can compress a three-week build into three nights. The limiting factor becomes learning how the model fails: negative prompts, upscaling, and lighting wording all behave counterintuitively.
An 8GB VRAM RTX 5060 Ti with 16GB RAM now serves as a local image-generation machine. The workflow compresses the model to about 12GB, generates a small image first, upscales at 2x, and runs one image at a time. That keeps memory usage under control and produces a 1792×2304 vertical image in roughly 30 seconds at 8 steps.
Three evenings of testing overturned three assumptions. Measured texture and brightness pointed to flat lighting as the cause of the fake look, not over-smoothed skin. 4x upscaling turned skin into sandpaper, so 2x is the ceiling. Negative prompts like 'don't include xx' were read and ignored, so the workflow avoids risky subjects instead of forbidding them. Lighting prompts need direction, bright areas, and shadows spelled out; a single word like 'backlight' produces flat front light.
The same seed plus the same person description keeps the same face across different clothes, backgrounds, and scenes, which turns the site into a way to shoot one person anywhere. Ten built-in subjects, three aspect ratios, and two portrait texture modes are available, and the page runs in a browser on a phone. Changing one word at a time is the habit that makes the loop work: alter one word, see what it does, repeat.
Measuring texture intensity and brightness turned a vague 'looks fake' complaint into two numbers, and the brightness gap pointed at flat lighting rather than over-smoothing. A quick measurement beats another round of guessing.
Compressed fast models amplify prompt errors: they guess, so a 4x upscale makes both correct and incorrect guesses equally sharp. The practical ceiling is the model's confidence; the upscale factor only exposes it.
A shorter feedback loop is the biggest speedup from AI coding: half-hour rounds became one-minute rounds, so a human can run hundreds of visual experiments in a single evening.
Seed-based identity control turns image generation into a usable photography workflow: one face can be moved across scenes and outfits without redrawing the person.