CreatorLoop

The interesting part of this project was not prompting. Stable Diffusion produces good car imagery with reasonable inputs. The difficulty is that every image costs real GPU seconds, GPU capacity is expensive and slow to acquire, and users expect an interactive experience from a workload that is anything but.

That tension - interactive expectations over batch-shaped economics - determined nearly every decision here.

A base model knows what a car looks like in general. It does not reliably know what a specific model looks like from a specific angle, and prompt engineering does not close that gap - it produces something plausible that an enthusiast identifies as wrong immediately. In automotive imagery, the audience is precisely the people who notice.

Training LoRA adapters gives the model that specific knowledge at a fraction of the cost of fine-tuning a full model, and adapters can be composed - a vehicle adapter with a style adapter - which keeps the number of trained artefacts manageable as the catalogue grows. Full fine-tuning per vehicle would not have scaled past the first dozen.

ComfyUI is usually seen as a graphical tool for experimentation. Here it serves as the execution engine behind an API, which is a deliberate choice: it means the pipeline that produced a result during experimentation is the same pipeline that runs in production, rather than a reimplementation that subtly diverges.

Generation workflows involve many stages - conditioning, sampling, upscaling, post-processing - and the version of every stage matters to the output. Keeping the graph as the artefact makes those workflows versionable, and makes it possible to reproduce an image a user generated last month.

GPU capacity cannot scale instantly. When demand spikes, requests wait, and the honest design question is what a waiting user sees.

The system is queue-based with visible position and progress rather than an indefinite spinner, because a wait with a known duration is tolerable and a wait without one reads as failure. Cold starts were the specific enemy: loading multi-gigabyte model weights onto a GPU takes long enough that scaling from zero produces timeouts during exactly the traffic spike that triggered the scale-up. Keeping a warm floor of capacity costs money and is cheaper than the alternative.

GPU time is the dominant operating cost and it scales linearly with usage, so cost per image was instrumented before it became a problem rather than after. The levers that mattered were caching intermediate results across similar requests, generating previews at lower resolution and only upscaling on request, and batching where latency budget allowed.

Generating everything at full resolution immediately is the intuitive design and it makes the unit economics fail at exactly the point the product succeeds.