The gap between an image everyone likes in a review and an asset that can go live is where most AI projects quietly lose their savings. Here is what sits in that gap.
01 — Selection and forensics
Choose the frame, then examine it properly at 100%. Hands, edges, reflections, repeated texture, geometry that does not resolve, perspective that does not agree with itself. Some of these are fixable in an hour. Some mean the frame is not the one, however good it looks small.
02 — Resolution and reconstruction
Generated output is rarely at campaign resolution, and naive upscaling invents detail that will not survive a billboard. Detail that matters gets rebuilt rather than interpolated.
03 — The built layer
Product, typography, logo, anything that has to be exact — modelled or set and composited in. This is usually the step that decides whether the asset reads as professional or as an experiment.
04 — Compositing and integration
Matching lighting, colour, grain, depth of field and lens character between generated and built layers, so the seam disappears. Ordinary postproduction craft, applied to a new kind of source material.
The last 10% of the image is 50% of the credibility.
05 — Grade and finish
One grade across the whole campaign. This is what makes twelve separately generated images look like one shoot, and it is the cheapest consistency you will ever buy.
06 — Versioning and delivery
Formats, crops, safe areas, languages, colour profiles, file specs. Unglamorous, entirely predictable, and the reason a campaign either ships on Monday or does not.
What this changes about planning
Budget the pipeline, not the generation. The generation step is genuinely much cheaper than what it replaced — which is precisely why the remaining steps now represent most of the cost, and why pretending they do not exist is the most expensive decision available.