AI image work is a design system, not a slot machine
Image generation & brand pipelines
The question: How do you get consistent, on-brand, rights-safe visuals out of image models?
One-off pretty pictures are easy; a repeatable visual system is the professional skill: consistent characters, brand palettes, usable formats, and a clean answer to "do we own this?"
What the lesson covers
The generation loop mirrors vibe coding: brief → generate → select → refine. Craft lives in the brief — subject, composition, lighting, lens/style, mood, negative instructions ("no text, no watermarks") — and in ITERATING on the best candidate rather than re-rolling from scratch. Reference images and style anchors (consistent seeds, style keywords, character sheets) are how campaigns stay coherent across dozens of assets.
Know your tools' temperaments: Midjourney for art direction and mood, Imagen/NanoBanana-class models for instruction-following and text-in-image, open Stable-Diffusion-family models for control (LoRAs, ControlNet) when you need exact layouts. The differences matter less than your pipeline: brief templates, review gates, and asset naming.
Rights and provenance are part of the craft, not an afterthought: model terms differ on commercial use; purely AI-generated work is generally not copyrightable (human creative contribution matters — your Lesson 18 knowledge); brand mascots and real people need explicit care; and disclosure duties (EU AI Act synthetic-content labelling, C2PA credentials) apply to published assets.
The pipeline that scales: prompt templates per asset type → generation batch → human selection against a brand checklist → upscale/retouch → labelled export with prompt + model logged. That log is your audit trail and your reuse library — the prompt library idea from Lesson 5, visual edition.
The organisational question this raises is who now owns visual quality, because the old answer stopped working. When producing an image took a day and a specialist, the specialist was the quality gate; when it takes twenty seconds and anyone can do it, the gate has to move or disappear. In practice it moves to two places: a written brand checklist specific enough that a non-designer can apply it, and one named person who signs before anything ships. Teams that skip both do not produce bad images — they produce inconsistent ones, which is worse, because inconsistency is what a brand is defined by the absence of. This is the seven-levels-of-design argument in operational form.
Key points
- Iterate on the best candidate with style anchors; don't re-roll from scratch.
- Tool temperaments differ (art direction vs instruction-following vs control) — the pipeline matters more.
- Rights: model terms, human-authorship for copyright, living-artist proximity, disclosure/C2PA.
- Log prompt + model per shipped asset: audit trail and reuse library in one.
- When anyone can make an image in twenty seconds, the specialist stops being the quality gate. It moves to **a written checklist a non-designer can apply plus one named signer** — or it disappears, and you get inconsistency rather than badness.
Framework — Brief → Generate → Select → Refine → Clear rights → Ship
Six stages; the two most skipped (select against a brand checklist, clear rights) are the two that separate professionals from slot-machine pullers.
The lab
Build a small, coherent, rights-clean visual kit.
Open this lesson, its lab and its quiz
Sources and further reading
- C2PA content credentials — C2PA
- Higgsfield — cinematic AI visuals — Higgsfield