The consistency problem
A single illustration is easy. A set of twelve in the same style is the hard problem. Research 10 + 15: prompt words drift; reference images do not. Any illustration skill must inject a brand style reference at every call.
Routing
| Model | Brand lock mechanism | Best for |
|---|---|---|
| Flux Pro / Flux.2 | reference_images[] (up to 8 in Flux.2) + brand LoRA |
Photoreal, stylized 3D, brand illustration sets |
| SDXL + brand LoRA | trained LoRA (6d recipe, ~5k steps on 20 images) | Bespoke brand style, open-weight |
| Recraft V3 | style_id (brand magic) |
Flat vector, editorial illustration |
| Ideogram 3 | style codes | Loose "same vibe" — not strict lock |
| Midjourney v6/v7 | --sref / --cref / --mref |
Concept work; no API, community wrappers only |
gpt-image-1 |
input_image[] |
Edit / composite flows |
First illustration in a set is human-gated. Once approved, its style becomes the reference injected into all subsequent generations.
Brand bundle injection
illustration prompt =
[SUBJECT + SCENE from brief]
+ [style anchor: "in the style of the provided reference images"]
+ [palette: exact hex list from brand]
+ [do_not list as positive anchors: "flat matte surfaces" not "no glossy plastic"]
+ [typography reminder: "no text, no labels"]
+ [technical constraints: aspect, resolution, composition]
+ reference_images[]: [style_ref_01.png, style_ref_02.png, (prior approved illustration).png]
+ LoRA handle / style_id / --sref
Prompt scaffold
An illustration of [SUBJECT: concrete noun phrase] in a [SCENE: clear action/context].
Composition: [centered | rule-of-thirds | off-center-left]. Subject occupies ~60% of frame.
Style: in the style of the provided reference images. Flat vector with soft gradients.
Line weight consistent with references.
Palette strictly limited to: [#hex, #hex, #hex, #hex, #hex].
Materials: matte surfaces, soft ambient lighting, no rim lights, no lens flare.
No text, no labels, no UI elements.
[aspect ratio]. 2048x1280 resolution.
Drop quality modifiers on Flux (no masterpiece, 8k — hurts adherence). Keep them on SD.
Post-processing
- Background removal only if asset is spot art (use BiRefNet for soft edges).
- Palette validation: K-means 8-color in LAB, ΔE2000 against brand palette, regenerate if max ΔE > 10.
- Composition validation: VLM rubric check against style references.
- Resize to target sizes (sharp premultiplied-alpha resize).
Full-set propagation
Workflow for generating N illustrations:
- Generate illustration #1 with the brand bundle and the brief.
- Human gates: accept / regenerate / tweak.
- On accept: add illustration #1 to the style reference set.
- Subsequent illustrations pull the whole augmented reference set → style locks progressively tighter.
- After 3–4 accepted illustrations, train a LoRA for even tighter lock on 20+ asset sets.
Output
illustrations/
├── empty-state-projects.png
├── empty-state-tasks.png
├── onboarding-welcome.png
└── meta.json # includes provenance + "set coherence score" across the batch