mohamedabdallah-14/prompt-to-asset/brand-consistency

Brand consistency

Engaged when a user is generating more than one asset for the same brand, or when an existing brand.json / brand.md is in play. Wraps asset_brand_bundle_parse, asset_generate_*, and validation.

BrandBundle shape

palette:
  - { name: primary,   hex: "#0A1F44", usage: mark }
  - { name: accent,    hex: "#FF6B6B", usage: highlight }
  - { name: neutral,   hex: "#FAFAFA", usage: background }
typography:
  primary:   { family: "Geist Sans", fallback: "geometric sans-serif, bold" }
  secondary: { family: "Geist Mono", fallback: "monospace" }
style_refs: ["assets/brand/ref-1.png", "assets/brand/ref-2.png"]
lora:       "assets/brand/brand.safetensors"       # optional Flux/SDXL
sref_code:  "--sref 1234567890"                    # optional Midjourney
style_id:   "uuid-of-recraft-style"                # optional Recraft
do_not:     ["drop shadows", "gradients", "photorealism"]
logo_mark:  "assets/brand/mark.svg"                # canonical mark for composition

Call asset_brand_bundle_parse({ source }) to build this from a brand.md, brand.json, DTCG tokens.json, or AdCP spec.

Palette enforcement per model (strongest → weakest)

Fallback that always works: K-means remap in LAB space + ΔE2000 validation. asset_validate flags drift; post-process recolors nearest palette entries.

Style reference management

• Canonical anchor — one .png or .svg that defines the brand look. Required for consistency across >3 assets.
• Accepted-asset promotion — after an asset passes validation + user acceptance, add it to style_refs[] so subsequent generations see richer brand context.
• CSD similarity score — Contrastive Style Descriptor embedding comparison. Threshold: ≥0.72 pass, 0.60–0.72 review, <0.60 fail. Tier-2 validation.
• Model-specific handles:
• Midjourney: --sref <image> (loose) vs --cref <image> (character lock) vs --mref <image> (object lock). Use --sw 250 for tight style, --sw 100 for loose brand.
• Recraft: style_id is a persistent UUID bound to a trained brand style.
• Flux + IP-Adapter: pass reference image; set ip_adapter_weight: 0.7–0.9 for style, lower for just color.
• SDXL + LoRA: trigger word in prompt + lora_scale: 0.7–1.0.

LoRA training workflow (≥20-asset sets)

ROI breaks even at ~20 assets. For smaller sets, IP-Adapter or --sref is cheaper.

1. Collect 20+ accepted assets (logo, icons, illustrations) in the brand style.
2. Call trainBrandLora MCP tool (Replicate/Modal endpoint, ~$50–$200, ~2 hours).
3. Store the returned .safetensors path in BrandBundle.lora.
4. Set trigger word (e.g. "acme_brand") + lora_scale: 0.8–1.0 in subsequent Flux/SDXL generations.
5. Re-validate style similarity — should jump to CSD ≥0.85 vs references.

Asset-type adaptation

Same brand expresses differently per asset. Honor the brand, but respect the medium:

Do-not list encoding

BrandBundle.do_not[] should never be forwarded as literal "no X" to Flux/Ideogram/Recraft. Translate per dialect:

do_not: ["drop shadows"]       → affirmative: "flat design, crisp edges"
do_not: ["gradients"]          → affirmative: "solid fills, flat colors"
do_not: ["photorealism"]       → affirmative: "flat vector illustration style"
do_not: ["rounded corners"]    → affirmative: "sharp geometric corners"

SDXL and Midjourney get literal negatives. Everyone else gets positive anchors.

Typography consistency

Composite, don't generate. Web fonts via Satori for OG images; SVG <text> with font stack for favicons (if required); brand-font text in Figma/Illustrator for logo wordmarks.

• typography.primary.fallback is what the model sees in the prompt ("geometric sans-serif, bold") — not the font file name. Diffusion samplers cannot read Geist.
• For Flux.2: typography: { font_class: "geometric sans-serif", weight: "bold" } can be passed in the JSON schema.

Validation across the set

After each generation, run in addition to normal tier-0/1:

tier_1 (set-aware):
  - palette ΔE2000 vs brand palette ≤ 10
  - CSD style similarity vs canonical anchor ≥ 0.72
  - stroke-width histogram consistent within set (σ ≤ 20%)
 
tier_2 (VLM-as-judge, if enabled):
  - "does this asset look like it comes from the same brand as [ref]?"
  - score 1–5; require ≥ 4 for acceptance

Convergence detection

When to stop regenerating:

• 3 of the last 4 generated assets cleared ΔE ≤ 10 AND CSD ≥ 0.72 on the first try → brand bundle is tight.
• 2 of the last 3 required post-process recolor → bundle is loose; consider LoRA training or add more refs.
• Any asset required 3+ regenerations after acceptance → brand brief is ambiguous; ask user to clarify.

Common mistakes

• Passing do_not[] literally to Flux (errors or weakens generation)
• Relying on prose color description alone for Ideogram (ΔE drifts >15)
• Skipping style_id on Recraft when brand lock is available (loses 50% of the value)
• Training a LoRA on <15 assets (overfits; produces artifacts)
• Never updating style_refs[] after accepted assets (lose compounding consistency)

Research

• docs/research/15-style-consistency-brand/15a-consistent-character-and-mascot.md — character/mascot consistency
• docs/research/15-style-consistency-brand/15b-style-transfer-sref-b-lora.md — --sref, IP-Adapter, LoRA
• docs/research/15-style-consistency-brand/15e-full-asset-set-consistency.md — full-set propagation
• docs/research/24-skills-for-p2a/05-brand-consistency-skill.md — full design spec