Every brand has a fingerprint. Not the logo or the hex codes in a style guide — something deeper. The way light falls in your product photography. The ratio of whitespace to content. Whether your headlines are punchy or poetic. The subtle warmth in your color grading that makes customers feel something without knowing why.
Style Genome™ is CANLAH AI's technology for capturing that fingerprint — not as a PDF that developers skim, but as a 768-dimensional mathematical vector that actively shapes every piece of content the system generates.
This article explains how it works — technically, but accessibly. No PhD required.
Why 768 Dimensions?
Humans describe brands in maybe 10-20 adjectives: "clean," "bold," "warm," "minimal." But the actual visual and tonal properties that create these impressions are far more nuanced. The difference between "premium minimal" and "cheap minimal" isn't one thing — it's hundreds of subtle signals interacting.
768 dimensions is the sweet spot we found through extensive testing. It's enough to capture the full complexity of a brand's creative identity without overfitting to noise. Each dimension represents a learned feature — not a human-labeled attribute, but a statistically meaningful axis of variation discovered by the model itself.
STYLE GENOME™ ARCHITECTURE
INPUT → Brand assets (logos, images, content, style guides)
↓ Multi-modal encoder
ENCODE → 768-dim brand vector (Style Genome™)
↓ Stored in isolated vector DB
GENERATE → AI creates content candidate
↓ Cosine similarity check
VERIFY → Score against brand vector (threshold: 0.85)
↓ Pass → deliver | Fail → regenerate with constraints
OUTPUT → On-brand content delivered to user
The Encoding Process
When you onboard with CanMarket, you upload your brand assets — logos, campaign images, social content, product photography, packaging shots, whatever defines your visual identity. You can also provide text samples (ad copy, taglines, product descriptions) to capture your verbal brand.
Our multi-modal encoder processes these assets through several specialized models:
Color distribution, saturation patterns, contrast ratios, composition geometry, lighting direction, depth-of-field preference, texture frequency
Sentence structure patterns, vocabulary complexity, formality spectrum, emotional valence, humor markers, cultural register
Whitespace ratios, text-to-image balance, grid preferences, margin consistency, element hierarchy patterns
Brand personality axes, value associations, category positioning signals, audience alignment markers
These modality-specific vectors are fused into a single 768-dimensional representation through a learned projection layer. The fusion isn't simple concatenation — it captures cross-modal interactions, like how your brand's color warmth relates to your copy's emotional tone.
The Verification Loop
Once your Style Genome™ is encoded, it acts as a persistent constraint on all content generation. Here's how:
When CanMarket generates a piece of content (an image, a social post, a campaign layout), that output is also encoded into a 768-dimensional vector using the same encoder. We then compute the cosine similarity between the generated content's vector and your Style Genome™.
"Cosine similarity in 768 dimensions captures nuances that no human style guide reviewer could articulate — let alone enforce consistently at scale."
If the similarity score exceeds our threshold (default: 0.85, configurable per workspace), the content passes. If it falls below, the system doesn't just reject — it feeds the deviation signal back into the generation process as additional constraints. The second attempt is typically much closer. By the third attempt, we achieve a 97% pass rate.
Why This Beats Style Guides
Traditional style guides are descriptive — they tell you what the brand should look like. Style Genome™ is generative — it actively shapes what the brand looks like in every output. The difference is enforcement:
Isolation and Security
Every Style Genome™ vector is stored in an isolated vector database partition. Your brand memory is never mixed with, influenced by, or accessible to other customers. Enterprise customers can choose their data residency region (US, EU, APAC). The vectors themselves are encrypted at rest with AES-256 and in transit with TLS 1.3.
Critically, your Style Genome™ is never used to train our base models. It exists solely as a constraint layer — shaping outputs, not inputs. If you delete your account, your Style Genome™ is permanently purged within 30 days.
🧬 Key Takeaways
- → Style Genome™ encodes brand identity into a 768-dimensional vector — covering visual, tonal, layout, and semantic features
- → Every generated output is scored via cosine similarity against your brand vector (0.85 threshold)
- → Failed outputs get regenerated with deviation-informed constraints — 97% pass by attempt 3
- → Unlike style guides, Style Genome™ enforces automatically and captures cross-modal interactions
- → Each customer's vector is fully isolated — never shared, never used for training, deletable on request
See Style Genome™ in action on your own brand.
Try CanMarket Free →