From Automation to Intelligence: Our next step in group image generation

At Relayter, our philosophy is clear: AI should be a multiplier for creative precision, not a compromise on brand integrity.

This conviction has guided our first major strategic investment in Artificial Intelligence, focusing on a specific, high-leverage problem that bridges the gap between automation and intelligence: The optimization of product group images.

Group images—the curated, multi-product compositions featured prominently in digital banners, flyers, and in-store displays—are a key visual element in a retail campaign. They serve as the immediate visual shorthand for a promotion and contribute to both the aesthetic and commercial success of the campaign.

The optimization of these images is a critical factor with a profound, direct impact across four core dimensions:

1. Premium Aesthetic & Brand Perception: The arrangement determines how sophisticated, luxurious, or contemporary a campaign is perceived, ensuring a consistent, high-quality visual standard that aligns with the brand’s aesthetic guidelines. (Combines ‘Premium Aesthetic’ and ‘Brand Perception’)
2. Visual Hierarchy & Focus: By controlling scale, placement, and layering, the composition directs the customer’s eye to the key ‘hero’ product or value proposition, ensuring that the most important products or price points are emphasized correctly. (Combines ‘Visual Hierarchy and Focus’ and ‘Visual Hierarchy’)
3. Campaign Attractiveness & Curated Perception: An intelligently composed group feels like a curated selection, managing visual flow and ensuring the banner is pleasing to the eye, thus creating compositions that are more engaging and stop the scroll. (Combines ‘Compositional Balance,’ ‘Curated vs. Crowded Perception,’ and ‘Campaign Attractiveness’)
4. Conversion Performance: Visuals that communicate value and intention more clearly correlate directly with improved click-through rates and sales.

Yet, despite this outsized importance, the creation of these critical visuals today remains overwhelmingly manual, repetitive, and often suboptimal:

• Manual Design: They are painstakingly built product-by-product in tools like Adobe InDesign.
• Fixed Rules: Some automation relies on rigid, pre-defined layout templates, which fail when product shapes or sizes change.
• Simple Grids: The default, fallback solution is often a simple grid, which wastes canvas space and lacks creative flair.

Creative professionals instinctively understand that composition is category-dependent. A group of sleek, tall perfume bottles requires a different compositional logic than a set of disparate cosmetic jars, or a stack of uniformly shaped beverage cans. This nuanced, category-specific intelligence is precisely where we see the largest, most immediate opportunity for meaningful, scalable value creation.

Our strategic departure from general-purpose generative AI is our focus on Layout Intelligence. We are not building a model to redraw, modify, or hallucinate product pixels. We are building a predictive model that operates purely in the domain of spatial logic.

Our output is not a new image file, but a set of geometric parameters:

(x1, y1, x2, y2) + z-index

In plain language, we predict the optimal:

• Placement: The exact canvas coordinates for each product.
• Scale: How large each product should appear relative to the others and the canvas size.
• Layering: The depth order (z-index) to ensure visually pleasing and non-overlapping arrangements.

This approach offers fundamental, non-negotiable guarantees:

The implementation of our Layout Intelligence is built on a robust, multi-stage architecture that integrates computer vision, machine learning, and deterministic safety checks.

1. Feature Extraction per Image

Every product image is analyzed upon ingestion, creating a rich feature profile that is stored and reused. This eliminates the need to re-analyze assets for every new campaign.

We compute two main classes of features:

• Visual Embedding: A high-dimensional numerical representation of the product’s visual style and content, generated using industry-standard, pre-trained vision encoders (e.g., ResNet or CLIP). This allows the model to understand visual similarity and category.
• Shape Features: Geometric data derived from the product’s silhouette (pack shot):
  • Aspect Ratio: Width-to-height relationship.
  • Visual Mass: The density of the product within its bounding box.
  • Bottom Anchor: The position of the base for baseline alignment.

Tight Bounding Box: The smallest rectangle that perfectly encloses the product shape.

2. Training data from real banners

The core strength of our model is that it learns from quality conceived and executed by human effort. We extract structured layout data directly from our customers’ historical, approved InDesign files.

Each historical banner is deconstructed into a structured training example, capturing:

• The dimensions of the Canvas Size.
• The Target Bounding Boxes and Product Positions.
• The Layer Order (z-index) established by the designer.

This ensures the model learns the real-world aesthetic, constraints, and creative intelligence of the customer’s own brand guidelines. It does not learn from synthetic or generic industry datasets.

3. A Layout model per team

This is a critical, foundational principle of Relayter’s AI strategy: Each Layout Model is trained exclusively on that team’s approved content.

This strict isolation strategy ensures:

• No Cross-Customer Data Sharing: Complete data privacy and security.
• No Style Blending: The model only generates layouts consistent with the client’s unique brand aesthetics; there is zero risk of “style drift” or blending with a competitor’s visual language.
• Full Data Sovereignty: Your creative layouts remain your proprietary intellectual property.

For the technological foundation, we utilize a dedicated platform for the management, training, and versioning of a high-performance layout model specifically for each customer’s environment. This ensures your layouts remain secure, versioned, and exclusively yours.

4. Context-aware set modeling

The model’s core intelligence lies in its ability to evaluate the entire set of products as a cohesive group, rather than treating them as independent items. We employ a permutation-invariant set architecture, enabling the model to understand:

• Relative Scale Relationships: How large product A should be relative to product B in this specific context.
• Visual Balance: Ensuring the composition is centered and weighted correctly on the canvas.
• Depth Ordering: Predicting the optimal layering to maximize legibility and visual depth.
• Category-Specific Patterns: For example, learning that beverages are often aligned on a single baseline, while small make-up items tend to cluster in tight groupings.

These highly nuanced compositional rules are learned from your historic executions, not coded as brittle, hard-coded constraints.

5. Deterministic safety layer

After the layout intelligence model generates its optimal prediction, a final constraint engine acts as a non-negotiable safety layer. This ensures that the generated output is always technically valid and ready for production:

• No Out-of-Bounds Placement: Products are confined to the designated canvas area.
• Correct Aspect Ratios: Prevents accidental stretching or squashing of the product images.
• Preserved Pixel Integrity: Final check to ensure all geometric parameters are honored.

The result is a composition that is simultaneously creative, brand-compliant, and technically flawless.

We have intentionally chosen group image optimization as our starting point because it represents a focused, high-ROI problem that maximizes immediate visual improvement while minimizing brand risk.

• Disproportionate Visual Impact: It influences the aesthetic quality and visual hierarchy of the banner more than almost any other single element.
• Repetitive Manual Labor: It is a time-consuming, repetitive task that designers are eager to automate, freeing them for higher-value creative work.
• Reliable Learning: The geometric parameters are predictable and can be learned effectively and reliably from historical data.

This focused application contrasts sharply with broad generative AI, allowing us to deliver immediate visual uplift and efficiency gains without introducing the inherent risks of visual errors or brand inconsistency.

Roadmap Overview

The integration of Layout Intelligence into the Relayter platform transforms how visual campaigns are executed.

You will achieve the ability to:

• Instantly generate campaign-ready, high-quality group visuals at the point of publication item generation.
• Maintain your unique layout style, ensuring every automated design adheres to your established brand guidelines.
• Keep full control of your data, with guaranteed data isolation and security.
• Avoid cross-brand training risks, eliminating the risk of aesthetic contamination from external data.
• Significantly improve the overall visual impact of your Banners at Scale, ensuring visual excellence across all promotions.

AI in retail marketing should not aim to replace the core creative expertise of your design team. It should scale it. Group image intelligence is our concrete first step in building marketing automation that not only functions efficiently but delivers visual content that is consistently exceptional.