GENERATIVE AI MEETS BRANDED CONTENT: A FIELD TEST REVIEW

Presented by Undercurrent AI, Advanced R&D and Ethical Consulting Partner

Executive Summary

Generative AI is exploding—and much of it remains speculative. In advertising, the stakes are higher than ever: more formats, faster timelines, and rising demands for content that scales without losing brand integrity. With so much hype surrounding AI’s creative potential, Undercurrent AI partnered with RiTE Media to test what this technology can actually deliver in a commercial context.

From late 2023 through early 2025, RiTE and Undercurrent XR has conducted real-world tests around separate use cases of AI’s capabilities in producing branded content, and came together in 2024 to collaborate on the latest use case presented here. Their goal was to determine whether a custom-trained LoRA (Low-Rank Adaptation) model could match or enhance traditional production methods—while maintaining legal and brand safety standards. The result: while 80% of AI-generated outputs fell short, 20% showed tangible, scalable value. This review outlines UndercurrentXR+AI methods, findings, tools, and takeaways that RiTE and Undercurrent discovered throughout this recent field test for agencies, brands, filmmakers, and technologists to glean from and develop insights to their own Generative AI journey.

Video summary of the AI-assisted content workflow.

Objectives

Test whether LoRA-trained diffusion models can emulate a traditionally photographed food commercial.
Explore AI’s capacity to generate high-volume deliverables from a minimal base of owned assets.
Evaluate the cost, quality, and control trade-offs between traditional and AI-assisted workflows.
Assess legal, ethical, and IP compliance standards around custom-trained models.
Highlight the emerging convergence of AI with cinematographic motion and framing control.

Methodology & Workflow

Phase 1 – Traditional Production

The process began with traditional production to establish a high-fidelity benchmark and proprietary training data:

Shot a tabletop burger commercial using practical food styling, lighting, and photography.
Captured dozens of high-resolution stills across multiple angles and lighting conditions.
These stills served as the foundation for AI training.

A professional food photographer setting up studio lighting in a kitchen for the burger photoshoot. — The Real Burger photoshoot for AI data input. Image courtesy of RiTE.

Which LoRA Base Model? (Selection Criteria)

SD 1.5

Pro: Widely adopted, good community support.
Con: Less photorealistic, requires advanced prompting.

SD XL

Pro: Enhanced realism, handles complex scenes well.
Con: Resource-heavy, potential compatibility issues.

FLUX (Selected for Test)

Pro: Excels at photorealism, needs fewer images (250-500).
Con: Highly resource-heavy, potential compatibility issues.

Phase 2 – Model Training & Generation

Trained a custom LoRA model using exclusively client-owned assets (no internet-sourced imagery).
Collaborated with Undercurrent AI to guide technical implementation.
Employed tools including ControlNet, IP Adapter, Depth-to-Image, and ComfyUI for enhanced fidelity and control.
Generated over 500 images across four different models.

A quick, abstract visual showing AI-generated elements resembling data or a fast, blurry movement. — GenAI skateboard teaser. Image courtesy of RiTE.

A clean, professionally rendered, rotating burger and fries on a white plate against a black background. — AI Modeled Real Burger. Image courtesy of RiTE.

Phase 3 – Motion Simulation

Captured camera motion using an iPhone and motion tracking software such as SynthEyes.
Applied this data to simulate camera movement in the AI generation pipeline.
This approach is nearing a phase of practical implementation for real-world commercial use.

A person using an iPhone to capture motion tracking data over a kitchen counter, focusing on a burger. — Orbiting the AI Modeled Real Burger with smartphone. Image courtesy of RiTE.

Key Findings (Strategic Outcomes)

Legal Review & IP Safety

All assets used for training were owned or created in-house.
The workflow was developed in tandem with legal counsel to ensure compliance with current IP and AI regulations.
No third-party datasets or scraped internet content were used.

1. Efficiency & Scalability

LoRA training was completed in a matter of hours. Assets were generated in 9:16, 1:1, and 16:9 without requiring reshoots, allowing creative teams to test new looks and variants rapidly.

2. Cost Implications

There was a significant reduction in cost for generating format-specific variations (especially social cutdowns). However, traditional shoots are still required for foundational product fidelity.

3. Quality, Motion, & Control

AI-generated imagery often degraded under close scrutiny (e.g., textures, food physics), struggling with realism. Conversely, motion simulations using SynthEyes showed promise, enabling precise camera control and motion consistency, which is vital for high-end commercials.

The 80/20 Reality (Usable Output)

Over 500 images were generated using four model variations. Roughly 80% of the outputs were unusable without significant cleanup.

The remaining **20%** delivered high-fidelity outputs that were:

Consistent with brand look and feel
Rendered in multiple aspect ratios
Legally clean and copyright-safe

These assets proved invaluable for supplemental deliverables, social content, and R&D previsualization.

Recommendations for Creative Teams & Clients

The core takeaway is that AI is not ready to replace human production—but it is ready to augment it in powerful, flexible ways.

Use AI as a post-production multiplier—not a full replacement for craft.
Train on your own assets to protect brand integrity and IP.
Combine traditional production and AI generation for maximum control.
Engage legal counsel early in pipeline design.
Begin testing AI motion and framing control tools for use in R&D and early-stage creative visualization.