Dianthus uses cutting edge AI technology to grow D2C brands – applying hyper-personalized marketing tactics and new AI-centric workflows and processes, Dianthus is the world’s only AI-first ecommerce company.


THE OPPORTUNITY
“As an AI-first company, Dianthus seeks out fellow deep learning researchers who can offer unique experience and perspective in applying and implementing cutting-edge solutions in AI and ML to partner with our in-house AI teams.”– Chris Litster, CEO
“AI development for unique products can be a risky endeavor without a disciplined approach to experimentation, broad domain knowledge and the ability to apply the latest research and architectures where needed. Neu.ro’s combination of PhD-level AI expertise and an extremely robust foundation in MLOps gave us the confidence that our project would be handled both innovatively and responsibly.”– Rob May, Co-Founder and CTO
“Neu.ro went above and beyond in terms of identifying and implementing cutting edge deep learning techniques to create a genuinely innovative solution. We look forward to continuing to advance the state of the art in the area of AI-generated visual marketing with them.”– Rob May, Co-Founder and CTO
D2C ecommerce has seen explosive growth in recent years, with Shopify GMV growing from $7.7bn in 2015 to $175bn in 2021. McKinsey recently called D2C “the best opportunity for innovative brands to build direct relationships with their customers”.
As ecommerce brands increasingly leverage predictive models to identify customer needs and behaviors at scale, they have been able to implement solutions for customer experience and service that incorporate real-time customer data through design, content and marketing. Dianthus is a leader in applying cutting edge AI to these issues and is so confident in its ability to grow D2C brands that it acquires or takes ownership stake in every brand it works with.
A proposed solution was to create a sophisticated AI computer-vision system that could generate unique photographs, including computer generated human or animal models against naturalistic backgrounds, that also incorporated the D2C product.
The Challenge
In order to achieve believable AI-generated product shots with digital influencers, a unique ML and data pipeline needed to be created to incorporate multiple processes, including:
- Background generation
- Identity generation
- 3d rendering
- Human positioning
- Product positioning
- Harmonization of all elements
This pipeline further needed to allow for results to progress through various stages of refinement to deliver a finished photo result that was attractive and natural-looking enough to share on social media.
A major challenge faced by Neu.ro was that different approaches were required for each of the elements involved in producing the finished product, including background generation, human face generation, human pose generation, influencer placement, product placement and camera angle selection.
Finding optimal approaches for each of these elements required an iterative process of experimentation and optimization.
An additional set of challenges was due to the high compute requirements for model training, selection and optimization. This required high performance GPUs to develop, train, iterate and optimize within the required time horizon. Additionally, there is the potential for this highly intensive process to have a negative impact on climate change via carbon emissions.
The ability to programmatically create compelling visual advertising assets for specific products and audience segments appropriate for social media channels is a holy grail for efficiently scaling brands online
100% of development and training was accomplished on Neu.ro’s 100% zero-emissions Green AI Cloud, powered by the latest NVIDIA GPUs
The Solution
Neu.ro set to work conducting research of possibly applicable techniques for dynamically composing unique images composed of people products and backgrounds.
Approaches tested during experimental phase included:
- CLIP guided optimization
- ESRGAN-based refinement
- CycleGAN
- MUNIT
- Diffusion model reprojection
- StyleGAN3 reprojection
- StyleGAN3 face recovery
- Face swap
- Contrastive Unpaired Translation
Toos used included:
- Neu.ro MLOps Platform
- PyTorch
- Tensorboard
- MLFlow
- Numpy/Scikit libraries
Real image datasets from commercially available sources were combined with synthetically generated images via Synthesis API with different camera views and configurations. Various rendering technologies were tested for quality, stability across platforms and ability to be automated at scale.
Additional research was conducted on reduction of manual input, including automated avatar and wardrobe creation and automated scene selection with product placement.
Furthermore, various personas were created and tested, including male, female, gen z and millennial figures.
The resulting solution incorporates technologies for human positioning, camera positioning, appropriate sizing of foreground and background elements, and product harmonization improvements, as well as numerous optimizations for speeding up the pipeline and limiting required human input into the process.