Cleverman
New York
Grooming / Personal Care / E-Commerce
AWS EC2 (GPU) Amazon S3 AWS Bedrock Flux Kontext IP Adapters ComfyUI FastAPI PyTorch Streamlit GCP Storage Juggernaut SDXL DreamShaper
Cleverman
New York
Grooming / Personal Care / E-Commerce
AWS EC2 (GPU) Amazon S3 AWS Bedrock Flux Kontext IP Adapters ComfyUI FastAPI PyTorch Streamlit GCP Storage Juggernaut SDXL DreamShaper
Cleverman is a New York-based personal care and grooming brand helping customers make more confident purchase decisions through technology. To power a personalized virtual try-on experience, Cleverman needed a productiongrade AI pipeline capable of visualizing hair and beard color changes across multiple angles, without sacrificing photorealism or identity accuracy. Avahi built a scalable GenAI pipeline using Flux Kontext with IP Adapters and a novel Image RAG approach, achieving perfect 5/5 identity preservation scores and processing speeds 8x faster than traditional diffusion models. The result is a seamless, customer-facing experience that drives purchase confidence and sets Cleverman apart in an increasingly competitive grooming market.
Cleverman is a New York City-based grooming and personal care brand focused on delivering personalized experiences to its customers. Operating at the intersection of beauty, e-commerce, and emerging technology, Cleverman’s platform enables consumers to explore and purchase grooming products tailored to their individual needs. With a lean, innovation-driven team, Cleverman has made AI-powered personalization a core pillar of its customer experience strategy, recognizing that helping customers visualize outcomes before purchase is one of the most effective ways to build trust and drive conversions.
For a grooming brand selling hair and beard color products, the purchase decision is inherently visual, and inherently uncertain. Customers are being asked to commit to a product based on a label swatch or a generic model photo that looks nothing like them. The gap between “what will this look like on me?” and what’s actually shown on a product page is a persistent source of hesitation, cart abandonment, and post-purchase returns.
Cleverman recognized that a personalized, AI-powered virtual try-on experience could close that gap, but executing it at production quality was far more technically demanding than it appeared. The challenge wasn’t simply recoloring a photo. It required generating consistent, photorealistic color transformations across eight distinct angles per customer image while maintaining precise identity preservation, accurate color matching, and clean mask precision. Any degradation in facial or hair detail — even subtle — would undermine customer trust rather than build it.
Early attempts using off-the-shelf image generation models revealed a critical limitation: existing tools produced identity drift, color inaccuracies, and inconsistent results that required significant manual touch-up. Deploying that kind of pipeline in a customer-facing product was not viable. Without a solution that could meet the quality bar required for a real commerce environment, Cleverman faced lower conversion rates, higher return volumes, mounting manual editing costs, and a growing competitive disadvantage as AI-powered try-on experiences became an expected feature in the beauty and grooming space.
AWS provided the compute backbone and AI infrastructure necessary to run a demanding, GPU-intensive generative AI workload at production scale. The processing requirements for multi-angle image transformation — particularly with high-fidelity diffusion-based models — made purpose-built GPU instances on AWS EC2 the right foundation, offering the flexibility to scale compute resources to workload demands without over-provisioning.
AWS Bedrock further extended the solution’s capabilities by providing access to managed foundation models within the same cloud environment, enabling Avahi to integrate additional AI capabilities without introducing external dependencies. Running the pipeline on AWS also aligned with established security and access control practices, with IAM roles governing resource access and keeping customer image data within a governed, controlled environment throughout the workflow.
Cleverman’s challenge required a partner with deep expertise in applied generative AI, specifically in computer vision, diffusion model architecture, and the nuanced problem of identity-preserving image transformation. Off-the-shelf solutions had already proven insufficient. What was needed was a team capable of engineering a custom pipeline from the ground up, with rigorous model evaluation methodology to back every technical decision.
Avahi brought exactly that. Rather than defaulting to a single model or approach, Avahi conducted a structured four-model evaluation across five test cases with quantified scoring across five performance dimensions before selecting the final architecture. This data-driven methodology, evaluating Flux Kontext, Juggernaut SDXL, DreamShaper, and alternatives against measurable benchmarks, gave Cleverman confidence that the chosen solution was genuinely the best available option, not simply the most convenient one.
Avahi’s ability to combine cutting-edge AI techniques (flow-matching, Image RAG, context-aware masking) with a scalable cloud architecture and a clean, handoff-ready codebase made them the right partner for a project where both engineering quality and client independence at the end of the engagement were non-negotiable requirements.
Avahi designed and built a production-ready GenAI image transformation pipeline purpose-built for Cleverman’s multi-angle virtual try-on use case.
Model Selection Through Rigorous Evaluation Before writing a single line of production code, Avahi ran a structured evaluation of four candidate models: Flux Kontext, Juggernaut SDXL, and DreamShaper. An additional alternative, across five test cases and five scored dimensions: visual quality, identity preservation, color accuracy, mask precision, and processing speed. This evaluation framework produced clear, quantified results that justified the final architecture selection and gave Cleverman a defensible, data-backed foundation for the technology choice.
Core Pipeline Architecture The production pipeline flows from a Streamlit-based UI, where images are uploaded and color selections are made, through a FastAPI backend that orchestrates ComfyUI workflows running on AWS EC2 GPU instances. Flux Kontext’s flow-matching architecture handles the core image transformation, replacing traditional iterative diffusion sampling with a more direct, efficient approach that dramatically reduced processing time. Transformed images are stored in GCP Storage, with the full pipeline designed for batch workflow automation to support scale.
Identity Preservation via Flux Kontext The hardest technical problem in the entire engagement was preventing identity drift during color transformation, the subtle but damaging degradation of facial features, hair texture, and individual characteristics that made earlier model attempts unsuitable for production. Flux Kontext’s context-aware editing architecture addressed this directly, achieving a perfect 5/5 identity preservation score across test cases and outperforming all evaluated alternatives by two to three points on this dimension alone.
Color Accuracy via Image RAG with IP Adapters Achieving precise, consistent color matching, particularly across eight different viewing angles, required a novel approach. Avahi implemented Image RAG (Retrieval-Augmented Generation for images) using IP Adapters, enabling the pipeline to reference and match target color characteristics with high fidelity rather than relying solely on text prompts or generalized model behavior. This technique produced a 4/5 color accuracy score and was central to delivering multi-angle consistency that held up under real-world product conditions.
Preprocessing, Masking, and Automation A dedicated preprocessing pipeline handled mask precision, isolating hair and beard regions with context-aware editing to ensure transformations applied cleanly without bleeding into skin, background, or other image elements. Batch workflow automation was built into the architecture from the start, enabling Cleverman to process multiple images and angles efficiently without manual intervention at each step, and positioning the platform for scale as their customer base grows.
Avahi delivered a production-grade GenAI pipeline that transformed Cleverman’s virtual try-on capability from a conceptual goal into a scalable, customer-facing product feature. By solving the identity preservation problem that had blocked earlier attempts — and pairing it with accurate color matching and clean mask precision across eight angles — the solution gives Cleverman’s customers a genuinely reliable, photorealistic preview of how a product will look on them before they buy. That capability directly addresses the conversion and returns challenges that motivated the project in the first place.
The architecture was built for independence — with full documentation, knowledge transfer, and a clean codebase — ensuring Cleverman’s team can maintain, extend, and scale the solution without ongoing external dependency.
Outcome Highlights:
Let’s explore your high-impact AI opportunities together in a complimentary session
Let’s explore your high-impact AI opportunities together in a complimentary session
Let’s explore your high-impact AI opportunities together in a complimentary half-day session
