Generative AI for Virtual Try-On and Personalization

Table of Contents

1. Introduction: Macroeconomic Context and the Technological Imperative
   • The Returns Crisis in Fashion Retail: The Economics of Loss.
   • A Paradigm Shift: From Static Product Pages to Generative Experiences.
   • The Role of AI in Transforming the Customer Journey (CJM).
2. The Technological Phenomenon of Google Gemini: Nano Banana as a Business Tool
   • Genesis and Naming: From a Late-Night Commit to a Global Brand.
   • The Gemini 2.5 Flash Image Architecture: Speed, Multimodality, and Reasoning.
   • Retail Use Cases: Identity Preservation, Prompt-Based Editing, and Text Handling.
   • Access and Integration Considerations Under Russia’s Sanctions Constraints.
3. Flux.1: A Sovereign Alternative and the Open Source Ecosystem
   • Black Forest Labs’ Technical Breakthrough: A Hybrid Transformer-and-Diffusion Architecture.
   • Model Family: Pro, Dev, Schnell — Choosing the Best Business Solution.
   • Inpainting Toolkit: The Role of Flux Fill and Redux in the Virtual Try-On Process.
   • ComfyUI as the Operating System for Generative Retail: Node and Workflow Architecture.
4. Infrastructure Sovereignty: Deploying AI Solutions in Russian Cloud Environments
   • The GPU Hosting Landscape in Russia: A Comparative Analysis of Yandex Cloud, Selectel, and Cloud.ru.
   • The Economics of Compute: An Analysis of GPU Rental Costs (A100, V100, RTX 4090) and Cost Optimization.
   • A Technical Deployment Guide: Docker, Environments, and Access Configuration.
5. Virtual Try-On (VTO) Architecture: From Concept to Code
   • The VTO Technical Pipeline: Segmentation (SAM), Positioning (DensePose), and Generation.
   • Telegram Integration: Building an MVP with Python Bots and a WebSocket API.
   • Interaction Scenarios: UX/UI Best Practices for Reducing Friction.
6. Implementation Economics: ROI, Conversion, and Unit Economics
   • How VTO Affects Metrics: Higher Conversion Rate (CR), Lower Return Rate (RR), and Increased Average Order Value (AOV).
   • The Psychology of Digital Ownership: The “Try Before You Buy” Effect.
   • Payback Modeling: CAPEX vs. OPEX When Implementing Generative Services.
7. The Legal Framework: Compliance with Federal Law 152 and Biometric Data Handling in Russia
   • Data Classification: Is a Try-On Photo Considered Biometric Data?
   • The Regulator’s Position: An Analysis of Roskomnadzor Guidance and Case Law.
   • Technical and Legal Safeguards: Localization, Anonymization, and User Consent.
8. Beyond Try-On: The Ecosystem of End-to-End Personalization
   • Generative Content: Automating Description Creation and SEO Optimization (Mindbox, Retail Rocket).
   • Predictive Analytics and Next Best Offer: How AI Anticipates Customer Wants.
   • Virtual Photo Studios and Content Production Optimization.
9. Case Studies and Market Practice: Lessons from Industry Leaders
   • Russian Pioneers: An Analysis of Lamoda and Wildberries Experience.
   • Global Benchmarks: Lessons from Amazon, Zalando, and Luxury Brands.
10. Conclusion: A Transformation Roadmap for Russian Entrepreneurs

1. Introduction: Macroeconomic Context and the Technological Imperative

The current landscape of Russian e-commerce is defined by an unprecedented level of competition and pressure on operating margins. The period of rapid, broad-based growth driven by an influx of new internet users and pandemic-era shopping habits is giving way to a phase focused on efficiency, loyalty, and the LTV (Lifetime Value) of each customer. In this environment, traditional product presentation methods—static studio photos and standardized text descriptions—are rapidly losing effectiveness, turning from a competitive advantage into a basic hygiene factor.

The Returns Crisis in Fashion Retail: The Economics of Loss

One of the most acute problems for entrepreneurs in apparel, footwear, and accessories remains the high return rate. According to industry reports, the return rate in online channels reaches 30–40%, significantly higher than in traditional retail (8–10%). The main reason for returns, accounting for up to 70–80% of cases, is a mismatch between the size or fit and the customer’s expectations formed from photos on models.

The economic consequences of this phenomenon are devastating. Reverse logistics includes the costs of shipping, receiving, quality inspection, repackaging, and often markdowns on items that have lost their retail appeal. On a global scale, retailer losses from returns are estimated in the hundreds of billions of dollars annually—a figure that reaches $550 billion by some estimates. For Russian businesses operating under difficult logistics across vast distances and rising courier service rates, the returns problem is becoming a threat to financial stability.

A Paradigm Shift: From Static Product Pages to Generative Experiences

The answer to this challenge is a technological transformation of the user experience. We are seeing a shift from transactional interaction models (“choose — pay — receive”) to immersive and generative. Virtual try-on (VTO) technologies based on generative neural networks are becoming the core element of this transformation.

The virtual try-on market is showing explosive growth. If the global market was valued at $10.93–11.38 billion in 2024, it is projected to grow to $108.5 billion by 2034, at a compound annual growth rate (CAGR) of around 25.8–26.4%. The drivers of this growth are not only large corporations but also the accessibility of the technology for small and medium-sized businesses. The emergence of powerful open models such as Flux.1 and accessible APIs from tech giants such as Google (the Nano Banana model) has democratized access to tools that were previously available only to players like Amazon or Walmart.

The Role of AI in Transforming the Customer Journey (CJM)

Artificial intelligence is no longer just a “black box” for analytics; it is becoming an active participant in content creation. Generative models can create personalized images in real time, allowing users to see a product not on an abstract model, but on themselves or on a digital twin closely matched to their measurements. This fundamentally changes the psychology of purchasing by creating a sense of ownership before payment, which, according to research, increases conversion by 30–94% and reduces return rates by 30–40%.

In this report, we will conduct a deep analysis of how Russian entrepreneurs can implement these technologies, overcome technical and legal barriers, and use the power of generative AI to achieve a qualitative leap in business performance.

2. The Technological Phenomenon of Google Gemini: Nano Banana as a Business Tool

Genesis and Naming: From a Late-Night Commit to a Global Brand

In the world of enterprise technology, product names are usually the result of long marketing research and focus groups. However, the story of the Nano Banana model is an example of how chance and viral momentum can determine a product’s success. Officially known as Gemini 2.5 Flash Image, this model was developed by Google DeepMind as a high-speed tool for image generation and editing.

The story behind the name began in July–August 2025, when the Google team was preparing the model for anonymous testing on the platform LMArena (Large Model Arena), a crowdsourced platform where users blindly compare the output of different neural networks. To submit the application, they needed a code name that would not reveal Google’s involvement. At 2:30 a.m., under deadline pressure, product manager Naina Raisinghani suggested the name "Nano Banana." It was a wordplay based on her own nicknames: friends called her "Naina Banana," and colleagues called her "Nano" because of her small stature and love of technology. The spontaneous decision was made.

When the model appeared on LMArena, it caused a sensation with the quality of its generation and editing capabilities, quickly taking the top spot in the rankings. Social media users picked up the funny name, and "Nano Banana" became a viral meme associated with cutting-edge generation quality. Google, showing flexibility, embraced the grassroots name by adding a banana emoji to the AI Studio interface and even releasing merch. Later, with the launch of Gemini 3 Pro Image, the brand evolved into Nano Banana Pro. For business, this case matters as an indicator: behind the "playful" name is one of the most powerful technologies on the market, which earned community recognition even before its official release.

Gemini 2.5 Flash Image Architecture: Speed, Multimodality, and Reasoning

From a technical standpoint, Nano Banana (Gemini 2.5 Flash Image) represents a major step forward compared with classic diffusion models. It is built on an architecture that integrates language understanding and visual context at a deep level.

1. Reasoning: The model does not just match pixels to text; it "understands" scene physics, lighting, and the spatial relationships between objects. This allows it to handle complex editing instructions, such as "move the object to the left and add a shadow consistent with the sunset light."
2. Speed (Flash): Model optimization for low latency makes it possible to generate images in seconds rather than minutes. This is critical for e-commerce use cases, where users are not willing to wait long for a try-on result.
3. High resolution and upscaling: Support for native 2K output and upscaling to 4K delivers quality sufficient for printing and professional use in marketing materials.

Retail Use Cases: Identity Preservation, Prompt-Based Editing, and Text Rendering

For virtual try-on tasks, Nano Banana offers a set of unique capabilities:

• Subject Identity Preservation: This is the holy grail of VTO. The model can preserve a person’s facial features, body type, and appearance characteristics while changing their clothing or surroundings. This solves the uncanny valley problem and makes try-on more personalized. The user sees themselves in the new outfit, not a lookalike mannequin.
• Intelligent editing (Inpainting/Outpainting): The ability to change parts of an image using natural language ("replace the jeans with a pencil skirt," "add a hat"). The model accurately understands object boundaries and blends new elements seamlessly.
• Multi-image fusion: The ability to combine several reference images. For example, you can provide a user photo and a product photo (packshot), and the model will generate a realistic image of the user wearing that product.
• Text generation: Unlike many competitors, Gemini 3 Pro (Nano Banana Pro) handles readable text generation on images exceptionally well. This opens up opportunities for creating branded content, ad banners, and mockups with text.
• Selfie-based try-on: A recent update made it possible to generate full-body images from a selfie. The model reconstructs the body and clothing, significantly lowering the entry barrier for users who don’t have a high-quality full-length photo on hand.

Access and integration challenges under Russian sanctions restrictions

Using Google Cloud services and the Gemini API in Russia comes with serious limitations. Direct access to gemini.google.com and Google AI Studio is often geo-blocked, and payment with Russian cards is not possible. Nevertheless, there are workable integration strategies for Russian entrepreneurs:

1. Network access: Using VPN services and proxy servers with US or European IP addresses is a basic requirement. For stable API operation in production, it is recommended to use corporate VPN gateways or deploy proxies on rented VPSs in friendly jurisdictions.
2. Payments and billing: Google AI Studio offers a generous Free Tier that lets you test hypotheses and build MVPs without linking a card. For scaling (pay-as-you-go), you need to use foreign bank cards or rely on account resellers. There are also specialized intermediary services (API gateways) that provide access to Google models through their own interface, accepting payment in rubles or cryptocurrency.
3. Risks: The main risk is a possible account ban by Google for violating the terms of use (bypassing regional restrictions). Therefore, building a business exclusively on the basis of Nano Banana can be a strategically vulnerable decision. This brings us to the need to consider open-source alternatives.

3. Flux.1: A Sovereign Alternative and the Open Source Ecosystem

Under conditions of uncertainty around access to Western proprietary APIs, open-source models (Open Weights) are becoming a strategic choice for Russian business. The leader in this area in 2024–2025 was Flux.1developed by Black Forest Labs, founded by former Stability AI employees.

Black Forest Labs’ Technical Breakthrough: A Hybrid Transformer and Diffusion Architecture

Flux.1 is a new generation of generative models using a hybrid architecture. It combines the advantages of transformers (used in LLMs for context understanding) and diffusion models (responsible for generating visual details). With 12 billion parameters, Flux.1 can generate images with exceptional detail, precise adherence to complex prompts, and a high degree of photorealism, outperforming many closed models.

Model family: Pro, Dev, Schnell

Flux.1 is available in three versions, allowing businesses to choose the right tool with flexibility:

• Flux.1 [pro]: A closed version available only through the API. It delivers the highest quality but carries the same risks as Gemini.
• Flux.1 [dev]: An open-weights version distilled from the Pro version. It offers high quality and efficiency, but the license restricts its use to non-commercial purposes (careful review of the terms is required for business use).
• Flux.1 [schnell]: An optimized, "fast" version (up to 10x faster than Dev). Distributed under the Apache 2.0 license, which allows unrestricted commercial use. It is the ideal candidate for building your own VTO services.

Inpainting Toolkit: The Role of Flux Fill and Redux

To deliver high-quality virtual try-on, it is not enough to simply generate images. Specialized tools are needed for editing and style transfer. The Flux ecosystem offers powerful solutions:

1. Flux Fill: A model specifically fine-tuned for inpainting (filling in areas) and outpainting (extending images). It lets you mask the clothing area in a user's photo and generate new apparel that fits lighting and body geometry perfectly.
2. Flux Redux: An adapter (similar to IP-Adapter in the Stable Diffusion ecosystem) that lets you use an image as a prompt. This is critical for VTO: instead of a text description like "red dress," the system takes a photo of the actual product as input, and Flux Redux transfers its visual characteristics (color, texture, cut) into the generation.
3. CatVTON: Specialized pipelines and wrappers for Flux, such as CatVTON, automate the clothing overlay process while ensuring high fidelity to product details.

ComfyUI as the operating system for generative retail

The de facto standard for working with Flux and building complex generative pipelines has become ComfyUI. It is a node-based graphical interface that lets you visually build image-processing logic without writing low-level code.

A typical VTO workflow in ComfyUI includes:

1. Load Image: Load the user and product photos.
2. Segmentation: Using models such as Segment Anything (SAM) or RMBG to automatically isolate clothing in the user's photo and create a mask.
3. Pose Estimation: Using DensePose or OpenPose to determine the person's pose so the new clothing fits anatomically correctly.
4. Generation: Flux Fill and Redux nodes generate the image within the masked area using product features.
5. Post-processing: Nodes for color correction and face restoration (Face Restore), if needed.

The advantage of ComfyUI is that the workflow you create can be saved in JSON format and used through an API, turning a visual diagram into a backend for a web service or bot.

4. Infrastructure Sovereignty: Deploying AI Solutions in a Russian Cloud

Using open-source models such as Flux.1 requires substantial computing resources, primarily GPUs. For Russian businesses, the optimal solution is to rent capacity from local cloud providers. This solves two problems at once: ensuring high performance and complying with data localization requirements.

GPU Hosting Landscape in Russia: A Comparative Analysis

Three key players stand out in the Russian market, providing GPU infrastructure for ML tasks:

1. Yandex Cloud (DataSphere) Yandex offers a mature ecosystem for ML development. The DataSphere service lets you run Jupyter Notebooks on powerful GPUs without needing to administer servers.

• Hardware: NVIDIA V100 (32GB) and A100 (80GB) cards are available.
• Features: Per-minute billing, which is cost-effective for batch processing (running the server only while the order queue is being processed). Convenient integration with S3 object storage for storing photos.
• Pricing: V100 usage starts at approximately $0.59 - $0.87 per hour (depending on vCPU and RAM configuration); A100 is significantly more expensive. Grants are also available for new users.

2. Selectel A provider specializing in bare metal servers and cloud solutions.

• Hardware: A wide range of cards, from consumer RTX 3090/4090 to professional A100/H100.
• Features: The ability to rent the hardware outright delivers maximum performance without virtualization overhead. Flexible server configurations.
• Pricing: Competitive pricing, especially for consumer-grade cards that are great for Flux inference (for example, the RTX 4090). Hourly billing and fixed monthly rental plans are available.

3. Cloud.ru (formerly SberCloud) The ML Space platform from Sber provides access to Christofari supercomputers.

• Hardware: Powerful NVIDIA A100-based clusters.
• Features: Focused on large enterprise workloads and model training. High scalability.
• Pricing: Billing can be complex, but there are attractive offers for long-term use.

Technical Deployment Guide

Deploying ComfyUI with Flux on a remote server (for example, in Yandex Cloud or Selectel) includes the following steps:

1. Preparing the environment: Rent an instance with a GPU (at least 24GB VRAM for comfortable Flux work, although optimized versions also run on 12-16GB). Install NVIDIA drivers (CUDA) and Docker.
2. Downloading model weights: Download the Flux.1 model files (schnell or dev), VAE, CLIP, and adapters (Redux/Fill) from Hugging Face. In Yandex Cloud, it is convenient to use S3 buckets to mount the folder with the models to the instance so you do not have to re-download them each time.
3. Launching ComfyUI: It is recommended to use a Docker container.
4. Bash

git clone https://github.com/comfyanonymous/ComfyUI.git
pip install -r requirements.txt
python main.py --listen 0.0.0.0 --port 8188

1. Key --listen 0.0.0.0 allows external access to the interface.
2. Security: Configure a firewall to restrict access to port 8188 or use SSH tunneling for a secure connection.

5. Virtual Try-On (VTO) Architecture: From Concept to Code

Building your own virtual fitting room is an integration task that connects the user interface with the power of neural networks.

VTO Technical Pipeline

The request processing flow for a try-on looks like this:

1. Data Ingestion: The system receives the user's photo (User Image) and the clothing photo (Garment Image).
2. Preprocessing (ComfyUI Workflow):
   • Segmentation: The neural network (for example, Node SegmentAnythingUltra) analyzes the user's photo and creates a mask for the area where the old clothing is located (for example, "torso", "legs").
   • Pose analysis: Node DensePose determines body position to understand perspective and body curves.
3. Generation (Inpainting):
   • Flux Fill: Takes the source image, the mask, and the encoded features of the new clothing. It performs generation (denoising), replacing the pixels in the mask with new ones that match the style of the Garment Image.
   • Flux Redux: Works together with Fill, ensuring the texture and details of the new clothing are as close as possible to the reference.
4. Postprocessing: Final color correction to harmonize the lighting of the new clothing and the background.

Integration via Telegram: Building an MVP

For the Russian market, a Telegram bot is an ideal channel for an MVP (Minimum Viable Product). It's inexpensive, fast, and familiar to users.

Stack: Python + python-telegram-bot (or aiogram) + websocket-client.

Bot workflow:

1. The user sends the /start command and uploads their photo. The bot saves the chat_id and file_id of the photo.
2. The user selects a product from the catalog (Telegram in-app Web App or inline buttons).
3. The bot creates a JSON request containing the ComfyUI workflow structure and inserts the paths to the user's photo and the product photo.
4. Backend Python script establishes a WebSocket connection to the ComfyUI server:
5. Python

ws = websocket.WebSocket()
ws.connect("ws://{}/ws?clientId={}".format(server_address, client_id))

1. The script sends the JSON prompt to the ComfyUI API (POST /prompt) and listens on the WebSocket for progress messages.
2. When generation is complete, the script downloads the finished image (GET /view) and sends it to the user in Telegram using the sendPhoto method.

This approach makes it possible to launch a fully working service in just a few days of development, using ready-made Telegram nodes in ComfyUI or by writing your own Python middleware.

6. Implementation Economics: ROI, Conversion, and Unit Economics

Implementing VTO is not just a marketing move, but an investment with a clear financial rationale.

VTO Impact on Metrics

Market data and implementation case studies show a significant impact on key E-commerce metrics:

ROI Modeling (Unit Economics)

Let’s consider a hypothetical clothing store with 5,000 orders per month.

• Average order value: RUB 5,000
• Current return rate: 30% (1,500 returns).
• Return cost (logistics + processing): RUB 500
• Losses from returns: RUB 750,000/month

VTO implementation scenario:

• 30% reduction in returns: Savings of RUB 225,000/month
• 10% increase in conversion: Additional revenue of ~RUB 2.5 million (at the same traffic level).
• VTO costs: GPU server rental in Selectel/Yandex (~RUB 30,000–50,000/month) + development amortization.

Bottom line: Even with conservative estimates, the investment pays for itself through logistics savings on returns alone, and the higher conversion becomes pure incremental profit.

7. Legal Framework: Compliance with 152-FZ and Working with Biometrics in Russia

Implementing technologies that work with face images requires strict compliance with Russian law, especially 152-FZ, “On Personal Data.”

Data Classification: Is a Photo Personal Data or Biometrics?

This is a key legal nuance.

• Personal data: A photo on its own is personal data. Processing it requires consent from the data subject.
• Biometric personal data: According to clarifications from Roskomnadzor, a photo becomes biometrics only if it is used to establish a person’s identity (identification) of a citizen.

For VTO: If the system uses the photo solely to overlay the clothing texture (“processing”), does not store biometric vectors, does not match the face against a database, and does not attempt to determine the user’s full name, then such processing is generally not considered biometric processing in the strict sense that would require transmitting data to the Unified Biometric System (EBS). This significantly reduces the regulatory burden.

Practical recommendations

1. Consent: The interface (bot or website) must include a checkbox: “I consent to the processing of my image for the purposes of virtual try-on.”
2. Data localization: Processing servers (GPU) must be physically located in Russia (a requirement for localizing databases of Russian citizens). Using Yandex Cloud or Selectel addresses this issue.
3. Data lifecycle: Do not store user photos longer than necessary for the generation session. Delete the source files and results immediately after sending them to the user. This minimizes the risk of leaks and complaints.

8. Beyond Try-On: An Ecosystem for Full-Scale Personalization

Generative AI is transforming not only try-on, but also the retailer’s entire content marketing operation.

Generative content and SEO

Platforms like Mindbox and Retail Rocketas well as specialized tools, use LLMs to generate unique product descriptions. This makes it possible to:

• Automatically create keyword-rich text for thousands of SKUs, improving SEO rankings.
• Adapt the tone of voice of descriptions for different audience segments (for example, “bold” for a youth collection, “restrained” for classics).

Predictive analytics and Next Best Offer

Russian CDP platforms (Customer Data Platforms) integrate AI to predict behavior.

• Retail Rocket uses algorithms to generate product recommendations in real time (“Next Best Offer”). AI analyzes consumption patterns (for example, buying children’s clothing in different sizes over time) and suggests items the customer is likely to need in the future.
• Mindbox uses ML to optimize send times for campaigns and build personalized product selections, increasing open rates and revenue. Integrating VTO data (which style “worked” virtually) into these systems will make hyper-personalized offers possible.

Virtual photo studios

Wildberries has already launched a “Virtual Photo Studio” tool that lets sellers generate product photos on models of different types without real photo shoots. This reduces production costs (photographers, studios, models) and speeds up time to market.

9. Case Studies and Market Practice: What the Leaders Are Doing

• Lamoda: One of the pioneers of AR try-on in the CIS. Implementing virtual shoe try-on (Wanna SDK) led to a 5% increase in category sales and improved purchase completion rates. About 100,000 people use the feature every month. The company also experimented with interactive billboards in Moscow.
• Wildberries: Actively implementing generative networks for content creation and virtual try-on, providing tools to millions of sellers.
• Zalando (Europe): Created an LLM-based virtual assistant that helps with size and style selection, reducing returns by 10%.
• Amazon: Uses generative AI to personalize product descriptions, highlighting features that matter to a specific user (for example, “gluten-free” or “suitable for running”).

10. Conclusion: A Transformation Roadmap

For Russian business owners in 2025–2026, adopting generative AI is a way to make a major leap in efficiency and customer service. These technologies are no longer limited to giants and are now accessible to small and midsize businesses.

Strategic steps:

1. Step 1: MVP on Telegram. Launch a bot with an API to ComfyUI (Flux). It’s fast, inexpensive, and lets you gather feedback.
2. Step 2: Infrastructure. Move to local GPU servers (Selectel/Yandex) to ensure stability, speed, and legal compliance.
3. Step 3: Data integration. Connect VTO data with CRM/CDP (Mindbox/Retail Rocket) to create a unified customer profile and hyper-personalized communications.
4. Step 4: Scaling. Roll out VTO in the web storefront and mobile app, using the experience and data you’ve already built.

The market is moving toward a hybrid experience where digital and physical realities merge. Those who adopt these technologies today will be market leaders tomorrow, while those falling behind will continue paying the logistics costs of endless returns.