Mayank Gaur
Development
New Delhi , India
Skills
Machine Learning (ML)
About
Mayank gaur's skills align with Programmers (Information and Communication Technology). Mayank also has skills associated with Engineering Managers (Engineering). Mayank gaur appears to be a low-to-mid level candidate, with 2 years of experience.
Work Experience
Virtual Try On
January 2024 - February 2024
- Developed a proprietary pipeline for B2B applications to generate high-quality, aesthetically pleasing virtual identitity datasets, addressing copyright concerns and ensuring loyalty-free usage without the need for individual photoshoots. • Implemented a data generation pipeline utilizing FaceFusion for face swapping and enhancement, coupled with CCSR, a diffusion-based upscaler, to upscale features of a source image representing a high-quality virtual identity, initially inferred from midjourney. Applied this upscaled image to a dataset containing a singular identity, resulting in the creation of a new dataset with consistent and coherent virtual identities across various poses. • Incorporated the outlined approach into a stable pipeline, resulting in the development of over 25 metamodel datasets of diverse ethnicities. These datasets are utilized to train models within the Virtual Try-On pipeline, providing B2B clients with a wide array of identity options to choose from.
Virtual Try On
March 2023 - Present
- Developed and implemented a Stable Diffusion XL (SDXL)-based pipeline to generate photo-realistic, full-body human images wearing selected attire. • Designed the pipeline to cater to both B2B and B2C markets, offering B2B clients cost-effective alternatives to traditional photoshoots and enabling B2C customers to engage in rapid and personalized virtual try-on experiences. • Conducted training of ControlNet, a Stable Diffusion-based model, with a well thought condition image as condition, for precise generation of human figures in various poses and clothing. Utilized a dataset comprising 600,000 web-scraped images. • Applied techniques such as LoRA (Low-Rank Adaptation) optimization and dreambooth to fine-tune SDXL's attention layers, resulting in the creation of consistent and distinct identities with minimal input data (20-30 images). Produced 80 high-quality models tailored for B2B applications, while establishing a foundation for B2C personalization.
Virtual Try On
July 2022 - March 2023
- Pioneered the development of a DL-based image-warping solution, tailored to transform fashion e-commerce imagery by seamlessly transitioning garments from dummies or standard models to premium models, ensuring photorealism and cost efficiency. • Created a matching module using 3D keypoints and DensePose to identify the optimal match from premium model images, ensuring precise model replacement. Also integrated skin color adjustments for high-quality, seamless e-commerce imagery. • Implemented a novel two-stage DL approach for garment draping, transitioning apparel from dummies to premium models via sophisticated image warping techniques, enhancing e-commerce imagery.
InfoBot
June 2021 - December 2021
- Filter (VaaanInfra Pvt. Ltd) • Enhanced operational efficiency by deploying a chatbot and developing a news classifier, optimizing workload management and facilitating access to relevant, real-time information for users. • Engineered an advanced end-to-end chatbot with heightened analytical capabilities, autonomously gathering and synthesizing data from diverse sources including websites, APIs, and documents. Requires minimal to no maintenance as it continuously trains itself. • Utilized web scraping to gather news from major sources, implementing a classifier to sift out articles unrelated to company domains.
Education
Bhartiya Vidyapeeth College of Engineering
BTech
January 2018 - January 2022
