E-commerce product photos generation
- #E-Commerce
Our client is an e-commerce company that needed a solution to generate professional product photos to showcase their products online. They wanted an easy and cost-effective way to create high-quality images without the need for expensive photographers and studios. Our team was tasked with creating a system that would allow them to achieve this goal.
- Computer Vision
Impact
- A system for tuning the StableDiffusion model on customer’s product photos
- A deployment pipeline for StableDiffusion that minimized GPU costs while supporting a large number of users
- Affordable and easy e-commerce product photo solution, saving money on professional photographers and studios, and speeding up sales
Services we provided
A system for tuning the StableDiffusion model
A deployment pipeline for StableDiffusion
Affordable and easy e-commerce product photo solution
Tech Stack
Pytorch
Python
Stable Diffusion
Huggingface Diffusers
DreamBooth
Runpod
Challenges and Solutions
🧐 Challenges
- Create a deployable and scalable system to fine-tune the StableDiffusion model for customer product photos.
- Achieve fast generation times by developing an efficient and scalable solution.
- Establish a cost-effective deployment pipeline for StableDiffusion, accommodating a large user base while optimizing GPU costs.
- Design the system to handle a substantial user load without compromising on performance or resource utilization.
💡 Solutions
- Conducted several experiments to identify optimal parameters for the StableDiffusion model
- Created and tuned prompts for each category, style, and environment available on client’s website
- Conducted experiments on input data to obtain necessary preprocessing and photo requirements from users
- Optimized model and prompts to ensure highest quality output
- Deployed system to serverless endpoints on RunPod
- Launched a container with model training per user to enable scaling and parallel execution of requests while minimizing GPU costs
User flow
1. User uploads a product photo.
2. The system processes the photo and places the product in a captivating environment.
3. A dynamic screen displays the product's results, enhancing the visual presentation.
4. The user receives a visually stunning product photo, showcasing the product's potential in an attractive setting, thanks to the system's smart algorithms.