InstaLoRA

LEO input video frame
LEO samurai man
LEO the business man
LEO the warrior
Kingston notcool
Kingston COOL

Inspiration

InstaLoRA is revolutionizing AI art creation by breaking down the barriers that have made personalizing LoRAs a complex task. Traditionally, this required a personal PC, a vast collection of diverse, captioned images, and significant technical expertise. We recognized the need for a simpler solution that eliminates the hassle and opens up AI art creation to everyone, making it as easy as using two short videos.

What it does

With InstaLoRA, creating your own personalized LoRA model is as simple as uploading two 3-second videos—a close-up and a more distant shot. Our platform quickly processes these inputs to deliver a custom LoRA model within minutes. This model enables users to generate unique AI art in any style they desire, for free, enhanced with contextually rich captions generated by our advanced script. No technical knowledge is needed, just a creative spirit.

How we built it

Our approach to building InstaLoRA was focused on user-friendliness without compromising on technical sophistication. We developed scripts to preprocess user-uploaded videos, utilized kohya-ss's sd-scripts for efficient LoRA model training, and integrated Hugging Face's StableDiffusionPipeline for high-quality image generation. The entire process is powered by Modal's serverless architecture, ensuring a seamless experience for the user while leveraging the power of an H100 GPU and the DreamShaper model for optimal performance.

Challenges we ran into

One of the main challenges we faced was automating the training and image generation process to make it accessible via a web interface, conserving GPU resources and ensuring the system is easy for anyone to use. Learning to navigate Modal and effectively integrate different technologies for a streamlined process involved a steep learning curve and required us to innovate continuously.

What we learned

The development of InstaLoRA has enriched our understanding of serverless architectures, personalized LoRA model optimization, and user interface design. This project has broadened our perspective on the potential of AI in creative industries and inspired us to continue pushing the boundaries of what's possible with technology.

What's next for InstaLoRA

ok, so the ui's a bit rough - we know. but that's an easy fix. the real win is getting instalora working its magic on modal. next up, we’re gonna pack in more styles, smooth out the user experience, and really dial up the efficiency. plus, we’re all about listening to what users want, so we’re ready to grow instalora into something even more epic.