AutoDrawLapse

Inspiration

Most generative AI tools today just "poof" a final image onto your screen. While the results look great, the actual magic of creating art, the rough sketching, the deliberate shading, the human element, is completely lost. I wanted to build something that bridges the gap between AI generation and the authentic, physical process of drawing.

What it does

AutoDrawLapse is an advanced generative art engine that transforms any static image into a highly engaging, stroke-by-stroke drawing timelapse. Instead of just applying a static filter, it mathematically simulates a human artist, analyzing the image and laying down thousands of individual, real-time pen strokes so you can actually watch the art being created.

How I built it

I architected a complex, three-stage computer vision pipeline. First, it breaks the image down into macro and micro regions. Then, it extracts a highly accurate structural skeleton of the subject. Finally, it uses a custom mathematical flow algorithm to plan out thousands of individual pen strokes that actually wrap around the physical 3D geometry of the image.

To actually implement this massive plan, I used the Codex CLI and GPT-5.6 Terra in MAX mode as my pair-programmers. Since I am operating on a free-tier account, Terra was the most capable agent available to me. I designed the overall logic and constraints, and Codex executed the heavy mathematical implementations and system architecture. To make it easy to evaluate, I wrapped the core CLI engine in a simple Gradio web UI.

Challenges I ran into

My original vision was to build this tool out to function like a SaaS web application. However, I ran into a lot of confusion trying to figure out how to actually deploy such a heavy engine as a solo developer. I intended to use platforms like Hugging Face and Modal, but I failed because of resource limitations. After struggling with these deployment logistics, I had to pivot and figure out a way to run the Gradio interface cleanly inside a Google Colab notebook instead.

Accomplishments that I'm proud of

I am incredibly proud that I managed to build a production-quality generative art tool using GPT-5.6 Terra. Since I am on a free-tier account and don't have access to the heavier GPT-5.6 Sol model, achieving this level of complex mathematical planning and architectural stability with Terra proves just how powerful the Codex CLI can be when guided by a strong technical plan.

What I learned

I learned that you have to be highly adaptable when deployment logistics get confusing. When a traditional web app deploy didn't work out, I learned how to engineer a secure Colab environment instead.

What's next for AutoDrawLapse

In the immediate future, I am going to use AutoDrawLapse as my own personal content creation tool to generate unique videos. Long term, my ultimate goal is to evolve this core engine into a fully fledged SaaS product so any creator can use it directly in their browser.

Built With

• codex
• gpt5.6
• python