If someone tells me AI creation is just “press a button and you’re done,” I’ll give them one of my recent client tasks.

“Make dominoes fall exactly like they do in real life.”

I tested 7 different AI models, ran dozens of generations, rewrote prompts over and over… and in the end, the client and I decided to skip the scene because today’s AI simply couldn’t produce realistic physics.

Sometimes the hardest part of AI filmmaking isn’t generating beautiful visuals. It’s discovering where the technology still reaches its limits. 🎬

Thrilled to see my latest R&D project, "Unfurl," featured in Stash Magazine! This piece is an exploration into the commercial potential of generative AI, asking: Can we bridge the gap between AI’s unpredictable nature and the precision required for production? The entire project lives at the intersection of human art direction and machine learning. To maintain deliberate control over the interplay of flowers and fabrics, I used Cinema 4D as the base structure for the generative layers (ComfyUI, Krea, Kohya). Huge thanks to Tom Parry for building such an incredible, tactile "unfurling" soundscape that perfectly mirrors the visuals. You can watch "Unfurl" and read more about the process in the Stash feature here: https://www.stashmedia.tv/vitor-teixeira-releases-new-personal-short-film-unfurl/

Congrats on the Stash feature! That tension between AI's unpredictability and production-ready precision is exactly the problem I think about with client work too. Curious what your actual pipeline looked like, pure generative output or a lot of manual cleanup after?