As many of you know, I’ve been deep in the trenches of testing AI video creation for academic projects at Duke University. My “litmus test” has always been trying to accurately animate molecular bonding—a project that dates back to the “Will Smith and spaghetti” days of AI.
The Challenge of Molecular Motion
Recently, I ran into a challenge across all of the video generation products: moving molecules. I had managed to get an OK result in “Draft Mode,” but I didn’t feel that I had control over the specific choreography. I reached out to the support team at Luma Labs, and the response from Chris in Community Support was incredibly refreshing and honest:
“I want to be upfront about something important: precise, frame-accurate animation—like directing specific molecules to move to exact positions and form specific bonds—is not really how AI video generation works right now… It’s fundamentally different from instructing 3D software like Blender or Maya where you have exact control over every object’s position and movement.”
Expert Tips for Better Scientific Renders
Rather than just pointing out the limitations, Chris provided a roadmap for how to push the boundaries of Dream Machine:
-
Ray3 Annotations (Scribbles): This feature lets you visually “draw” motion paths directly on your image. It gives the AI much stronger directional cues than text alone, which is vital for complex structures.
-
Don’t Re-create a Working Draft: I mentioned that I had a good result in Draft Mode but lost it when I tried to re-generate at higher quality. Chris noted, “You don’t need to do that!”. Use the HiFi Option: When you have a Draft result you like, select the HiFi button at the bottom right to upscale that exact output to 720p or 1080p without “rolling the dice” on a new generation. (Steve Note: It will cost more to do you testing at that resolution than draft mode)
-
Prompt Engineering for Science: For a clean look, try using: “fixed camera viewpoint, a detailed 3D molecular model slowly rotating on a clean background, scientific visualization style”.
-
Simplify and Chain: Instead of one massive prompt with simultaneous instructions, break it into single-action segments and use the Extend feature to chain them together.
The Road Ahead: Luma in Education
What impressed me most wasn’t just the technical advice, but the genuine interest in the educational application of these tools. To that end, I’m excited to share that I’ll be signing up for the Luma Education Program to explore how we can bring these capabilities to more educators and students at Duke.
If you’re an educator looking to dive in, you can check it out here: Luma for Education.
A huge thank you to Chris and the Luma team for the “goodwill” credits and, more importantly, for the expert guidance. It’s a reminder that even in a world of automated models, human support is what truly moves the needle.
