Martini, Gaussian Splats and the Art of AI Cinematography
For three years, AI video has had a strange problem. You don't really direct the camera. You ask for a camera move and hope for the best.
Type "slow dolly in, 35mm, shallow depth of field" into a prompt box, hit generate, and see what comes back.
But a director on a real set doesn't describe a camera move and leave the rest to chance. You find the frame, choose the lens, block the movement, decide what gets revealed and when, then shoot it.
Martini is trying to bring some of that control into AI video.
Give it a single still image and it reconstructs the scene as a navigable 3D space using Gaussian Splats. You can move a virtual camera through it, choose your lens, set keyframes and work out the move before generating the video.
So instead of prompting for a shot, you block one.
Last month at FaiGHT CLUB, we made the case for keeping AI filmmaking local. This month, we're going back into the cloud to test whether more creative control makes that trade worthwhile.
The question for the night is simple:
If you control the camera, how much of the shot do you actually control?
🎬 What You'll Work On
"What changes when you can move the camera?" (20 min)
A filmmaker's introduction to Gaussian Splats, what they allow you to do with a single image, and why that matters for image-to-video.
Gavin will walk through Martini's approach and compare it with the prompt-based camera controls we've been using in AI video.
We'll also pick up where last month's local AI session left off. What are we giving up when we move back to a cloud tool, and what are we getting in return?
"One Frame. Five Moves." (70 min)
Everyone starts with the same still frame.
Teams take that image into Martini and choreograph different camera moves through it: a creep-in, a reveal, an orbit, a push-through and a wildcard.
You choose the lens. You place the camera. You set the keyframes and timing.
Then we generate the shots and put them side by side.
Same frame. Five directors. Five different shots.
If the camera move changes how we read the scene, we want to see how much of that difference came from the person behind the controls, and how much still came from the model.
Debrief & Share
Which moves worked? Which ones broke the illusion? Where did the reconstruction fall apart?
Did blocking the shot give us more control than prompting for it, or just a different kind of unpredictability?
And where could this actually fit into a filmmaker's workflow, whether for previs, pitching, client work or finished shots?
⚔️ The Five Rules
Rule 1: Make, Not Talk
No endless AI theory. Build actual shots in two hours that you can test Monday morning.
Rule 2: Teamwork = Human Open Source Bot
No passengers, no creative dictators. Everyone contributes, everyone rotates. Everyone gets a turn behind the virtual camera.
Rule 3: Don't Hoard Your Subscription Credits
Fast hours, API tokens, pro access. Pool your compute power for the team. Generosity wins.
Rule 4: No Subscription Required
You won't need your own Martini subscription to participate. [Confirm access arrangement before publishing.]
Rule 5: Leave Egos at the Door
Director, DP, editor, producer, fresh grad. Everyone is equal at the table. Pure collaboration, pure learning.
⚠️ Important Notes
Bring a laptop with internet access and a Chrome-based browser.
Bring one still image you'd love to fly a camera through. It could be a location photo, a frame from your own footage or a generated image. We'll pick the battle frame or frames on the day of the workshop.
All experience levels welcome, but come ready to collaborate and experiment.
In the event of oversubscription, Lorong AI members will be prioritized.
🗣️ Facilitated by Gavin Lim (AI Film Lab)
Gavin Lim is founder of AI Film Lab, a Future of Work community exploring the intersection of AI technology and visual storytelling. A seasoned TV producer, director and filmmaker with nearly thirty years of experience, Gavin leads practical exploration of AI tools that enhance rather than replace human creative judgment.
AI Film Lab has conducted AI filmmaking hackathons with Singapore media practitioners and led the first AI filmmaking hackathon at the Philippines' inaugural AI Festival (2025). Their approach centres on iteration over perfection and collaborative learning over isolated expertise.
More About the Series
FaiGHT CLUB is a monthly AI filmmaking lab bringing together professional filmmakers, creatives, and producers for hands-on exploration of AI tools that enhance rather than replace human craft. Each session tackles a specific production challenge through collaborative team sprints, guest case studies, and practical workflow development.
Tinker Time is Lorong AI's hands-on meetup combining guided workshops with open collaboration time for builders to share code, troubleshoot together, and experiment with AI development techniques.
Get involved: Learn more about Lorong AI | Speaker Sign-up | WhatsApp Community | LinkedIn | X