NVIDIA, Google, and World Labs Just Changed AI Filmmaking Forever
Caleb from Curious Refuge breaks down a wave of new AI tools that can turn a single image into an entire explorable 3D city — and a new Google screenwriting tool that looks wild.
I have been geeking out on AI tools for a while now — using them for scripts, images, automation, all of it. But I just watched Caleb from Curious Refuge break down a wave of new releases and even I had to pause and go "wait, that's real?" Multiple times.
The short version: the gap between a single image and an entire explorable 3D world just got dramatically smaller. Here's what's actually happening.
AI filmmaking is about to feel completely different — not because of video generation, but because of world generation.
NVIDIA's LRA 2: From Image to Interactive World
In the video, Caleb explains that NVIDIA released a tool called LRA 2 that does something genuinely mind-bending. You give it one starting image, and it generates a full 3D world from it — including the geometry, the depth, the environment. Then you can explore that world with camera movement, and connect it to NVIDIA's Isaac SIM model to make it interactive.
The practical demo Caleb shows is a 3D scan of a home that you can actually walk through using AI-driven interactions. And they've demonstrated it at city scale — you can zoom out to a full aerial view of San Francisco and zoom back into street level.
As Caleb puts it in the video, this fundamentally changes how you approach storytelling in AI film. Right now, maintaining environmental consistency across shots is one of the biggest unsolved problems — you generate scene A, then try to generate scene B in the same world, and they never quite match. 3D world generation means you build the set once, then direct your shots within it. That consistency problem basically disappears.
NVIDIA isn't alone. Tencent Hunan (the team behind Minimax) released an open-source version of essentially the same capability around the same time. Caleb notes this pattern keeps repeating: a major closed-source release, followed almost immediately by an open-source alternative that costs a fraction as much. The pace of this is wild even by AI standards.
World Labs Spark 2: City-Scale 3D Environments
World Labs builds the underlying architecture that a lot of image-to-3D-world tools are built on. Their new render engine, Spark 2, scales those environments in a way that works online — so instead of being a desktop-only research demo, it's usable in a browser.
Caleb shows a demo of Hobbiton from Lord of the Rings rendered in Spark 2 — walkable, explorable, surprisingly detailed. He also shows a full-city San Francisco render where you can zoom from satellite view down to street level. It's not fully real-time yet, but the fact that it composites a whole city into a coherent explorable environment is genuinely impressive.
The Google Screenwriting Tool
Beyond world-building, Caleb also covers a new Google AI screenwriting tool that he describes as "crazy impressive." I'm not going to oversell it — what stood out from the video is that Google is bringing AI into the writing phase of film production, not just the generation phase. For creators who are currently bouncing between writing tools, image generation, and video tools, having Google-grade AI assist with the screenplay layer could be a real workflow unlock.
What I'm Actually Thinking About This
I've been using AI tools in my own content workflow for a while — Midjourney for images, various tools for scripts and automation. The thing that strikes me about this wave of releases is that the barrier between "having an idea" and "having a visual world" is collapsing in real-time.
When I first started exploring AI image tools, consistency was the killer. You could generate one beautiful image, but generating a coherent series of images with the same characters, environment, and lighting was basically impossible without a lot of manual work. World generation tools fix the root cause of that problem.
Caleb also covers a real workflow challenge in the video: using Seed Dance via Runway to remove unwanted characters from a generated scene. After eight attempts and $32 in generation costs, he got the result. His honest take was that it's better to generate your base footage exactly right the first time than to try to fix it in post with AI — each iteration adds quality loss and cost. That's a lesson that applies to AI tools in general: prompt well upfront, and you'll spend a lot less time (and money) fixing mistakes.
The Bottom Line
If you're in any kind of content creation — not just filmmaking — the world-generation tools Caleb covers in this episode are worth watching closely. We're moving from AI video (impressive clips) to AI environments (full 3D sets you can direct). That's a different kind of capability.
Watch the full Curious Refuge video for all the demos. And if you're not already following that channel, it's one of the cleaner weekly roundups for what's actually moving in AI filmmaking versus what's just hype.