A $15M ad campaign wrapped in 40 hours for $20K — and it passed the brand's internal quality review.
What Is It?
AI video startup Luma just launched Luma Agents — a creative AI system that generates text, images, video, and audio end-to-end in one place.
Here's the thing — until now, AI creative work meant juggling separate tools: Midjourney for images, Sora for video, ElevenLabs for audio. Each had its own prompting style, and stitching results together was entirely on you. Luma Agents collapses that whole pipeline into a single system.
The technical core is a new architecture called Unified Intelligence. The first model, Uni-1, is a decoder-only transformer that processes language tokens and image tokens in one sequence simultaneously — think of it as reasoning and rendering at the same time. CEO Amit Jain compares it to an architect: just as an architect drawing a line simultaneously simulates structure, light, and spatial dynamics, Uni-1 handles both reasoning and rendering in a single forward pass.
Luma was founded in 2021 and is based in Palo Alto. It's raised a total of $1.1B at a $4B valuation. Last November it pulled in $900M from HUMAIN — a subsidiary of Saudi Arabia's PIF — and announced plans to build a supercluster data center in Saudi Arabia. Investors include Andreessen Horowitz, NVIDIA, AWS, and AMD Ventures.
What Changes?
You could always make content with AI. But most tools basically said "here are 100 models, learn to prompt each one." In Jain's words, the current state of creative AI is one where you spend most of your time "orchestrating tools" rather than actually creating.
| Traditional Multi-Tool Approach | Luma Agents | |
|---|---|---|
| Workflow | Prompt each tool separately → manually merge outputs | One brief, end-to-end automation |
| Context | Context lost when switching tools | Full project context maintained |
| Quality control | Humans review every output manually | Agent self-evaluates and iterates |
| Model selection | You need to know each model's strengths | Automatic optimal model routing |
| Collaboration | Passing asset files back and forth | Real-time collaboration on a multiplayer board |
The real-world examples are impressive. According to LBB, a small South African agency with fewer than 20 people completed a Mazda MX-5 campaign in just two weeks — covering the car's evolution through the '80s, '90s, and 2000s. Sourcing actual vehicles, shooting, and post-production would normally take months.
Another global client shot a master video the traditional way, then used Luma to localize it for 150+ markets and languages. Craft by humans, scale by AI — that's the hybrid production model Luma is proposing.
The list of external models Luma Agents currently integrates is substantial: Ray3.14, Google Veo 3, Sora 2, Kling 2.6, Nano Banana Pro, Seedream, GPT Image 1.5, ElevenLabs, and more. The agent automatically picks and routes the best model for each step of the job.
Getting Started
- Sign up for Luma
Create an account at lumalabs.ai. API access is available too, but starting with the app is the easiest way in. - Plan with Brainstorm mode
Start with natural language like "create a product launch campaign." The agent expands your ideas and suggests creative directions. Nothing gets generated yet — this is the strategy phase. - Build with Create mode
Once you've locked in a direction, switch to Create mode. The agent generates images, video, and audio — handling model selection, routing, and iterative refinement automatically. - Refine with feedback
"Make the colors warmer," "go with direction #2" — give feedback like you're talking to a colleague. The agent keeps full context and adjusts accordingly. No prompt engineering required. - Scale & deploy
Take your finished assets and generate variations for different platforms, languages, and markets. Collaborate with teammates in real time on the multiplayer board.
Key Takeaway
Customers retain 100% IP ownership of generated content. The platform includes automatic copyright risk review, human-intervention audit trails, and a mandatory human review workflow before publishing.
