soro2

Six months ago, my co-founder and I were sitting in a coffee shop, frustrated. We'd just spent $3,000 on a 30-second product demo video, and the back-and-forth with the production team took three weeks. There had to be a better way.

That's when we started building Sora2.

The Problem We're Solving

Video content dominates the internet, but creating it remains painfully slow and expensive. Most creators face the same dilemma: either spend thousands on professional production or settle for mediocre quality. We wanted to eliminate that trade-off entirely.

What Makes Sora2 Different

The breakthrough came when we focused on physics simulation. Earlier AI video tools could generate moving images, but they looked... off. Objects would float unnaturally, lighting would shift randomly, and movements felt robotic .

We rebuilt our model from scratch to understand real-world physics. When you generate a video of water splashing or fabric moving, it behaves like actual water and fabric. The lighting stays consistent across frames. Shadows fall where they should .

How It Actually Works

The interface is dead simple. You type what you want to see, or upload an image you want to animate. Our system breaks down your request, understands the physical properties of objects involved, and generates video that respects those properties .

We also added something we call "multi-shot" capability. Instead of just one continuous clip, you can create sequences with different camera angles and perspectives—like an actual film production .

Early Results

We launched a private beta two months ago with 200 creators. The feedback has been wild:

A marketing agency replaced their entire video production workflow

An educator created a full semester of animated lesson content in one weekend

A small business owner generates daily social media videos without touching a camera

One user told us they'd been paying $500 per video ad. Now they're creating 20+ variations per campaign and testing what actually converts .

The Technical Challenge Nobody Talks About

The hardest part wasn't the AI model—it was making generation fast enough to be useful. Nobody wants to wait 45 minutes for a 10-second clip. We spent months optimizing our pipeline. Now most videos generate in under 3 minutes .

Audio sync was another nightmare. Video and sound need to match perfectly, or the whole thing feels uncanny. We trained our model to generate them together, not separately, so footsteps land when feet hit the ground and dialogue matches lip movements

What We're Learning

The use cases surprise us daily. We built this thinking about marketing teams and content creators. But we're seeing:

Real estate agents creating property walkthroughs from floor plans

Game developers prototyping cutscenes before hiring animators

Teachers explaining complex concepts with custom animations

Nonprofits creating awareness campaigns on zero budget

Turns out when you make video creation accessible, people find problems you never imagined.

The Honest Struggles

Not everything's perfect. Generating human faces still needs work—we're conservative about that to avoid misuse. Long-form content (over 60 seconds) remains challenging. And sometimes the AI interprets prompts in unexpected ways, which can be hilarious or frustrating depending on your deadline [1].

We're also navigating the ethics carefully. We built in watermarking and usage tracking. Commercial rights are clearly defined. We reject prompts that could create misleading content.

Where We're Headed

Right now, Sora2 runs entirely on the web. We're building API access for developers who want to integrate video generation into their own products [3]. Mobile apps are coming next—imagine generating videos directly from your phone.

The bigger vision? We think every piece of written content should have a video equivalent. Blog posts, documentation, educational materials—all of it could be visual without requiring production teams.

Why I'm Sharing This

I've learned more building Sora2 than in my previous ten years as a developer. A few things that might help other founders:

Ship before you're ready. Our beta was embarrassingly buggy, but user feedback shaped everything that mattered.

Talk to users constantly. We do 5-10 user calls every week. The insights are gold.

Technical moats matter. Our physics simulation took months to perfect, but competitors can't easily replicate it .

Distribution beats features. We spent too long perfecting the product before thinking about how people would find it.

What I'd Love Feedback On

How do you think about AI-generated content in your workflow?

What would make you trust AI-generated videos for professional use?

What video creation problems are we missing?

Happy to answer any questions about the tech, the business model, or the journey so far.