Luma's Ray3 Modify interpolates video between start and end frames

Anyone who has ever worked on a video set knows the nightmare of reshoots. The actor nails the emotional delivery, but the client hates the wardrobe, or the background looks wrong, or you are missing two seconds of footage to connect a scene.

AI video models could fix this in theory, but in practice, running footage through AI usually wipes out the actor's subtle facial expressions and eye contact.

Luma, the AI video company backed by a16z, just launched a new model called Ray3 Modify to fix that exact problem.

Keeping the acting, swapping the scene

Generative video can make breathtaking clips, but getting it to do exactly what a director wants has always been like herding cats.

Luma claims Ray3 Modify directly tackles the challenge of preserving human performance while applying AI effects. Creators can take existing footage and feed the model a character reference image. The AI then converts the human actor into that character while keeping the original motion, timing, eye line, and emotional delivery intact.

This means studios can shoot actors against a simple backdrop, then use AI to transport them to any fictional world, swap their costumes, or tweak their appearance across multiple shots without losing identity continuity.

As Luma co-founder and CEO Amit Jain put it, creative teams can capture real performances and immediately modify them in post, effectively reshooting scenes with AI instead of dragging an entire crew back on set.

Directing between start and end frames

Ray3 Modify also introduces a feature that video editors have desperately needed: start and end frame control.

Instead of praying that a text prompt moves characters where they need to go, users can provide an opening frame and a closing frame. The model then generates the transitional footage in between.

This gives creators precise control over camera moves, character transitions, and scene continuity.

The new model is live on Luma’s Dream Machine platform. The company, which is locked in tight competition with rivals like Runway and Kling, first launched basic video modification tools in June 2025.

This release also follows some massive financial momentum. In November, Luma announced a $900 million funding round led by Humain, an AI firm owned by Saudi Arabia’s Public Investment Fund, with participation from a16z, Amplify Partners, and Matrix Partners. The startup is also working with Humain to build a massive 2GW AI cluster in Saudi Arabia.

Why it matters

We think Ray3 Modify represents a mature pivot in generative video, moving away from parlor tricks and toward real Hollywood production pipelines.

For commercial directors, indie filmmakers, and brand studios, this is immediately useful. You no longer have to burn budget on physical location changes or minor wardrobe fixes. Shooting cheap reference footage with real actors and letting AI handle the heavy visual lifting dramatically lowers production risk.

It also shifts the competitive battleground. Text-to-video novelty has worn off; the race between Luma, Runway, and Kling is now entirely about controllability. Whoever gives directors frame-level precision wins the high-margin enterprise market.

Still, the boundary between an actor's performance and an AI-generated likeness is getting very blurry. While Luma's approach is technically smart, it is bound to stir up fresh debates around digital performance rights and actor compensation down the road.