Luma Ray3 Modify Keeps the Performance and Renegotiates the World Around It

AI video tools have made the effect the easy part. What they still struggle with is not wrecking the take you already paid for.

Most video generation tools treat the source footage as raw material for a new clip. Luma's Modify Video tool takes a different position: the performance you filmed is the valuable part, and the world around it is up for negotiation. The pitch, first made when the feature launched inside Dream Machine in June 2025, has not changed. What has changed is the model underneath it, twice, and the controls wrapped around it.

What the tool does now

The consumer app now labels the feature Video To Video. The API calls it a video_edit request, and it runs on Ray 3.2, which Luma launched on June 9, 2026. Ray 3.2 brought keyframe control, up to 16 keyframes per clip at arbitrary positions, performance tracking across up to eight simultaneous faces, native HDR generation with 16-bit EXR export, and full API access for the first time. Duration extends to 20 seconds at 1080p.

The Modify tool itself has grown into something closer to a full editing request than the three-preset restyle it started as. Outputs come in 540p, 720p, or 1080p, at 5 or 10 seconds. The API caps source video at 18 seconds and 200 MB, and allows up to 64 guide images pinned along the video. Luma's guidance describes the model as reading performance signals from the source, including pose, facial expression, and scene structure, and using them to decide what to preserve and what to rebuild.

The strength slider is a production decision

The control that matters most is the strength setting, and it is not a quality slider. It sets the balance between source fidelity and generative freedom, and Luma organizes it into three bands.

Adhere keeps the model close to the original contours, subject placement, materials, and environmental detail. It is the right end of the scale for relighting, recoloring, light retexturing, small wardrobe adjustments, and effects that should sit on top of an otherwise approved shot. It is also the safest first test when the source contains choreography, precise product handling, lip movement, or a deliberate camera move.

Flex is the practical starting point for noticeable wardrobe swaps, stylized live action, period changes, environment updates, and character replacements that retain human proportions. It gives the prompt and reference enough influence to become visible without immediately discarding the source performance.

Reimagine loosens the source geometry so the model can invent a new character silhouette, environment, or visual language. This is the band for puppeteering creatures, turning a performer into a material-based character, or shifting an ordinary location into a fantasy world.

There is a specific failure mode worth knowing. At stronger Reimagine settings, the model is less constrained by original edges and may infer movement from the transformed subject's pose. That freedom can introduce zooms or shifts instead of preserving the recorded camera path. In other words, the more you ask the model to reimagine, the less you should expect the original camera move to survive.

A workflow that matches the tool

The useful discipline here is to write a lock list before generating. Name subject identity, pose timing, camera path, wardrobe, background, lighting, product geometry, text, and audio, and mark each one as preserve, modify, or irrelevant. If five elements must stay unchanged and only a jacket color should move, start low. If the performer must become a glass creature inside a different world, start high. The best result is usually the lowest setting that clearly completes the requested change, because more freedom increases the chance of camera drift or unstable detail.

The comparison method follows from that: hold the source, crop, prompt, and references constant, then render a small bracket at one lower, one middle, and one higher setting, and judge them against the lock list. Record the winning setting alongside the source, prompt, and references so later shots can reproduce the decision.

Why Luma is positioned as infrastructure

Luma has been explicit that Ray is professional infrastructure priced for production volume, not a consumer app. The context makes the choice look deliberate. The company raised a $900 million Series C led by HUMAIN, opened a London office, and runs enterprise Luma Agents deployments at Publicis, Adidas, and Mazda. The Mazda relationship produced a concrete deliverable in April 2026 when the Johannesburg agency Boundless used Luma Agents to deliver the automaker's first AI-produced commercial in under two weeks.

That is the most credible production-deployment signal any AI video platform has put on the board, and it depends on the same promise Modify makes. A commercial uses real talent, approved frames, and a director's intent. A tool that regenerates the whole shot is unusable in that pipeline. A tool that preserves the performance and changes the world around it can be slotted in.

Ray 3.2 and Ray 3.14 are parallel sub-models in the Ray3 family. Ray 3.14, which shipped in January 2026, handles duration-change and loop workflows, where its fixed-length output is an asset rather than a constraint. Ray 3.2 handles standard video-to-video transformation and multi-keyframe guidance.

Where it sits against the rest of the stack

Modify is a restyling tool, not a general editor, and that boundary is worth respecting. For tasks it was not built for, teams still reach for Seedance, Kling, or Wan, often through the same API key on an aggregating platform. What Luma is selling is a specific capability inside that stack, not a replacement for it.

The place where it stands apart is the combination of motion preservation and keyframe control. A first-frame-only tool commits you to a starting image. Modify commits you to an entire recorded performance, and then lets you place up to 64 guide images along it. That is closer to how an editor thinks about a shot than how a prompt writer thinks about a clip.

The caveats that matter before you budget

Audio is the first one. On third-party endpoints, the Ray 3.2 video-to-video output comes back silent, so you re-attach the original track in your editor after the restyle. Luma's own launch material does not address audio handling for the consumer app, which means a short test clip is the safest way to find out how the Dream Machine version treats sound.

Cost is the second. On one cloud provider, a 5-second run at 720p costs about $1.188 with no discount active, and the 10-second version doubles that. The billing model has moved from pixels to clips, which is easier to predict but means the cost of a bracket of three test renders is real money at scale.

Breadth is the third. Modify is a restyling tool, not a general editor. For tasks it was not built for, you are still going to Seedance, Kling, or Wan, often through the same API key.

What to watch

The claim to test is preservation. If Modify holds a performance through complex motion and a deliberate camera move at a Flex setting, it earns its place in a commercial workflow. If it drifts the camera whenever the prompt gets ambitious, it stays a stylization tool with a professional label. The next couple of model revisions will settle which one it is.