← Back to blog
AiAbout 6 min read

Adobe's Project Live Set Edits the Clip While It Plays

Published Nov 14, 2026
Adobe's Project Live Set Edits the Clip While It Plays

The demo at Adobe MAX Sneaks on November 11 runs a black cat sitting on a floor. Four prompts are stacked over it: give the cat a chef's hat, replace the cat with a shiba inu, replace it with an alligator in a Miami-print shirt, replace it with a puppy in a tuxedo. The playhead keeps moving while the substitutions resolve. Nothing renders to a file and comes back. The frame changes while you watch it.

Adobe calls this Project Live Set. It is a research preview, not a product, and the company did not describe the models behind it or give a ship date. Gautham Mysore, who leads Adobe's video and audio research group, framed the intent as bringing generation and editing into one continuous experience so a creator can see a result, react to it, and steer what happens next. He also called it a sketching tool for moving images, which is the most accurate description anyone has offered so far.

What changed in the interface

Generative video has spent two years as a batch process. You write a prompt, wait, get a clip, watch it, and discover the third second is wrong. Then you either regenerate the whole thing and hope the good parts survive, or you move to an editor and try to hide the bad stretch with a cut or a mask.

Live Set inverts the order. You start from a clip that already exists, and you edit the clip while it is running. The unit of work stops being the generation and becomes the playback position.

A loop of transparent film under a magnifier, its frames shifting from cool blue to warm amber.

That sounds like a UI detail. It changes what you can evaluate. Judging a generated shot means watching it in motion, with its own timing, and most current tools ask you to judge it from a still or a short loop before you commit to a full render. If the edit happens on the timeline, the judgement happens where the mistake is visible.

The shipping part: generation inside Premiere

The research preview got the headlines. The November 10 Creative Cloud announcement is the part people can use this month.

Adobe put Generate Video inside the Premiere timeline through a tool it calls Generative Media. The editor selects a gap, and the system reads the surrounding edit for context before generating b-roll to fill it. Whatever comes back lands in the timeline and can be refined in place. Model choice is not limited to Firefly: the panel also offers partner models from Google, Kling, Runway, and Luma.

That last sentence is the quiet strategic move. Adobe spent years positioning Firefly as the reason to stay inside Creative Cloud. Opening the timeline to competitors' video models accepts a simpler truth: editors do not want a vendor's model, they want the shot that fits the gap.

Generative Remove arrived in beta alongside it. You mask an unwanted object, track it across the clip, and click remove. Adobe's examples are a boom mic, a logo, a background distraction. The claim is that the fill is realistic and the rest of the frame is left alone, which is the claim every removal tool makes and the only one that matters.

The audio updates follow the same logic of removing handoffs. Enhance Audio splits a mix into dialogue, music, ambience, and effects so each can be balanced on its own. Separate Sound Source cleans noisy dialogue and breaks it into components. Separate Crosstalk puts overlapping speakers onto independent tracks. Color Mode, previously covered, is now generally available.

Why the timeline is the right place for this

A prompt box knows nothing. A timeline knows the shot before, the shot after, the lens, the grade, the camera movement, and the fact that the actor is holding a cup in their left hand.

Every serious generative video workflow has been built around that gap. You generate in a chat window, download, import, and then spend the next hour matching the result to footage it was never told about. The generated clip is judged on its own merits and then fails on contact with the sequence.

Putting generation where the context lives removes a category of failure rather than a step. It also makes a smaller promise than "AI edits your film". The tool fills a gap you defined. You still decide that the gap should exist.

What Live Set still has to prove

Three things, and Adobe has answered none of them publicly.

The first is cost and latency at real resolution. Real-time substitution on a compressed preview is a different engineering problem from real-time substitution on a 4K master. Every demo of this kind looks instant because the demo is built to look instant.

The second is temporal consistency. Swapping a cat for a shiba inu in a locked-off shot is the easy case. The hard case is a subject that turns, crosses behind another object, and comes back. The demo shows four substitutions at once, which is a display of throughput, not of persistence.

The third is what happens to the rest of the frame. Generative edits tend to leak. You ask for a hat and the lighting on the ear changes, or the shadow under the cat shifts two degrees. Adobe's shipping removal tool promises the untouched region stays untouched, and that promise is testable in a way the sneak is not.

There is also a licensing question sitting under all of it. Premiere now offers partner video models from four companies with four different training data positions and four different commercial terms. An editor picking a model from a dropdown is making a legal choice, and nothing in a dropdown menu communicates that.

The pattern across the November releases

Look at what shipped in the same ten days. Krea opened a beta of Bots, agents that take on whole creative projects and keep house rules in memory. ElevenLabs folded lip sync into dubbing so a foreign-language cut is one job instead of two. Melius opened Ditto, which swaps people and objects inside a finished clip. Google added five tools to Flow, one of which turns a single phone take into several camera angles while keeping the performance.

None of these are new models. Every one of them is a change to where the work happens.

The model race of the past two years produced a strange side effect: capability arrived faster than the places to put it. You could generate a convincing eight-second shot and still spend an afternoon getting it into an edit. The current round of releases is about closing that distance, and the companies closing it fastest are the ones that already own the surface where the decision gets made.

Adobe owns the timeline. That is a strong position and an uncomfortable one, because owning the surface means the models can be rented from anyone, including the companies Adobe just invited into its own panel.

What to do with this now

Live Set is not something to plan around. Sneaks previews have a long history of arriving as something narrower two years later, and Adobe explicitly declined to commit.

The Premiere features are real. If you cut video for a living, the interesting test this month is whether generating b-roll inside the edit produces fewer rejects than generating it outside. Whether the b-roll looks good is a question that settled itself a year ago. My guess is that the number of takes you throw away falls more than the quality of each take rises, and that the whole argument for putting models in timelines rests on that gap.

Related articles