Twenty seconds. Four shots. Now on a 16 GB Mac.

LTX 2.5 · Multi-shot · Lip sync

A real music-video take, across every supported Mac.

This is not four clips pretending to be one performance. LTX generated all 481 frames continuously—and Kynd now makes this Social recipe available from 16 GB upward.

Actual Kynd output · 20 seconds · 481 frames · 512×512 · sound on

Most AI music-video workflows make you choose between two fragile options. Generate one long shot and accept a motionless camera, or generate several short clips and try to hide the changes while identity, lighting and lip sync reset at every join.

Kynd’s multi-shot LTX 2.5 route takes a third path: describe several distinct camera setups inside one continuous generation, keep the real audio on one unbroken timeline, and let the model carry the performance across every change of view.

The camera cuts. The song does not.

Four discrete setups from the same generated take

The prompt names four visual beats:

• A wide establishing shot that shows the performer, guitar and camper-van interior.
• A straight-on close-up that brings the mouth and expression forward.
• A left profile framed against the window.
• A low angle looking up beneath the performer’s chin.

The camera does not glide between them. It is repositioned, so the image changes with the grammar of an edit rather than the wobble of an impossible camera move. Because all four setups live inside one generation, the model can preserve the character, clothes, instrument, set and light as the viewpoint changes.

Three ingredients hold the performance together.

1 · A character reference

The approved image establishes the performer before generation begins: face, materials, wardrobe and overall design. It gives every setup the same identity to return to.

2 · The real vocal track

Kynd does not ask the model to imagine a generic song and add sound later. The supplied vocal stem is locked to the generation timeline. The mouth is conditioned by the actual words, rhythm and phrasing the viewer hears.

3 · A camera plan written as time

The prompt does more than describe a pretty frame. It tells the take when each setup should take over, giving LTX a simple edit plan while the audio continues beneath it.

Together, those inputs turn a prompt into something closer to direction: who is performing, what they are performing, and where the camera should be.

Why the lip sync survives the cut.

Traditional post-production lip sync often solves each clip separately. That works until the edit lands mid-word: one shot closes a mouth while the next opens on a different phoneme, and the performance feels stitched together even when every individual clip looked acceptable.

Here, the audio clock never restarts. LTX receives one continuous vocal condition for one continuous 481-frame take. A camera change can alter the size and angle of the face, but it does not create a new performance timeline.

That is the important distinction. The shots change inside the performance; the performance is not rebuilt around the shots.

The memory wall was never the creative limit.

The old resident route tried to keep too much of the pipeline in memory at once. Kynd changed the engine instead of shrinking the idea.

Kynd now loads the text encoder only for the phase that needs it, streams transformer blocks through denoise, and decodes the video in temporal tiles. In a full-length 481-frame validation of that memory-critical path, denoise plus decode peaked at 9.15 GB. The product still reserves a conservative 13 GB envelope and opens Social at 16 GB, leaving room for macOS.

Social now works across the board: 16, 24, 32 and 64 GB. The model, weights, steps, frame count and audio conditioning do not change between tiers. A larger Mac adds headroom; it does not unlock a better version of the film.

Same creative tier. Smaller memory footprint.

• Music videos can develop. A performance does not need to remain trapped in one composition for the whole lyric.
• Dialogue can have coverage. Establishing shots, close-ups and profiles can share one spoken timeline.
• Identity gets a better chance. One continuous take avoids restarting the character from fresh noise at every edit.
• The audio remains the spine. Camera choices serve the rhythm without breaking the vocal condition.
• The edit starts in the prompt. The creator can think in shots and beats before committing to a long render.

Most importantly, none of this is a special “16 GB quality” mode. The memory work changes how the pipeline is held, not what the creator is allowed to make. It does not replace an editor; it gives the editor a more coherent piece of material to begin with.

What the 16 GB breakthrough does—and does not—mean.

Sixteen GB has less spare room, not a lesser creative tier. Kynd unloads the chat model before the render, streams transformer blocks instead of keeping all of them resident, and decodes the video in temporal tiles. Those memory techniques preserve the same generation recipe across supported Macs.

Two-stage HQ is a separate route. The 16 GB promise here applies to the one-stage Social Fast Path. The optional two-stage / refined path has a different memory profile and remains gated to larger machines.

A prompt is direction, not a frame-accurate edit decision list. LTX can understand timed camera language, but it may place a transition a little earlier or later than requested.

Lip sync begins with the source. A clear, isolated vocal gives the model a better signal than a dense master where the voice competes with drums and instruments.

Continuity is stronger, not magical. A single take helps identity and set coherence, but generative video can still alter small details as time passes.

Creativity should begin with the shot, not the system requirements.

The breakthrough is not simply that an AI model can show four angles. It is that those angles can belong to the same person, in the same place, performing the same real song without the mouth losing its place—and that the full experience can now begin on a 16 GB Mac.

That moves local video one step away from novelty clips and one step toward filmmaking: reference, performance, coverage and timing—planned together, rendered together, kept on your Mac.

Also useful: How Kynd brought LTX 2.5 to 16 GB Macs → · Why RAM should not decide what you can create →