LTX Video Editing Models: The Essentials
Last updated: September 15, 2026
The LTX editing family from Lightricks works on footage you already have instead of generating from scratch: Extend Video adds new seconds to a clip, Retake regenerates a targeted segment, Quality Render to Real turns a 3D or game render into photoreal video, and Video to Video restyles a whole clip under structural guidance. Every model here takes a video as its main input and keeps the parts you did not ask it to change.
Which Model Should I Use?
Model | What it does | Best for |
Adds continuation footage to a clip | Up to 4K, up to 20 extra seconds, Match Input sizing | |
Adds continuation footage to a clip | Simple 6, 8, or 10 second extension at the start or end | |
Regenerates a segment of a clip | Fixing performance, timing, or framing without a full re-render | |
Regenerates a segment of a clip | Choosing exactly what to replace: audio, video, or both | |
Converts a 3D/game/CG render to photoreal | Turning a viewport capture or engine preview into a finished plate | |
Restyles an entire clip | Applying a new look across a whole video with structure control (canny, depth, or pose) |
Extend Video: Seamless Continuity
Extend Video expands existing footage while keeping character identity, lighting, and art style consistent, so the new frames flow naturally from the source clip.
LTX-2 19B Extend Video
The LTX-2 19B Extend Video model adds up to 20 seconds to the start or end of a clip, at up to 4K, with a Match Input option that copies the source's resolution and aspect ratio automatically. Video Strength (0 to 1) controls how closely the new frames stick to the original clip's visual data, Context Duration sets how much of the preceding video the model reads before generating, and camera motion presets (Dolly, Jib, or Static) can steer the new segment. When Generate Audio is on, the extension gets synchronized soundscape to match.
LTX 2.3 Pro Extend Video
The LTX 2.3 Pro Extend Video model keeps the same idea simpler: pick 6, 8, or 10 seconds of new duration and whether it lands at the start or end of the clip, describe only the new action in the prompt.
To keep continuity on either model: set Aspect Ratio and Resolution to Match Input, describe only the new action (not the whole scene) in the prompt, and raise Video Strength on the 19B model if the new segment feels detached from the source.
Retake: Cinematic-grade Editing and Shot Retakes
Retake regenerates only the segment you target, replacing audio, video, or both, while the rest of the clip, and its motion, lighting, and lip sync, continues untouched.
LTX-2 Retake
The LTX-2 Retake model operates on clips up to about 20 seconds and typically modifies segments from 2 to 16 seconds. File uploads are capped at 100 MB and 20 seconds; output is 1080p (4K uploads get downscaled).
Use it to fix a gesture or timing mismatch, try an alternate camera version, shift the emotional tone of a moment, or polish a rough cut, all without rebuilding the whole shot.
LTX 2.3 Pro Retake
The LTX 2.3 Pro Retake model adds an explicit Retake Mode choice: Replace Audio only, Replace Video only, or Replace Audio and Video. Set a Start Time and Duration (2 seconds minimum) to bound exactly which part of the 1080p clip gets touched.
Quality Render to Real: CG and Game Renders to Photoreal
The LTX 2.3 Quality Render to Real model is built for one job: taking gray clay, CG, or game-engine previews and turning them into photoreal motion. It follows the timing, camera, and staging of a render you already have, then re-skins it with real-world materials, lighting, and detail. An optional First Frame Image fixes the materials and lighting of the opening frame; Intensity (Light, Strong, or Strong V2) controls how far the result departs from the original render; Detail Refine is an optional second pass that doubles width and height for sharper texture; audio can be generated in the same run.
Slider comparison: 3D render (left) vs. photoreal generated result (right), boy walking a rain-soaked alley at night · Open on Scenario
Slider comparison: 3D render (left) vs. photoreal generated result (right), orc on a smoky battlefield with Strong V2 intensity · Open on Scenario
Slider comparison: 3D render (left) vs. photoreal generated result (right), hilltop restaurant flythrough · Open on Scenario
Restyle: Video to Video
The LTX-2 19B Video to Video model restyles an entire clip while keeping its structure: a reference video conditions the output, with an IC-LoRA Modality control (canny edges, depth, pose, or detailer) available for stricter structural guidance, plus an optional first-frame image and independent camera motion presets. It is a legacy model kept for this one use case: a full-clip restyle rather than a targeted extend or retake.
Parameters
video
Required on every model in this family: the clip to extend, retake, restyle, or convert to photoreal.
prompt
On Extend Video, describe only the new action for the added segment, not the whole scene. On Retake, describe the edit you want in the targeted window. On Render to Real, describe the realistic result you want; a negative prompt is pre-filled with common artifacts to avoid.
videoStrength
19B Extend Video and 19B Video to Video. 0 to 1: how closely new or restyled frames must match the input clip's visual data. Higher locks in more of the source.
extendDuration / duration / extendMode
Extend Video's length controls. The 19B model takes a free extendDuration up to 20 seconds; 2.3 Pro takes a fixed duration of 6, 8, or 10 seconds plus an extendMode of start or end.
retakeMode, startTime, duration
2.3 Pro Retake. retakeMode chooses what gets replaced (audio, video, or both); startTime and duration (2 seconds minimum) bound the edited window.
resolution
Auto (Match Input) is recommended on the 19B Extend and Video to Video models so the output exactly matches the source's dimensions. Retake outputs 1080p regardless of the source (4K uploads are downscaled).
modality / conditioningVideo
19B Video to Video only. IC-LoRA structural control: none, canny, depth, pose, or detailer, with a matching pre-processed conditioningVideo when a specific control map is available.
intensity
Quality Render to Real. Light, Strong, or Strong V2 (default): how far the photoreal result departs from the source render.
enableDetailRefine
Quality Render to Real. An optional second pass that doubles width and height for sharper detail; pair with detailRefineStrength to control how much it reworks the image.
cameraMotion
Available on the Extend Video and Video to Video models: apply Dolly, Jib, Static, or (on 19B) other named presets to the generated portion.
generateAudio
On Extend Video and Quality Render to Real: generate a synchronized soundtrack for the new or converted footage in the same run.
Use Cases
Lengthening a hero shot: Extend Video adds a natural continuation when a clip needs to run longer than the original take.
Client revision passes: Retake applies a targeted fix (a gesture, a line reading, a camera angle) without rebuilding the whole shot, keeping review cycles fast.
Game and CG previz to final plate: Quality Render to Real converts an engine viewport capture into a photoreal shot for a trailer or cutscene.
Whole-clip restyling: Video to Video reskins an entire take (a new art direction, a different lighting mood) while keeping the original staging via structural conditioning.
Dialogue and audio swaps: 2.3 Pro Retake's Replace Audio mode changes a line or sound cue without touching a single visual frame.
Tips for Better Results
Set Aspect Ratio and Resolution to Match Input on Extend and Video to Video. This is the single most reliable way to avoid a visible seam between original and new footage.
On Extend, describe only the new action. The model already has the preceding context; re-describing the whole scene competes with that context instead of building on it.
Raise Video Strength if a continuation feels detached. It is the direct dial for locking new frames to the source clip's look and motion.
Keep Retake windows tight. A 2 to 6 second edit window lands more reliably than trying to retake a long, fast-changing passage in one pass.
Feed Render to Real a first-frame reference when the look matters. Fixing materials and lighting on frame one anchors consistency for the rest of the clip.
Use the pre-filled negative prompt on Render to Real rather than clearing it: it targets artifacts specific to render-to-photoreal conversion.
Reach for a conditioningVideo on Video to Video (canny, depth, or pose) when the restyle needs to preserve exact structure, not just rough composition.
Known Limitations
Retake output is capped at 1080p. A 4K source gets downscaled; higher resolutions are not yet available on this model.
Retake uploads are capped at 100 MB and 20 seconds, with edit windows from 2 to 16 seconds.
Render to Real's resolution trades off against length: 720p covers clips up to about 6 seconds, 480p up to about 15 seconds.
LTX-2 19B Video to Video is a legacy model (tagged as superseded), kept specifically for full-clip restyling with structure control; it is not the right choice for a simple extend or retake.
Very fast or complex action across a Retake or Extend boundary may need more than one iteration to match the surrounding continuity.