LTX 2.3 Inpainting, Retake and Extend Explained

Nine seconds are perfect. One second is wrong. Almost everyone's instinct is to change the seed and generate the whole clip again — which is the most expensive possible response to a small problem.

LTX 2.3 has an editing story, but it is not the one image-model habits predict. There is no "mask this region and regenerate" node in the shipped inference pipelines. What exists instead is a temporal editing tool that is arguably more useful for video: mask a time window, regenerate only that, keep everything else.

This guide covers what you can genuinely edit today, and — because search results are full of confident claims otherwise — where "inpainting" actually appears in the official codebase. Sourced from the LTX-2 pipeline documentation, the ltx-trainer package, and LTX-Desktop, checked July 2026.

Retake: masking time instead of space

RetakePipeline is the editing tool LTX 2.3 actually ships, and the mechanism is worth understanding rather than treating as a black box.

Per the docs, it is single-stage generation that:

1. Encodes the source video and audio into latents
2. Applies a temporal region mask marking [start_time, end_time] for regeneration
3. Denoises only the masked region from a text prompt
4. Preserves content outside the time window

Required CLI arguments: --video-path, --start-time, --end-time.

Two capabilities that make it more than a convenience:

Independent video and audio regeneration. Separate regenerate_video and regenerate_audio flags mean you can fix a line of dialogue while leaving the picture untouched, or fix a visual glitch and keep the audio you already approved. In a joint audio-video model, that separation is not obvious — and it is exactly what editing real material requires.

Either checkpoint. It can use the full model with CFG guidance, or the distilled model with a fixed sigma schedule. Cheap fixes on distilled, careful fixes on dev.

The constraints will reject unprepared footage: the source video's frame count must satisfy 8k+1 (97, 193, …) and its resolution must be a multiple of 32. Camera footage or output from another model almost always needs re-encoding first. Do that step before you start debugging arguments.

Retake is also exposed in LTX Desktop as "video edit generation," so you can use it without touching the CLI — though note the app lists Retake among its paid API features.

Where inpainting actually lives

Here is the correction worth the price of admission.

Search for "LTX 2.3 inpainting" and you will find tutorials implying a masked-region video inpainting workflow. In the official codebase, inpainting appears in a specific and different place: the trainer.

The ltx-trainer documentation describes a flexible conditioning framework encompassing text-to-video, image-to-video, video and audio extension, inpainting, outpainting, and IC-LoRA variants. The shipped configs include audio_inpainting_lora.yaml, audio_extend_lora.yaml, audio_suffix_lora.yaml, and video_extend_lora.yaml.

Read that carefully, because the distinction matters:

• Inpainting is a conditioning mode you can train an adapter for. It is part of the training framework.
• It is not one of the eleven shipped inference pipelines. There is no inpaint.py alongside retake.py.

So the honest state of play: if you want spatial inpainting on LTX 2.3 today, you are either training an adapter for it, or you are using a community node graph that composes the pieces itself. What you get out of the box is temporal editing via Retake.

That is not a downgrade — for video work, "regenerate this second" solves more real problems than "regenerate this rectangle." But it is different from what the search results promise, and knowing which you have saves an evening.

Iterating on shots rather than repairing them? Editing tools matter most once you are committed to a take. If you are still exploring, generate quick variants at ltx23.app and save the repair workflow for the shot you actually keep.

Extending clips instead of generating longer

The other editing need — "make it longer" — has its own answer, and it is not "raise the frame count."

Longer generations mean more tokens, more attention cost, and more room for identity and motion to drift. The official framing points to extension: video extension and audio extension / suffix are supported conditioning modes in the trainer's config set (video_extend_lora.yaml, audio_extend_lora.yaml, audio_suffix_lora.yaml).

Practically, the workflow that behaves well:

1. Generate a short clip you are happy with
2. Extend from it rather than regenerating longer from scratch
3. Retake any window that goes wrong
4. Assemble in an editor

Which, notably, is how LTX Desktop is shaped — it ships a video editor interface and editing projects, not just a generate button. The official tooling assumes you assemble rather than one-shot.

Keyframe conditioning is an editing tool too

Worth including because it is filed under generation and behaves like editing.

The conditioning documentation distinguishes two image-conditioning methods:

• Replacing latents — replaces the latent at a specific frame with your encoded image. Strong control over specific frames.
• Guiding latents — adds the image as a guiding signal rather than replacing it. Better for smooth interpolation between keyframes.

That is a real editorial choice: pin a frame exactly, or nudge toward it. KeyframeInterpolationPipeline uses the additive guiding approach for smoother transitions, which is why it feels different from plain image-to-video even on the same input.

If you need a shot to hit an approved frame at a specific moment, that is replacement conditioning — and it is closer to editing than to prompting.

A repair workflow that scales

And the preparation that prevents the most common failure: make your source legal first — resolution a multiple of 32, frame count on the 8k+1 pattern. Every editing pipeline that takes a video enforces it.

FAQ

Does LTX 2.3 support inpainting? Inpainting is a conditioning mode in the official trainer's framework, not a shipped inference pipeline. The out-of-the-box editing tool is RetakePipeline, which masks a time region rather than a spatial one.

How do I fix one bad second of a clip? RetakePipeline with --start-time and --end-time. Everything outside the window is preserved.

Can I regenerate only the audio? Yes — Retake exposes independent regenerate_video and regenerate_audio control.

Why is my video rejected by Retake? Frame count must be 8k+1 and resolution a multiple of 32. Re-encode first.

How do I make longer videos? Extend from a short, good clip — video and audio extension are supported conditioning modes — rather than pushing frame count in one generation.

Is Retake available without the CLI? Yes, LTX Desktop exposes it as video edit generation; the app lists it as a paid API feature.

Bottom line

LTX 2.3's editing model is temporal, not spatial: mask a time window with Retake, extend rather than over-generate, and use replacing-vs-guiding latent conditioning when a specific frame must land. Spatial inpainting exists as a training-side conditioning mode, not as a ready-made pipeline — and any tutorial claiming otherwise is describing something it built itself.

Adopt the repair habit and your cost per usable clip drops immediately. And for the exploration phase before any of that, ltx23.app is the cheapest place to iterate.

Sources

Verified July 2026 against primary documentation:

1. LTX-2 available pipelines documentation — RetakePipeline mechanism, flags, constraints, checkpoint options
2. ltx-trainer documentation — conditioning framework including inpainting, outpainting, extension
3. ltx-trainer configs — video_extend_lora, audio_extend_lora, audio_suffix_lora, audio_inpainting_lora
4. LTX-2 conditioning documentation — replacing vs guiding latents
5. Lightricks/LTX-2 — official GitHub repository — pipeline list and package layout
6. Lightricks/LTX-Desktop — Retake as video edit generation, editor interface, paid API features
7. Lightricks/LTX-2.3 — Hugging Face model card — ÷32 and 8k+1 constraints
8. ComfyUI official LTX-2.3 tutorial — native templates and what they cover
9. LTX-2 installation & CLI flags — running the retake module
10. ltx-core package documentation — conditioning utilities
11. LTX-2 technical paper (arXiv 2601.03233)

Capabilities described are those exposed by the official pipelines at the time of writing; community node graphs may compose additional behavior.