Can LTX 2.3 generate missing frames before an existing video, or do I need a different temporal inpaint workflow?

Short Answer

In short, LTX 2.3 can add frames before an existing clip, but the cleanest route is usually a start-extension workflow, not ordinary inpainting. The important caveats are frame-count alignment, resolution multiples, enough context from the source clip, and manual review for continuity at the join. If you are using a ComfyUI guide-node setup, think of the guide as conditioning placed on a generated timeline, not as magic negative-time frames; the breakdown below shows how to test it without wasting many renders.

The confusion comes from the word "inpaint." In most video tools, inpainting means replacing pixels or a time segment inside an existing clip. What you want is more like a prepend, reverse extension, or temporal outpainting task: generate what happened immediately before frame 0, then make the transition into the original clip look intentional.

Choose The Right Operation

If you only need a few seconds before the clip, use an extension mode that explicitly supports adding video at the beginning. LTX’s own API documentation describes an extend operation that can generate frames at the start or end of a video while using context frames from the input. That is closer to your goal than a retake operation, because retake is designed to regenerate a selected section that already exists inside the source video.

If you are staying inside ComfyUI or another local graph, the practical equivalent is to build a longer output timeline. Put the original video, or at least its first usable context frames, later in the timeline at a valid frame position. Then ask the model to generate the earlier frames that lead into that reference. After that, composite or splice the generated lead-in with the original clip, because many AI video workflows will still touch more frames than you expect.

For social content, this matters because a bad temporal join is very visible in a vertical short. A viewer may not know what an LTX guide frame is, but they will notice a sudden identity shift, camera jump, audio mismatch, or warped hand at the first preserved frame. Before creating a final post, draft the clips in YumCut only after you have a stable source clip and a short list of candidate pre-rolls to compare.

Why Guide Frames Can Fail

The 8n+1 rule is not just a random node requirement. LTX’s open-source documentation says frame counts should follow the 8n+1 pattern, and LTX’s own retake guidance explains that these constraints come from how the model processes video latents. If the source clip, guide clip, or generated timeline is off by a frame, the conditioning can land in the wrong place or fail in a way that looks like a prompt problem.

Resolution is another easy trap. LTX’s documentation repeatedly requires video dimensions to be divisible by 32. If your input is a screen capture, phone export, or cropped vertical clip with odd dimensions, normalize it before debugging guide placement. Use a boring, low-resolution test first, because high-resolution tests make every mistake slower and more expensive.

The deeper issue is directionality. A guide frame at index 0 naturally tells the model what the first frame should look like. Generating frames before that means the original reference no longer belongs at index 0; it belongs after the new lead-in. If the node or pipeline does not support negative indices, you have to shift the reference forward in the generated timeline or use an explicit start-extension endpoint.

A Practical Test Workflow

Start with a tiny experiment: choose a 3-5 second source clip, trim it to a valid frame count, and normalize it to a supported aspect ratio and resolution. Then run one test where the extension is only two seconds. A short lead-in makes it easier to see whether the model understands motion direction, subject identity, camera angle, and lighting before spending time on a longer version.

Prompt the missing beginning as a lead-in, not as a generic scene. For example, describe the camera position, subject motion, and exact action that should end at the source clip’s first moment. Avoid asking for too many new events before the join. The more story you add, the more likely the model is to invent objects or motion that cannot plausibly resolve into the reference video.

Review the join frame by frame. Check the last 12-24 generated frames before the original clip begins, not just the exported MP4 at normal speed. If the subject or camera drifts, try more context, a shorter extension, a stronger first-frame or guide-frame condition, or a simpler prompt. If the generated pre-roll is good but the join is rough, sometimes the best fix is a small manual transition: a cut on motion, a whip pan, a blur, a sound hit, or a one-frame flash that hides unavoidable drift.

When YumCut Helps

YumCut is useful after the technical generation step, when you need to turn the working clip into repeatable short-form output. Once you have a few candidate pre-rolls, test the workflow in YumCut by building versions with different opening captions, pacing, and call-to-action overlays. That is where faceless video workflows save time: not by guaranteeing perfect temporal inpaint, but by making iteration faster once the source material is stable.

Keep manual review in the loop for any AI-generated frame extension. Use AI for options, rough continuity, captions, and fast variant production; use human judgment for identity consistency, factual claims in overlays, brand safety, and whether the first second actually makes people want to keep watching. If the extension is meant for a client, paid ad, tutorial, or product demo, preserve the original clip separately and document what was generated.