LTX Retake Section 🔁🎬

AI video is a game of whack-a-mole: you finally get a clip where the lighting works and the motion is right, and then one three-second window is garbage - a face that melts, a hand that turns into spaghetti, a sentence that mumbles. The instinct is to regenerate the whole thing and pray the rest survives. This node is the smarter move: it tells the LTX API to re-render only the section you mark, leaving the parts that worked alone.

It hits /v1/retake with your video, a start time, and a duration, and the API redoes that window. Pro models only, and this is the most surgical tool in the pack - the counterpart to LTXExtendVideo, which adds new footage at the ends. Retake edits from the inside.

Inputs and outputs

Required: api_key, video_url, start_time (seconds, where the bad section begins), and duration (the section length, minimum 2 seconds).

Optional:

• prompt - describe what should happen in the retaken section. This is your lever for fixing the specific problem: "her hand returns to her lap" instead of a generic re-roll.
• model - ltx-2-3-pro or legacy ltx-2-pro, no fast tier.
• mode - the interesting one. replace_audio_and_video (default, the whole section), replace_video (keep the existing audio, redo the picture), or replace_audio (keep the video, redo the sound). If the only sin is a mumbled line, re-rendering just the audio is dramatically cheaper than redoing both.
• resolution - limited to 1920x1080 or 1080x1920. The tooltip says it's auto-detected if omitted, but you're stuck at 1080p, which is fine - the surrounding footage isn't going to be meaningfully sharper than that anyway.

Outputs: frames (IMAGE), video (VIDEO, native preview), fps (FLOAT, fixed at 25).

Install

ComfyUI Manager (search "comfyui-ltx-node"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/PauldeLavallaz/comfyui-ltx-node
pip install requests pillow

Restart, paste your key from ltx.video/api-keys.

The same URL gotcha, and one cost note

Like the extend node, video_url wants a hosted video - an HTTPS URL or an ltx:// storage URI, not a local file path. Get the clip somewhere reachable before you wire this in. Cost-wise this is friendlier than it looks: you pay for a retake, which is one short render, versus paying to regenerate a full clip. The replace_audio mode is the budget MVP - if the picture survived and the track didn't, retaking just the audio is the cheapest fix in this entire pack. Just don't expect the section boundaries to be invisible; check the joins, and don't be afraid to nudge start_time by half a second if there's a visible seam.