LTX 2.3 / extendVideo

LTX 2.3's video extension node. Give it a source clip and a prompt, and it continues the video - same look, same motion vocabulary, new frames - running entirely on Civitai's cloud through the official civitai/civitai-comfy-nodes pack. It's the LTX answer to the eternal "five seconds isn't enough" problem, and it's one of the simplest nodes in the family to understand: the one new dial is num_frames, which says how much footage to add.

Extension is where LTX has always been quietly good. The 0.9.5 release way back added frame interpolation and video extension, LTX-2's "continue a video" behavior got called scary-accurate for carrying voices and motion across the seam (the KB's ltx-video doc has the receipts), and 2.3's rebuilt VAE made the hand-off noticeably cleaner. On Civitai's hosted 22B engine, you get that continuation without the 32GB+ VRAM the local model officially wants. The trade, as always with this pack: per-run Buzz cost and your source clip living on Civitai's servers while the job runs.

How it works

Standard orchestration plumbing - videoGen workflow with the LTX 2.3 engine, submit to the Civitai Orchestration API, long-poll, download the MP4 as a VIDEO. The source video is uploaded with the request and conditioned on as the first segment; the hosted Comfy workflow generates num_frames more and stitches the extension on. Frame-count rules for LTX matter even here: LTX 2.3 wants frame counts divisible by 8 plus 1, which is why the default sits at 121 (and the 9–481 range's bookends land on the formula).

Inputs and outputs that matter

• prompt (required) - what happens next. Continuation quality tracks prompt quality: describe the ongoing action, not a new scene.
• source_video (required) - the VIDEO socket with the clip to extend.
• negative_prompt - what to avoid.
• num_frames - default 121, range 9–481. More frames = longer extension = more cost and queue time. This is the input to spend your time on.
• duration - the extension's seconds (tooltip: 3 or 5); note it coexists with num_frames, so the engine resolves length from both.
• width / height - default 1280×720; divisible by 32.
• fps - default 24; steps - default 20 (8–50); guidance_scale - default 4; generate_audio - default true.
• model - 22b-dev / 22b-distilled.
• loras - CIVITAI_LORAS socket; diffusion_model - transformer override for fine-tunes.
• quantity - 1–10 extensions per job, distinct seeds.
• api_config - auth override.

Outputs: video, workflow_id, raw_json.

Installing it

ComfyUI Manager → Custom Nodes Manager → "Civitai Comfy Nodes" (publisher civitai) → Install → restart, or clone into ComfyUI/custom_nodes and pip install -r civitai-comfy-nodes/requirements.txt. Only requests; no models.

Common issues

• Auth & Buzz. Civitai account with Buzz; CIVITAI_API_TOKEN for headless installs.
• Seam jumps. Extensions can drift at the cut point. Lower guidance_scale a touch or add a describing prompt; LTX 2.3 is better here than 2.x, but it's not perfect. Keeping num_frames modest and extending iteratively (extend the extension) usually gives cleaner results than one giant jump.
• Upload cost. The source clip uploads to Civitai each run; short and compressed is kind to your uplink and your patience.
• Don't confuse duration and num_frames. Both set length; if you set one and the other still says default, you may get more (or less) footage than you expected. Set both deliberately.
• Early preview. Behavior may change without notice; pin the version.
• Cost. Longer extensions bill like longer jobs - read the per-run cost on the node.