LTX Audio to Video (Replicate Fixed) 🎤✅

Same trick as the pack's main audio-to-video node - image plus audio becomes a talking video - but the rendering happens on Replicate instead of on ltx.video's own API. If you already have a Replicate account with credits and a r8_... token, this is the variant you reach for. If you don't, the native LTXAudioToVideo node is the simpler first stop.

That "(Replicate Fixed)" in the name is worth explaining, because it's the reason the node exists. An earlier incarnation of this idea tried to jam the image and audio into the Replicate call as base64 data. That blows up - huge payloads, serialization edge cases, jobs that fail for no visible reason. This node "fixes" it the boring, correct way: it uploads the image and the converted audio to a public URL first, then hands Replicate those URLs. Every other wrapper since should've learned the same lesson.

How it works

The flow mirrors the native node, with a different backend:

1. Your image becomes a JPEG (auto-resized to a max of 1920px).
2. Your audio (an AUDIO tensor - connect a Load Audio node) becomes an MP3 via ffmpeg → torchaudio → pure-Python fallback.
3. Both files go up to catbox.moe as anonymous temporary hosts.
4. The node POSTs to api.replicate.com/v1/models/lightricks/ltx-2.3-pro/predictions with Prefer: wait, then polls until the prediction succeeds (up to about 4 minutes).
5. It downloads the finished MP4, saves it to ComfyUI's output/ folder, and decodes the frames.

Inputs and outputs

Required: replicate_api_key (the r8_... token - the node is clear that the Replicate key is what it wants, not an ltx.video key), image, audio, and prompt.

Optional and worth knowing:

• model - here you do get ltx-2-3-fast as an option alongside ltx-2-3-pro, so this is the audio node to use if you want cheap quick tests.
• aspect_ratio - 9:16, 16:9, or 1:1. Note the native node is stuck at just portrait/landscape; this one adds square.
• fps and seed - seed -1 means random.

Outputs are the pack standard: frames (IMAGE), video (VIDEO, native preview), and fps_out (FLOAT, defaults to 25).

Install

Same pack, same story - 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, grab a Replicate token from your account settings, and paste it in.

The honest tradeoffs

You're paying Replicate per prediction, and your files pass through catbox.moe, a public anonymous host - so don't feed it anything you'd mind the internet briefly hosting. The native node uploads to LTX's private cloud storage instead, which is the more private path. Pick Replicate when you have credits to burn and want the fast model or square aspect; pick the native node when you want one less third party in the chain.

Failures are mostly boring: a wrong or missing token raises immediately, and a long job can hit the poll timeout. Re-run, and if catbox is having a bad day the upload step throws before anything is charged.