LTX 2.3 / audioToVideo

The most distinctive node in the LTX 2.3 family, and the one with no real local-Civitai equivalent: you hand it an audio track plus a prompt, and it generates video synchronized to that audio. It runs on Civitai's cloud through civitai/civitai-comfy-nodes, so the audio you feed it and the clip it returns are the whole transaction - billed in Buzz, no model downloads.

Synced audio is LTX's signature. It's the family trait that turned LTX-2 into the lipsync and music-video pick of the open-video world (the KB's ltx-video doc has a whole arc on it - voice cloning from seconds of audio, mouth shapes that feel real, and the 2.3-era audio cleanup). This node is that capability pointed the other way: audio in, not just audio out. You can make a talking-head clip anchor to a subject reference, or generate visuals that follow the rhythm and energy of a track.

How it works

Standard orchestration flow - videoGen workflow, submit, long-poll, download - with the LTX 2.3 engine running a Comfy workflow server-side. Two things are special. First, source_audio is a required AUDIO socket, so you need an audio-producing node (or the pack's audio loaders) wired in; it's uploaded to Civitai with the job. Second, there's an audio_to_video_attention_scale knob - the tooltip says it's "an audio-to-video cross-attention boost applied via LTX2AttentionTunerPatch. Higher values make audio drive the video more strongly." That's the dial to reach for when the output is ignoring the beat.

Inputs and outputs that matter

• prompt (required) - what the video shows.
• source_audio (required) - the AUDIO socket driving the generation.
• negative_prompt - what to avoid.
• duration - tooltip: "Duration in seconds (3 or 5)".
• width / height - default 1280×720; keep divisible by 32.
• fps - default 24.
• generate_audio - default true; whether the output carries an audio track.
• guidance_scale - default 4; steps - default 20 (8–50).
• model - 22b-dev / 22b-distilled.
• loras - CIVITAI_LORAS socket.
• diffusion_model - CIVITAI_AIR transformer override for community fine-tunes (e.g. SulphurAI/Sulphur-2-base).
• quantity - 1–10 videos per job, distinct seeds per slot.
• reference_image - optional; per the tooltip, "e.g. talking-head subject. When provided, the video is anchored to this image." This is the string input for the subject reference.
• image_guide_strength - 0–1, default 0.7; strength of that image anchoring. Only used when reference_image is set (stage 1; stage 2 is fixed at 0.5).
• audio_to_video_attention_scale - 0–10, default 2; how strongly audio drives the video.
• api_config - auth override.

Outputs: video, workflow_id, raw_json.

Installing it

Shared pack: ComfyUI Manager → Custom Nodes Manager → "Civitai Comfy Nodes" (publisher civitai) → Install → restart, or git clone https://github.com/civitai/civitai-comfy-nodes.git into ComfyUI/custom_nodes and pip install -r civitai-comfy-nodes/requirements.txt. Just requests; nothing to download.

Common issues

• Auth & Buzz. Civitai account with Buzz; CIVITAI_API_TOKEN for headless.
• Audio upload. source_audio uploads to Civitai. Long tracks mean a heavier upload and a longer pre-job wait.
• Sync too weak or too strong? That's the audio_to_video_attention_scale knob - low values drift from the audio, high values make audio dominate.
• Don't expect it to transcribe. Audio-to-video uses the track as a driver, not a transcript; prompt adherence for specific spoken content is still the model's weak axis. Test before committing to a full lipsync pipeline.
• Early preview. Behavior may change without notice; pin the version.