LTX 2.5 / audioToVideo

Most of the LTX 2.5 nodes start from text or an image. This one starts from audio. You wire in a track through the source_audio socket, and the model generates video that moves to it - the tooltip's phrase is "audio-to-video cross-attention boost," and the whole node is built around how strongly that audio drives the visuals. If you've ever watched a clip whose beats, lip movements and gestures land exactly on the soundtrack, this is the node that does that - and the two knobs unique to it are the ones worth understanding before you spend any Buzz.

Like every node in civitai-comfy-nodes - Civitai's official ~160-node pack generated from its Orchestration API spec - it's a cloud job. LTX 2.5 runs through Comfy workflows on Civitai's fleet (the tooltip even names the LTX2AttentionTunerPatch under the hood), billed in Buzz. Submit, long-poll, get a native VIDEO tensor.

Inputs that matter

• prompt (required) and source_audio (required AUDIO socket) - the content and the beat it follows.
• audio_to_video_attention_scale - default 2.0. The headline knob: higher values make audio drive the video harder. Too low and the visuals ignore the track; too high and motion gets slaved to every transient.
• reference_image (STRING) + image_guide_strength (default 0.7) - anchor the video to a subject, e.g. a talking-head still. When a reference image is provided, image_guide_strength controls the stage-1 conditioning (stage 2 is fixed at 0.5 per the tooltip).
• duration (3–20s, default 5), width/height (1280×720), fps (24), seed, guidance_scale (4), steps (20) - the usual LTX 2.5 surface.
• generate_audio - on by default; relevant here because LTX's native audio can be re-synced rather than dropped.
• model (22b-dev), loras, diffusion_model (fine-tune override), quantity.

Outputs: video (VIDEO), draft_cache_audio (AUDIO), draft_cache_video (VIDEO), workflow_id and raw_json (STRING).

Install and auth

Standard for the pack - official, Comfy Registry:

# ComfyUI Manager: search "Civitai Comfy Nodes"
cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Restart, authenticate via a Civitai Auth node, CIVITAI_API_TOKEN, or sidebar OAuth.

Common issues

Attention scale is a sweet-spot dial, not a "more is better" one. Start at the 2.0 default and move in small steps; slamming it up produces twitchy, over-slaved motion. reference_image is a STRING field (URL/data URL/base64), not an IMAGE socket - same gotcha as the first/last-frame nodes; people try to wire a Load Image in and it won't connect. Audio quality in the source matters - the model syncs to what it's given, and a muddy track gives muddy sync. LTX audio remains the family's fragile side; if the output audio wobbles, that's a known weakness, not your settings. Your audio and frames leave your machine, per-second billing applies, and the pack is early preview - behavior can change without notice.