Audio Extender (segment + overlap)

The audio side of LTX-2 segment extension

When you extend an LTX-2 video segment by segment, every pass needs audio that lines up with the frames it's generating - but the "unique" part of a segment is only the new frames, not the overlap. This node is the audio bookkeeper for that loop: it slices the timeline's audio into two chunks - the conditioning_audio (which includes overlap context) and the segment_audio (just the unique region) - so the model gets the right audio to condition on and the stitcher gets the right audio to lay down.

The controls that matter

Three things shape the slice:

• mode - how much context surrounds the unique region: left_context_only (default), right_context_only, symmetric_context, or no_overlap. For extension, left context is almost always what you want: the model needs to know what just played, not what's coming.
• left_overlap_s (0.5) / right_overlap_s (0) - the actual seconds of context. These should match your video overlap in time so audio and frames stay in step.
• link_mode - how the chunk is tied to the generation timeline. use_timeline_cursor (default) trusts an explicit cursor from the extension math; use_generated_frames / use_extension_frames derive timing from the pass's frame counts; independent_segments ignores the timeline.

The optional numeric inputs are the real interface to the pack's extension system: segment_index, segment_duration_s, video_frames, generated_frames, extension_frames, timeline_cursor_frames, segment_start_frames, and effective_unique_frames - most of them fed by IAMCCS_AudioExtensionMath (the tooltips name it explicitly). clamp_policy (soft_clamp default) decides what happens when the requested range overruns the audio track: trim to bounds, or raise (strict).

Outputs

conditioning_audio and segment_audio (the two slices - wire conditioning into the LTX audio conditioning path and segment audio into the stitcher), segment_duration_s_out / conditioning_duration_s_out, and sample-level bounds: segment_start_sample / segment_end_sample / conditioning_start_sample / conditioning_end_sample. The report string confirms the math.

In the workflow

This is a helper for the pack's LTX-2 extension and SuperNode pipelines - the SuperNodes requirements list it as an expected helper node. It's not something you'd drop into a random graph for fun; it shines when you've got a multi-pass extension loop and your audio keeps drifting out of sync. The classic failure is mismatched numbers: if left_overlap_s says 0.5s but your video overlap is 9 frames at 24fps (≈0.375s), the audio and frames fall out of step over many segments. Keep the overlap values consistent with the video side and the report output will confirm it.

Install is the pack install - ComfyUI Manager, search "IAMCCS", or git clone https://github.com/IAMCCS/IAMCCS-nodes.git into custom_nodes, restart. No model downloads. It expects standard ComfyUI AUDIO objects, so anything that loads audio (VHS, core loaders) feeds it fine.