When audio and video latent lengths differ, LTXVConcatAVLatent adjusts the audio to fit the video: longer audio is trimmed and shorter audio is zero-padded. Before using the result, decide whether losing the long-audio tail is acceptable; for short audio, check that the padded region remains unmasked so the model can generate it. If batch, channel, or other incompatible dimensions prevent fitting, stop and resolve the inputs rather than assuming the node will repair them.
Decide and prepare the inputs
- Identify the actual
video_latentandaudio_latentconnected to the node. Check their displayed type labels; node color alone is only a visual clue and does not establish a compatible input type. - Record the lengths as shown by the workflow or by your own inspection method. Compare them without assuming a particular unit or inferring a frame-to-audio conversion that the documentation does not specify.
- If audio is longer, identify what content lies beyond the video’s fitted duration. Proceed only if losing that portion is intentional. If it contains a word, musical cue, or other material you need, shorten or otherwise prepare the audio before concatenation, or revise the video length.
- If audio is shorter, decide whether the remaining interval should be generated. After concatenation, inspect the output
noise_maskwhere one is available and confirm the padded area is unmasked. The node documentation says this region is left unmasked so the model can generate it; do not treat zero padding as a preserved recording of silence. - Confirm the inputs can be fitted before running the workflow. The documented failure cases include disagreement in more than one dimension and disagreement in batch or channel dimensions. If these occur, stop and make the latent inputs compatible upstream; do not count on automatic resizing or batch repair.
The node combines the streams in a nested tensor. If the video latent already contains audio, the node keeps its video stream and replaces its audio with the supplied audio latent. This makes it important to verify which audio you intend to retain.
Hypothetical acceptance worksheet
The figures below are assumed observations for planning, not measured ComfyUI results. “Length units” means values reported by the creator’s inspection method; the documentation does not define a conversion formula here. Use the table to record the decision and then inspect the actual output.
| Case | Video length | Audio length | Expected length action | Review and decision |
|---|---|---|---|---|
| Long audio | 120 units | 150 units | Trim audio to fit video | Assume the last 30 units contain a required spoken phrase; reject this pairing unless that loss is acceptable or the inputs are revised. |
| Short audio | 120 units | 90 units | Zero-pad the shorter audio | Inspect the padded interval in the output mask; accept for generation only if it is unmasked and generation there is intended. |
| Dimension mismatch | 120 units | 90 units | Length adjustment may not be possible | Assume batch differs; stop and correct upstream rather than treating the length match as sufficient. |
This worksheet is a creator’s comparison method, not an extra node feature or a guarantee about exact output dimensions. The documented contract covers trimming, zero-padding, mask handling, and errors for incompatible shapes; it does not specify how to interpret a custom inspection tool’s length units.
Limits and what to do when it fails
Trimming is irreversible for the audio portion excluded from the fitted latent. Zero-padding supplies no original audio in the added interval. The documented unmasked padded region is available for model generation, but that is not a promise that generated audio will match a particular continuation or sound natural. Review the result against the intended scene and dialogue.
When the node reports that the inputs cannot be fitted, note which dimensions differ and compare batch and channel first. Then prepare compatible latents upstream and retry. If the output mask is absent, the documentation says a missing input mask is replaced with an all-ones mask matching that input’s samples when a combined mask is formed; it does not describe every downstream model’s behavior. Check the output that your workflow actually passes onward. The node documentation does not specify a separate control for preserving trimmed audio or changing the padding policy.
Common questions
Does the node stretch long audio to match the video? No. The documented behavior is trimming, not time-stretching. Prepare a different audio latent if timing changes are required.
Does zero-padding mean the added interval is silent audio? It means zeros are added to the shorter latent. The documentation says the padded area is unmasked for model generation, so do not assume it is a fixed, preserved silence.
What if the video latent already has audio? The node keeps the video stream and substitutes the supplied audio stream. Confirm that the supplied audio is the one you want before combining.
Sources
Sources and next step
This guide interprets public documentation and uses explicitly hypothetical acceptance examples, not application tests. Controls and availability can change; verify the cited manual against your installed version.
For work created in Imagild, carry the input requirements and acceptance checklist into the film workspace. The external software controls discussed above belong to their respective products.