To check a separated LTXV audio-visual latent, first confirm the documented positional rule: the input samples batch must have at least two elements, with the first used for video and the second for audio. Then use available upstream construction records, metadata, or an inspection tool that actually exposes identity to trace video_latent and audio_latent. If no such evidence is available, record the output identity as unverified. For a present noise_mask, the official rule is that it is split in the same way; independently observing that association requires a tool or metadata that exposes it. A connected output only proves that a value was routed; it does not prove content identity or independently verify mask pairing.
Establish the input contract and trace each output
The ComfyUI node documentation defines the split along the first, or batch, dimension: element one supplies video and element two supplies audio. The input must have at least two elements there. The documentation does not specify how extra elements are handled, so do not assume they become additional outputs or are preserved.
Use the node’s LATENT type labels to identify the input and output ports. Color can be a visual clue, but verify the displayed type rather than inferring it from color alone. Before wiring downstream processing, write down what the first two input elements represent. If their order is unknown, resolve that upstream; the separator cannot infer intent from content.
- Inspect the incoming latent’s
samplesbatch dimension and confirm it contains at least two elements. - Record the expected identity of element one and element two, using the upstream workflow’s known construction or metadata where available. Do not guess from a port name.
- Connect the combined latent to
LTXVSeparateAVLatent; label or otherwise track its two outputs as expected video and expected audio. - Before connecting downstream nodes, use the documented positional mapping together with known upstream construction records, metadata, or an inspection tool that actually exposes the relevant identity to trace each output to its input element. If the available evidence does not establish that mapping, record the unsupported output identity as
unverified; do not claim to have directly inspected latent contents. Record mask presence and mask correspondence separately.
For the mask behavior, the official node reference says a present noise_mask is split in the same way as samples. It does not define a separate mask output, a mask-viewing procedure, or a particular mask shape, element count, or downstream effect. Therefore record two separate results: the documented rule is that the mask follows the same first-element/video and second-element/audio split; independent observation of that association is not observed unless the workflow tools actually expose it. Do not claim that the mask pairing was independently verified merely because the node ran successfully. As an author-recommended cross-check, compare any exposed mask metadata or tensor slices with the already recorded input element identities; this is a suggested inspection method, not an official software test.
Use a before-and-after contract worksheet
The following is a hypothetical review record, not a report of software testing. Assume an upstream workflow deliberately places video in batch element one and audio in element two, and the combined latent has a noise_mask. Record observations you can actually establish. “Unknown” is a useful result when the available viewer does not expose mask details.
| Check | Before separation | After separation | Acceptance decision |
|---|---|---|---|
| Batch identity | Element 1: intended video; element 2: intended audio | video_latent traces to element 1; audio_latent traces to element 2 | Combine the documented positional mapping with known upstream construction records, metadata, or an inspection tool that actually exposes identity; otherwise mark each unsupported output identity unverified |
| Minimum input size | Two elements confirmed along samples batch dimension | Not applicable | Pass the documented minimum; do not infer behavior for further elements |
| Mask presence | noise_mask present in the combined latent | Mask status inspected for each returned latent | Record present, absent, or unknown; do not claim alignment if it cannot be inspected |
| Mask order | Same element-one/video, element-two/audio convention | Official rule: the mask is split in the same order; independent association: observed or not observed | Mark rule-supported, independently not observed when no tool exposes the association; mark verified only when an actual inspection supports it |
| Routing | One combined LATENT reaches the separator | Two labeled latent outputs reach their intended consumers | Routing confirms connections, not content identity |
If this hypothetical inspection showed the two outputs connected but no evidence of which input element each contained, mark identity as unverified. Do not upgrade that result to a pass because the graph runs.
Interpret limits and choose the next check
If the input has fewer than two batch elements, it does not meet the documented prerequisite. Check the upstream node that constructs the combined latent and establish the intended order before separating again. If the batch has more than two elements, the reference only explains the first two; inspect the upstream workflow and relevant node behavior instead of assuming how extras are treated.
If noise_mask is absent, record it as absent; there is no mask alignment to verify for that input. If it is present but your inspection tools do not reveal its association with each result, report mask pairing as unknown. The documentation states that it is split in the same order, but does not describe a dedicated mask-viewing procedure or guarantee later processing outcomes. Use an inspection method that exposes latent contents where available, or consult the relevant upstream/downstream node documentation. Do not substitute visible output quality for proof of internal element identity.
Common questions
Does the node identify video and audio by looking at their contents?
The documented rule is positional: first batch element for video, second for audio. It does not describe content-based detection.
Does a visible `noise_mask` port have to be connected separately?
The reference describes noise_mask as part of the latent data and says it is split when present. It lists the node input and outputs as LATENT, not a separate mask port.
What if the batch contains a third element?
The documented minimum is two, and the reference assigns the first two roles. It does not say what happens to further elements. Verify behavior from the relevant workflow and node implementation rather than extending the two-element rule by assumption.
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.