For a packed LTXV-AV latent, choose video_cfg and audio_cfg by changing one at a time while holding the model, conditioning, and other sampling settings constant. Start from the node’s documented defaults, 3.0 for video and 7.0 for audio, as reference settings rather than proven optimums. Keep the values equal when you intend to use one overall guidance scale; use different values only when you want to control the two modalities separately. First verify that the latent really contains separate video and audio components.
Decide what to change and how
The LTXVDualCFGGuider documentation describes separate guidance for video and audio in a packed LTXV-AV latent. Its inputs accept values from 0.0 to 100.0, but that range is not a recommendation that any particular value will improve a result. Set your decision target before sampling: for example, whether the video follows the described motion, or whether the audio matches the requested sound. Judge those outcomes separately.
Use this controlled comparison:
- Check the incoming latent’s actual type or metadata and confirm it is a packed LTXV-AV latent with both modalities. A node’s color is only a visual clue; use the displayed type label or other available type information.
- Fix the model, positive and negative conditioning, seed if available, and remaining sampler settings. Record the values so that a comparison changes only the guidance under investigation.
- Make a reference run with the two documented defaults, or with the current settings if you already have a baseline. Record visible motion and audio characteristics independently.
- To assess video guidance, change
video_cfgwhile leavingaudio_cfgand all other settings fixed. Compare the video result against the same prompt and baseline. - Restore the baseline, then change
audio_cfgalone. Compare the sound independently; do not infer an audio effect from an overall impression of the clip. - If separate control is the goal, test a pair with different values and record both modalities. If you set equal values, treat that run as the single-scale comparison, not as evidence about independent audio control.
This one-variable-at-a-time sequence is an editorial experiment design, not a requirement imposed by the node. It makes the observations easier to attribute.
Acceptance worksheet
The entries below are a hypothetical worksheet, not measured outputs or promised effects. “Observation” means what the creator should note after reviewing their own samples; it does not predict what a setting will produce.
| Case | Latent type | video_cfg | audio_cfg | Equal? | Video observation | Audio observation | Decision |
|---|---|---|---|---|---|---|---|
| A: reference | Packed LTXV-AV | 3.0 | 7.0 | No | Note whether requested motion is present | Note whether requested sound is present | Keep as baseline only |
| B: video-only change | Packed LTXV-AV | 4.0 | 7.0 | No | Record any difference from A | Record whether sound also changed | Consider the video setting only if its target response is preferable |
| C: audio-only change | Packed LTXV-AV | 3.0 | 6.0 | No | Record whether motion changed | Record any difference from A | Consider the audio setting only if its target response is preferable |
| D: equal values | Packed LTXV-AV | 5.0 | 5.0 | Yes | Review the video result | Review the audio result | Interpret as one overall scale |
| E: boundary check | Non-packed or non-AV latent | 4.0 | 7.0 | No | Record the applicable output | Do not use this run to judge independent audio CFG | Confirm the input type before drawing an AV conclusion |
Use your own prompts, outputs, and acceptance criteria in the observation cells. For example, a creator might mark whether a described camera movement is recognizable and whether a requested ambience is present, without scoring either as objectively better. Do not treat these hypothetical values as an optimum, or assume the table establishes how much any output should change.
Limits and next steps
The documentation says that equal or very close guidance values are handled as one scale for the whole latent. It does not specify a numerical threshold for “very close,” so avoid relying on a borderline pair when the distinction matters. Choose clearly equal values for a single-scale run, or values that are plainly different for a separate-control comparison.
The node’s separate-modality behavior applies to a packed LTXV-AV latent. For an input that is not such a latent, the documented boundary is that only video_cfg is used. Do not interpret that case as a failed audio adjustment: first identify the actual latent type. If the type is uncertain, resolve that uncertainty before comparing CFG settings. If a controlled comparison shows no useful change, record that result and retain the setting that best meets your own criteria; the documentation does not promise that changing guidance will improve conditioning adherence.
The node creates a guider output for a sampler. Its documentation does not establish a universally best pair, guarantee a particular visual or sonic response, or describe a scoring method for judging outputs. Those remain choices for the creator and the specific generation task.
Common questions
Should I begin with the documented defaults?
They are a practical reference point because the node lists them, not an assurance that they suit every prompt or model.
Does a higher value necessarily make one modality better?
The documentation defines guidance strength and the modality it applies to, but does not promise a monotonic improvement. Compare the actual video or audio response against a stated criterion.
What if I want one shared CFG value?
Set the two inputs to the same value and interpret the run as a single overall scale. For an input without separate video and audio components, the documented behavior is to use video_cfg only. See the ComfyUI node documentation for the input and latent-type boundary.
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.