The intended video duration is set by frames_number ÷ frame_rate; LTXV Empty Latent Audio uses both inputs to derive the number of audio latents it creates. Connect an Audio VAE for the latent dimensions, and treat the resulting length as a generation-time budget, not as a promise that later audio and video outputs will align perfectly. The official node documentation does not specify the exact conversion formula from video frames to audio-latent count, so do not infer that count from duration alone.
Choose inputs from the target duration
Start with the visual shot you intend to generate, not the node defaults. The video duration implied by the inputs is a planning calculation: for example, 97 frames at 25 fps correspond to 3.88 seconds. That arithmetic describes the frame-rate budget; it does not establish the generated audio’s exact playable duration.
- Decide the intended video frame count and frame rate. Use values for the same shot, not a frame count from one plan and a rate from another.
- Calculate the implied video duration as
frames_number ÷ frame_rate. If it differs from the brief, revise one of those inputs before generating. - Connect the intended Audio VAE. The node requires it; its configuration supplies latent channel and frequency dimensions.
- Set
batch_sizeto the number of audio latent samples you want in the batch. It controls the batch dimension, not the duration calculation. - Inspect the output’s type label and shape where your workflow exposes them. Compare planned duration and the audio/video outputs after generation; do not assume matching input timing guarantees matching final media duration.
The documentation lists frames_number and frame_rate as required inputs and says their values determine num_audio_latents. It describes the output shape as (batch_size, z_channels, num_audio_latents, audio_freq). The channel and frequency dimensions come from the Audio VAE, while the frame-count and rate inputs determine the audio-latent count. See the LTXV Empty Latent Audio node reference.
Use a duration-budget worksheet
The rows below are hypothetical planning cases, not measured node outputs. They hold frame count constant to show why frame rate matters to the intended video duration. The “latent count” column is deliberately left as a check against the actual node output: the documentation gives the dependency but not a count formula.
| Hypothetical inputs | Implied video duration | Batch and VAE plan | Check and decision |
|---|---|---|---|
| 100 frames, 25 fps | 4.00 s | Batch 1; connect the Audio VAE selected for this workflow | Record the node’s reported num_audio_latents; do not assume its value from 4 seconds |
| 100 frames, 30 fps | About 3.33 s | Batch 1; keep the same VAE to isolate the timing-input change | The planned video is shorter than at 25 fps; revise frame count if the brief still calls for 4 seconds |
| 100 frames, 25 fps | 4.00 s | Batch 3; same VAE | Expect three batch entries, not three times the duration per entry; verify the exposed shape |
| 100 frames, 25 fps | 4.00 s | Batch 1; a different compatible Audio VAE | Check its configured channel and frequency dimensions; do not attribute those shape changes to frame rate |
For a four-second target, 25 fps implies 100 frames, while 30 fps implies 120 frames. This is a planning conversion, not a claim about a particular node output. The worksheet separates timing choices from batch and model configuration so a mismatch has a useful first place to investigate.
Limits and what to do when lengths differ
The node reference identifies the inputs and output dimensions, but does not document the precise rounding, temporal compression, or frame-to-audio-latent conversion rules. It also does not guarantee that generated audio and video will have identical playable durations. Treat an observed latent count as configuration-specific; do not turn one workflow’s output into a universal formula.
If the planned durations disagree, first recalculate the video budget and confirm that the node received the intended frame count and rate. Then verify that an Audio VAE is connected and inspect the latent type and dimensions. A node’s color can be a visual clue, but use the displayed input and output type labels to confirm connections. If the latent shape looks plausible but final media durations differ, compare the generated outputs in the downstream workflow. Audio/video concatenation may crop or pad to fit, but that is a later operation, not a correction made by this empty-latent node. The separate LTXV AV concat node reference describes that later length-adaptation stage.
Common questions
Does `batch_size` make each audio sample longer? No. The documented shape assigns it to the batch dimension; num_audio_latents is derived from frame count and frame rate.
Does changing 25 fps to 30 fps always change the latent count by a predictable amount? The node documentation says both inputs determine that count, but gives no exact conversion or rounding rule. Check the output for the specific configuration.
Can I leave Audio VAE disconnected while planning? No. The input is required, and the documentation says the node raises an error if it is absent.
Sources and next step
- LTXVEmptyLatentAudio - ComfyUI Built-in Node Documentation
- LTXVConcatAVLatent - ComfyUI Built-in Node Documentation
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.