Start with start_percent=0 and end_percent=1 as the full-range baseline. The ComfyUI LTXV Reference Audio node applies identity guidance only when its scale is above zero and the current sampling step falls within the selected sigma range. A narrower range is an experiment in when that guidance acts, not a substitute for changing its strength. The documentation does not identify one range as universally best.
Set up a controlled range comparison
Use this sequence when you have a reference clip and want to know whether limiting the guidance interval is useful for your specific generation:
- In ComfyUI, locate
LTXVReferenceAudioand identify its inputs by their displayed type labels. Node color can be a visual clue, but use the actual labels to confirm connections. The node encodes the reference audio into conditioning and patches the model for identity guidance. - Choose one model, one reference clip, one
identity_guidance_scalegreater than zero, and one fixed set of other sampling conditions. The node documentation recommends a reference clip of about five seconds; it also warns that shorter or longer clips may degrade identity transfer. This is guidance, not a guarantee for an individual clip. - Generate a baseline with
start_percent=0andend_percent=1. Record the range and scale used. Keep the output and its prompt available for comparison. - Make further runs that change only
start_percentandend_percent. For example, compare a proposed early interval and a proposed late interval against the full range. Treat these as experimental ranges you select, not documented presets. - Listen to each result using the same target passage. Record identity cues and audible content separately: a voice may sound more or less like the reference while words, timing, or other audible details also change.
The node exposes the range controls only when advanced options are enabled. If you cannot see them, check that interface setting before treating the controls as unavailable. For additional node behavior, see the LTXV Reference Audio documentation.
Use a range acceptance worksheet
The entries below are a hypothetical evaluation, not generated results or a prediction of what any range will do. The example assumes the creator chose these intervals and scale for comparison. Mark observations from your own outputs; do not treat a subjective identity judgment as a measured score.
| Run | Range and fixed scale | Hypothetical identity observation | Hypothetical content observation | Decision to test next |
|---|---|---|---|---|
| A | 0–1, scale 3.0 | Reference-like vocal quality is noticeable | Target words are intelligible; a pause differs from the intended pacing | Keep as the comparison baseline |
| B | 0–0.5, scale 3.0 | Identity impression is less consistent across the passage | Words remain intelligible; the pause sounds closer to the intended pacing | Repeat this range with another approved passage before preferring it |
| C | 0.5–1, scale 3.0 | Identity impression seems closer on the final phrase | Final phrase timing differs; no claim about the rest of the voice follows from this alone | Inspect the full passage and compare content separately |
This worksheet separates two questions: whether identity seems preserved and whether the generated audio still serves the intended content and delivery. If a run is hard to judge, note the uncertain phrase and repeat the comparison with the same reference and target passage rather than changing several inputs at once.
Interpret results and choose a next step
A result from one voice, prompt, model, or sampling setup does not establish a generally optimal sigma interval. The documentation defines the range as the period in which identity guidance is active; it does not promise that an early or late interval improves identity, wording, pacing, or sound quality. Treat your comparison as local evidence for the setup you held constant.
If none of the ranges gives an acceptable result, first confirm that the scale is above zero: at zero, identity guidance is disabled. Then verify the range values and ensure the runs differed only in those values. Check whether the reference is around the recommended five-second duration, remembering that the documentation warns of possible degradation for shorter or longer clips. The node resamples the reference to the audio VAE sample rate when the rates differ, so a rate mismatch alone is not a reason to infer that you need a different sigma interval. If results remain unsuitable, report the observed identity and content issues separately; the documentation does not specify a guaranteed corrective interval.
Common follow-up questions
Does `0–1` mean the reference is applied to every conditioning step?
It defines the full documented sigma range for identity guidance. Guidance still requires a scale above zero and a sampling step within that range. The reference is also encoded into conditioning by the node.
Should I change the scale and range together?
Not when diagnosing which interval matters. Hold the scale fixed while comparing ranges; test a different scale in a separate comparison so strength and active period do not become confounded.
Does a better identity impression mean the generated audio is better?
Not by itself. Judge identity cues separately from words, timing, and other content relevant to the target. The worksheet is a proposed review method, not an official metric or a claim that ComfyUI supplies an identity score.
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.