For LTXV Reference Audio, prepare a clean speaker sample of about 5 seconds: the ComfyUI documentation recommends roughly that duration and warns that shorter or longer clips may weaken identity transfer. The node encodes the reference into conditioning; identity guidance is a separate, optional model path. Start with identity_guidance_scale at 0 when you want to avoid its extra forward pass, then consider enabling it if voice identity needs more emphasis. The reference is an input, not the final generated audio.
Choose the reference and guidance path
- Select a clip whose speaker identity is the intended reference. Record its duration and what it contains, such as one speaker speaking clearly or speech mixed with music. The documentation recommends about 5 seconds; it does not specify a required sentence or guarantee a result.
- Connect the clip to
reference_audioand the LTXV Audio VAE toaudio_vae. Check the input type labels in ComfyUI rather than relying on node colors. If the audio and VAE sample rates differ, the node resamples the reference to the VAE’s rate. - Make a baseline with
identity_guidance_scaleset to0. Reference audio is still encoded into positive and negative conditioning; zero disables the extra guidance pass, not the reference input itself. - If the baseline does not meet your identity goal, make a separate comparison with guidance enabled. The documented default scale is
3.0, but treat it as a setting to evaluate, not a promised improvement. Guidance adds an extra forward pass without the reference at each applicable sampling step, so it also changes the computation path. - Record
start_percentandend_percentwith each version. They define the sampling range in which guidance is active, and the controls may only appear when advanced options are enabled. This guide does not prescribe interval experiments.
Keep reference and generated media labeled separately in your project notes. Listen to the generated output as its own deliverable; do not count the uploaded clip as proof that the requested voice was generated or preserved.
Use a two-version acceptance card
The following is a hypothetical comparison worksheet, not a software test or expected outcome. Its observations illustrate how to make a decision without assuming that guidance will improve identity.
| Version | Reference record | Guidance record | Hypothetical listening observation | Decision |
|---|---|---|---|---|
| A | 5.0 seconds; one speaker, clear speech | Scale 0; range recorded as 0.0–1.0 | Output is understandable, but the reviewer says the speaker sounds less similar than intended | Keep as the no-extra-pass baseline |
| B | Same 5.0-second reference | Scale 3.0; range recorded as 0.0–1.0 | Reviewer hears a closer resemblance, but the result is not independently verified | Consider B for this brief; do not claim guaranteed identity transfer |
For a real review, include the VAE sample rate, the reference’s audible contents, the generated file being judged, and the reviewer’s specific observation. Compare versions against the same approved target and listening conditions; this is an editorial control for the comparison, not a behavior guaranteed by the node documentation.
Limits and next steps
The documentation describes voice-identity transfer, but gives no identity score, acceptance threshold, or guarantee that a particular clip or scale will produce a desired likeness. It also does not define a best duration beyond recommending approximately 5 seconds. Avoid inferring that extra detail, a particular phrase, or a longer clip will necessarily help.
If the zero-scale version is not suitable, first review the reference recording’s duration and contents and confirm the audio_vae connection. Then compare a guidance-enabled version while keeping the reference and other generation inputs unchanged. If neither version meets the brief, document the mismatch and revise the reference or production requirements; do not present the source clip as the generated result. The documentation says the extra pass applies only when the scale is greater than zero and a sampling step falls within the recorded range. It does not promise that enabling guidance will fix an unsuitable reference.
Common questions
Does a scale of zero ignore the reference audio? No. According to the node documentation, the audio is still encoded into conditioning; zero disables the additional identity-guidance pass.
Is the reference clip the audio output? No. It is the input used for identity conditioning. Assess the generated audio separately.
Should I tune `start_percent` and `end_percent` here? Record their values because they determine when enabled guidance is active. Detailed interval tuning is outside this guide’s scope.
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.