ID-LoRA One-Stage Sampler

Everything upstream just prepares. This is the node that does the work: it runs the one-stage denoising loop with all the guidance turned on, decodes the video, decodes the audio through the vocoder, and hands you a ComfyUI VIDEO ready for the built-in Save Video node. The name is accurate - one resolution, one pass, no upscaler involved.

The two inputs that make ID-LoRA what it is

Both are optional, and both are easy to miss because they live in the optional section rather than the required one:

• first_frame (IMAGE) - a photo of the face you want to animate. The node encodes it, center-crops it to your output size, and injects it into the first frames of the video latent as clean conditioning that the denoiser keeps fixed.
• reference_audio (AUDIO) - the voice sample. It's encoded and prepended to the audio latent as a reference prefix. Leave it disconnected and you get whatever voice the prompt implies - a generic speaker, no identity transfer. The entire "this face sounds like this person" trick needs this wire.

The generation controls

• num_frames (121 default) and num_inference_steps (30 default) - clip length and denoising budget. 121 is the standard LTX frame count. Drop steps to iterate faster, but this pipeline runs the dev checkpoint, which wants its steps; the distilled-model 8-step shortcut doesn't apply here.
• auto_resolution (on) + max_resolution (512) - reads the aspect ratio off your first frame and sizes the output, capping the long side. Turn auto_resolution off and height/width take over (keep them multiples of 32).
• video_guidance_scale (3.0) and audio_guidance_scale (7.0) - separate CFG for the two modalities. The audio knob runs hotter by default, which matches where the identity signal lives.
• seed - your iteration currency. LTX's prompt adherence is still its weakest axis, so seed-hunting cheap generations beats a prompt-engineering marathon.

How the identity transfer actually works

With the reference audio connected, each denoising step runs the transformer twice on the audio path - once with the reference prefix, once without - and the difference, scaled by identity_guidance_scale from the loader, is pushed onto the audio gradient. That's the mechanism behind "sounds like the reference," and it's why that knob lives on the Model Loader instead of here. STG and the audio-video bimodal CFG from the loader stack on top of the plain video and audio CFG.

Output

VIDEO - frames plus embedded audio at the vocoder's sample rate - which you wire into Save Video. Don't forget that last node; a sampler output with nothing connected after it just runs and vanishes.

Installing and troubleshooting

Shared pack install - see the ID-LoRA Model Loader article. This is the node where VRAM dies: the heavy transformer loads here (the loader's deferred load_models() finally fires), so the first run after any loader change takes minutes just to load. If it OOMs, the fixes live upstream on the loader - int8, lower max_resolution, fewer frames - not here. And the usual pack-wide reminder: deprecated, native LTXVReferenceAudio is the supported path, this is for old workflows.