ID-LoRA Two-Stage Sampler

This is the flagship sampler of the pack - the one in the demo videos, the 1024x1024 output. It runs the two-stage pipeline that the Two-Stage Model Loader set up, and it's the reason that loader exists at all.

How the two stages go

Stage 1 generates video and audio together at your target resolution with full guidance - CFG, STG, identity transfer, A/V bimodal CFG all on. Then the stage-1 models are freed, the video latent is upsampled 2x, and stage 2 refines at double resolution with the distilled LoRA alone: no guidance, three fixed steps, audio frozen from stage 1. It's the same pattern as a distilled checkpoint wanting CFG near one and eight steps - the refinement pass is cheap, unguided polish on top of the expensive stage-1 base.

The practical consequence is the one number you'll keep tripping on: whatever height and width you set are stage-1 numbers. Output is 2x them. 512x512 in, 1024x1024 out. With auto_resolution on, max_resolution caps stage 1's long side and the final output doubles it - 768 becomes 1536, if your VRAM and patience survive it.

Inputs

Same shape as the one-stage sampler: pipeline, conditioning, seed, num_frames, num_inference_steps (30 for stage 1; stage 2's three steps are fixed and not exposed), frame_rate, video_guidance_scale (3.0), audio_guidance_scale (7.0), auto_resolution, and the optional first_frame and reference_audio. First frame and reference audio behave exactly as they do in the one-stage node - same identity-transfer mechanism, same generic-voice risk if you forget the audio wire.

The one new knob:

• hq_mode (default on) - swaps the Euler stepper for the res2s second-order sampler in both stages. Higher quality, slower, and noticeably heavier on VRAM because of the midpoint evaluation. If you're fighting OOM, the README lists this as one of the first things to sacrifice.

Where it hurts

VRAM and time, in that order. Two transformers load sequentially, the latent gets upsampled to 2x, and res2s adds a second-order step on top. Non-quantized at high resolution the README asks for 48GB and suggests 80GB. The escape hatches, in rough order of pain: int8 on the loader, lower max_resolution, fewer frames, then hq_mode off. The one-stage sampler is the sane place to iterate; come back here only when you want the final render.

Installing and troubleshooting

Shared pack install - see the ID-LoRA Model Loader article, and the Two-Stage Model Loader article for the two extra files (upsampler + distilled LoRA, ~8GB). The usual reminder applies to the whole pack: deprecated, native LTXVReferenceAudio in current ComfyUI is the supported route, and these nodes exist for reproducibility and old workflows.