ID-LoRA Two-Stage Model Loader

This is the sibling of the ID-LoRA Model Loader with two extra boxes on the form, and those two boxes are the whole point. Where the one-stage loader builds a single-pass pipeline, this one loads the machinery for the two-stage pipeline that made the pack's demo videos look good: generate at your target resolution with full guidance, then upsample 2x in latent space and refine with a distilled LoRA. 512x512 becomes 1024x1024. 768 becomes 1536.

The two extra inputs

• upsampler_path - the LTX-2.3 spatial upsampler (ltx-2.3-spatial-upscaler-x2, ~1GB) from ComfyUI/models/upscale_models/
• distilled_lora_path - the ~7.1GB distilled LoRA (ltx-2.3-22b-distilled-lora-384.safetensors) that stage 2 runs on

Everything else mirrors the one-stage loader: checkpoint_path, the auto-detecting text_encoder_path, the ID-LoRA lora_path + lora_strength, quantize, and the three guidance scales (stg_scale, identity_guidance_scale, av_bimodal_scale). Output is the same ID_LORA_PIPELINE type, so the Prompt Encoder downstream doesn't care which loader fed it.

How the two stages split the work

Stage 1 runs the full guided pipeline - CFG, STG, identity guidance, A/V bimodal CFG - at your target resolution. Then the stage-1 models are freed, the video latent is upsampled 2x, and stage 2 refines it with only the distilled LoRA and no guidance, in three fixed steps, with the audio frozen from stage 1.

That's the standard distilled-model pattern, and it's worth internalizing: distilled weights get their guidance baked in during training, so you run them at very few steps with CFG near one instead of at seven. Same reason a distilled checkpoint wants 8 steps where the dev model wants 30. The stage-2 refinement here is exactly that - fast, cheap, unguided polish on top of the expensive stage-1 base.

The cost

Like its sibling, this node defers actual model loading to the sampler, so all you get back is a pipeline object - but a pipeline object that claims roughly 75GB of disk (the two-stage download adds the upsampler and distilled LoRA to the ~67GB one-stage set) and a lot of VRAM later. Non-quantized, the README recommends 48GB, and 80GB for high-res two-stage output. int8 on quantize plus a modest max_resolution gets you closer to 24–32GB, at the cost of some quality on the very pass you added the extra models to get.

Installing it

Shared pack install - see the ID-LoRA Model Loader article for the full walkthrough. The differences here: the download script fetches the upsampler and distilled LoRA too (add ~8GB to your disk budget), and you symlink them into upscale_models/ and loras/ respectively. Same transformers<5 pin, same ltx packages from the ID-LoRA repo.

Should you use it?

Only if you actually want the upscale. The two-stage loader is the heaviest node in the pack, and for iterating on prompts and seeds the one-stage loader is a much cheaper place to start - same technique, lower resolution, minutes instead of an eternity. And the standing reminder applies: the whole pack is deprecated, with native LTXVReferenceAudio in current ComfyUI the recommended path. This node exists for reproducibility and for old workflows, and that's about it.