ID-LoRA Model Loader

Let's get the elephant out of the room first, because the pack's own README leads with it: this node is deprecated. ID-LoRA - the technique that takes a photo of a face plus a short audio clip of a voice and generates a talking-head video where that face speaks in that voice - now ships natively in ComfyUI core as the LTXVReferenceAudio node (PR #13111, courtesy of Kijai). The authors' own advice is "update ComfyUI and don't use these nodes."

So why read on? Because you grabbed a workflow that references IDLoraModelLoader and it's sitting there red in your graph. Fair enough. Here's what it actually does, and why it's so heavy.

What it is

The one-stage ID-LoRA pipeline starts here. This node points the pipeline at three big pieces of the LTX-2.3 stack and hands you one reusable object:

• checkpoint_path - the LTX-2.3 22B dev checkpoint (~44GB) from ComfyUI/models/checkpoints/
• text_encoder_path - the Gemma 3 12B text encoder directory. Leave it empty; the node auto-scans text_encoders/ for a folder named gemma
• lora_path + lora_strength - the actual ID-LoRA weights (id-lora-celebvhq-ltx2.3 or id-lora-talkvid-ltx2.3, ~1.1GB each). none skips the LoRA entirely; strength runs 0–2, default 1.0

The single output, ID_LORA_PIPELINE, feeds straight into the ID-LoRA Prompt Encoder. That's the whole job: this node is a pipeline factory, not a generator. Nothing gets generated until a sampler runs.

The inputs that actually matter

• quantize (none / int8 / fp8) - your VRAM valve. int8 is the sweet spot for fitting a 24GB card with a modest quality hit; fp8 saves a little more but adds artifacts. At 48GB+ just leave it none.
• stg_scale, identity_guidance_scale, av_bimodal_scale - spatio-temporal guidance, speaker-identity transfer strength, and audio-video bimodal CFG. Defaults of 1.0 / 3.0 / 3.0 are sane starting points; you tune these for likeness-vs-naturalness, not out of the box.
• lora_strength - the general LTX advice about not overcooking LoRAs applies here too. Push it too hard and you trade natural audio for speaker likeness.

One mechanism detail worth knowing, because it explains a lot of the pack's behavior: despite the name, the loader does not load the 22B transformer when it runs. It builds the pipeline and defers load_models() to the sampler - after the prompt encoder has done its job and freed the text encoder. That ordering is the only reason this thing survives on a 24GB card. ComfyUI caches this node's output aggressively: the first queue takes minutes while models load, and later runs reuse the pipeline for free. Don't touch it between generations.

Installing it (the real way)

This is not a one-command pack, and ComfyUI Manager will only get you partway (search "ID-LoRA-LTX2.3-ComfyUI" and it clones the repo, but the models and the version pins are all manual). The README documents a multi-repo install:

git clone https://github.com/ID-LoRA/ID-LoRA.git
bash ID-LoRA/ID-LoRA-2.3/scripts/download_models.sh models/   # ~67GB for one-stage
pip install 'transformers>=4.52,<5'          # pin BEFORE installing anything else
pip install -e ID-LoRA/ID-LoRA-2.3/packages/ltx-core
pip install -e ID-LoRA/ID-LoRA-2.3/packages/ltx-pipelines
pip install -e ID-LoRA/ID-LoRA-2.3/packages/ltx-trainer
cd ComfyUI/custom_nodes && git clone https://github.com/ID-LoRA/ID-LoRA-LTX2.3-ComfyUI

Then symlink the downloaded weights into ComfyUI's models/ folders (checkpoints, loras, text_encoders) so the dropdowns can see them. The Gemma encoder needs a HuggingFace login and its license accepted before the download script will fetch it.

Where people get burned

• transformers 5.x - the number one dependency conflict. Recent ComfyUI installs pull transformers 5, which ID-LoRA breaks on. Pin to 4.x before installing the ltx packages, and be ready for it to fight your other custom nodes.
• OOM - the 22B transformer is greedy. Enable int8, drop max_resolution to 512, cut num_frames to 121. The README's own math says int8 + 512px + one-stage lands on a 24GB card.
• Deprecation drift - the native LTXVReferenceAudio path is the recommended route now, so this pack is more likely to rot than get fixed as core moves on.

If you're starting fresh and no workflow is forcing you into these nodes, save yourself the install and update ComfyUI instead.