HyperLoRA Uni Loader

The HyperLoRA pack has a two-node dance - HyperLoRAConfig builds a config, HyperLoRALoader consumes it - and that's genuinely annoying when the defaults are right anyway. HyperLoRAUniLoader is the answer: it's both nodes fused into one, with the recommended settings baked in. You pick three things from dropdowns and get the same HYPER_LORA bundle the full setup produces.

What it is

Under the hood it literally calls HyperLoRAConfigNode.execute() and then HyperLoRALoaderNode.execute() - the source just wires them together and hardcodes the config values the released models expect: resampler.dim 1024, num_id_tokens 16, lora_rank 8, has_base_lora false. Same machinery, fewer ways to break it. That makes it the right first choice for anyone who just wants a working face-LoRA pipeline.

The inputs

• image_processor / image_encoder - the CLIP ViT-L/14 processor and model you downloaded, picked from folders under models/hyper_lora/. "Not found!" = model folders not set up.
• encoder_types - clip, arcface, or clip + arcface. Default to clip + arcface; that's what the released models were trained with and what the standalone script uses.
• face_analyzer - antelopev2 from models/insightface/models/, needed whenever arcface is in the mix.
• model - which HyperLoRA variant: sdxl_hyper_id_lora_v1_fidelity (better likeness) or sdxl_hyper_id_lora_v1_edit (better editability).
• dtype - fp16 is the right default; bf16/fp32 if you have a reason.

Output is the single HYPER_LORA bundle that feeds HyperLoRAUniGenerateIDLoRA directly, or the individual FaceAttr/IDCond/Generate nodes.

Installing

ComfyUI Manager → search "ComfyUI-HyperLoRA" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/bytedance/ComfyUI-HyperLoRA
pip install -r requirements.txt

then restart ComfyUI. The requirements (diffusers, insightface, onnxruntime, opencv-python, transformers) are the heavy part - insightface in particular is a historically painful install. Then the model files:

mkdir -p models/hyper_lora/clip_processor/clip_vit_large_14_processor
mkdir -p models/hyper_lora/clip_vit/clip_vit_large_14
mkdir -p models/hyper_lora/hyper_lora
mkdir -p models/insightface/models

CLIP processor + ViT from openai/clip-vit-large-patch14, the HyperLoRA weights from bytedance-research/HyperLoRA (unzipped under models/hyper_lora/hyper_lora/), and antelopev2 from the InsightFace model zoo (manual Google Drive download, unzip under models/insightface/models/).

Gotchas

The same "Not found!" trap as the rest of the pack - if any dropdown is empty, your folder layout is off; fix the paths before touching settings. And remember the surrounding rules of the workflow: SDXL checkpoint only, RealVisXL v4.0 recommended, the prompt has to start with fcsks fxhks fhyks, , and stop_at_clip_layer set to -2 (which needs the ComfyUI_ADV_CLIP_emb nodes if you're following the example workflows). If you see metadataIncompleteBuffer on load, a downloaded model file is corrupt - re-fetch it rather than debugging settings.

The one tradeoff: the Uni loader fixes has_base_lora to false, so if you're playing with a variant that ships a separate base resampler, you'll need the full Config + Loader pair after all.