Load LoRA Model Index (NH)

Load LoRA Model Index (NH) (class NH_LoraModelIndex) is the model-only half of the NH LoRA-index pair. Where NH_LoraClipIndex patches both the diffusion model and the CLIP, this one patches just the MODEL with one LoRA from a menu of up to 64 - for when you're applying a LoRA that doesn't need its CLIP half touched, or you're deliberately separating the two. Same switching-loader idea, one less input to feed.

How it works

List your LoRAs in the lora_1 … lora_64 dropdowns (lora_count decides how many slots are active), then choose via select_mode:

• index - index (1-based) picks the slot to apply.
• boolean - enable_N toggles; the first enabled slot wins, and if none are enabled it returns a blocked (LoRA-free) model so the workflow doesn't crash.

strength_model is the blend strength. Outputs are model, filename, and index. Note the output list: no clip, because this node only touches the model - if you need both halves patched together, use NH_LoraClipIndex. The dropdowns read straight from models/loras, same as the core loader.

Where you'd actually use it

When a LoRA was trained model-side only, or when you're doing a Model-only LoRA stack and want the CLIP handled elsewhere - a common pattern with certain style/architecture LoRAs that shouldn't affect text encoding. It's also the right pick when your downstream only needs a MODEL input and feeding a patched CLIP would add confusion. Like its sibling, it shines for A/B runs: one graph, drive index from a slider or seed, and flip between LoRAs per run.

Installing

ComfyUI Manager → search NH-Nodes → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/jetthuangai/NH-Nodes.git
cd NH-Nodes
pip install -r requirements.txt

Restart and refresh the browser page. It reads whatever's in models/loras - nothing downloads.

Gotchas

The 1-based index trips people up again - slot lora_1 is index 1, not 0. New LoRAs in models/loras need a browser refresh to show up in the dropdowns. And boolean mode applies only the first enabled slot; no stacking. One genuine limitation versus the core loader: there's a single strength_model shared across all slots, so you can't store per-LoRA strengths in one node - if your LoRAs need different strengths, that's still separate loaders or a slider feeding the strength input.