JLC LoRA Loader - Multi-Model / Block Weight

Most LoRA loaders default to patching both the diffusion model and the text encoder, and a lot of the time that's unnecessary. If you're on a pipeline where LoRAs are pure style/content (or you've already got your conditioning dialed and don't want a LoRA's text-encoder weights anywhere near it), the extra CLIP handling is just wiring you have to route around. JLC LoRA Loader - Multi-Model / Block Weight is the pack's per-slot block-weight loader with the CLIP side stripped out: up to ten LoRAs, each with its own MODEL strength and its own block vector, and a single MODEL output.

It's the MODEL-only sibling of the CLIP + Block Weight variant, from the same jlc-comfyui-nodes pack by J. L. Córdova (Damkohler). Everything about the dynamic-slot design carries over: ten predeclared slots, slot_count controls visibility and is authoritative, hidden slots are ignored but their values survive in the workflow JSON, active slots apply sequentially.

The MODEL-only difference

The meaningful behavior difference is in the loader core, and it's worth understanding before you trust it: MODEL-only loaders filter out clearly text-encoder/CLIP-side keys from each LoRA file before applying it to the model. So if you load a LoRA that was trained with CLIP weights, this node doesn't touch the text encoder and doesn't make you wade through misleading warnings - it just applies the model side. Ambiguous keys are preserved so genuine MODEL-side issues still surface. That's the pack's explicit design, not a side effect.

Each slot's inputs are exactly three:

• lora_XX - dropdown over your models/loras folder
• strength_model_XX - the LoRA's force on the diffusion model (−10 to 10; 0–1 is the normal band)
• block_vector_XX - that slot's per-LoRA CSV profile, e.g. 1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1

Vector semantics are the pack standard: vector[0] is the base ratio for unmatched MODEL keys, vector[1:] maps across block indices (input_blocks, middle_block, output_blocks, double_blocks, single_blocks), the last value repeats if the vector runs out, zeros suppress patches, nonzeros scale that slot's strength. Since there's no CLIP here, nothing is block-weighted but the MODEL side - which is everything this node touches.

The single model output feeds your KSampler. Wiring is short, and that's the point of this variant: for a stack of pure image-model LoRAs on Flux or SDXL, there's no CLIP to route, no strength to forget to zero, one cable out.

Install and gotchas

ComfyUI Manager (search JLC ComfyUI Nodes) or:

cd ComfyUI/custom_nodes
git clone https://github.com/Damkohler/jlc-comfyui-nodes.git

Restart ComfyUI after install/update. No model downloads, no extra pip dependencies.

Two things to remember. First, default strengths are 0.0 - the node is a pass-through until you actually set values. Second, per-slot vectors are the visually dense end of the block-weight family; if every LoRA in the stack is going to share one profile anyway, the Shared Block Weight MODEL-only variant gives you a single vector for the whole stack and a much calmer node. Pick this one when individual LoRAs need genuinely different treatment - and if you find yourself wanting text-encoder influence too, the MODEL+CLIP version of this node is the same thing with clip re-added.