Filtered Random LoRA Loader (LBW)

This is the advanced member of the RandomLoRALoader family, and the reason it exists is a problem anyone who stacks LoRAs has felt: a style LoRA doesn't just change the style. It can drag composition, structure, and character features along with it, because a normal LoRA patch hits the whole U-Net at once. The LBW variant adds LoRA Block Weight - control over which U-Net blocks the LoRA is allowed to affect - on top of the keyword-filtered random pick.

In practice: an "anime watercolor" style LoRA applied with Style Focused preset touches the OUTPUT blocks only, so you get the brushwork without the composition getting hijacked. A pose LoRA with Structure/Composition Only keeps the layout and leaves the style alone. You can even chain two instances - pose first, then style - and get a result that keeps each effect in its lane. When most of the conversation around LoRAs is "mixing them is unpredictable," this is a genuine attempt at dialing down the chaos. (LBW itself traces back to the sd-webui-lora-block-weight theory; the automatic SD1.5/SDXL detection is this pack's nice touch.)

How it works

Same engine as the other two nodes: one folder, keyword filter, random pick, trigger words pulled from metadata, strengths applied to MODEL/CLIP. The LBW part adds a per-LoRA weight list - a comma-separated string of numbers, one per U-Net block, 20 elements for SDXL (BASE + IN:9 + MID:1 + OUT:9) and 17 for SD1.5. The node detects which architecture each LoRA belongs to by inspecting the keys inside the file, so you don't pick a model type yourself. Mismatched weight counts get auto-fixed: too few pads with 1.0s, too many truncates.

You don't usually type those numbers, though. The weight_mode enum gives you:

• Normal (All 1.0) - plain LoRA application, LBW off. This is your default.
• Style Focused - OUTPUT blocks only. Style without touching structure.
• Character Focused - balanced IN+MID+OUT. Character features stay intact.
• Structure/Composition Only - INPUT+MID only. Layout without the style bleed.
• Balanced / Soft - gentle application, for general-purpose stacking.
• Preset: Random - picks one of the four at random each run.
• Direct Input - your own lbw_input weights (SDXL: 20 numbers, SD1.5: 17).

One LBW-specific behavior worth knowing: the positive_text output emits the block weights in <lora:name:model:clip:lbw=1,0,...> syntax, which is what makes it play nicely with Wildcard Encode if you're building a fully dynamic workflow. And since this inherits the filtered node wholesale, keyword_filter, filter_mode, search_in_metadata, and the whole metadata-trigger-word machinery are all here.

The inputs that actually matter

• lora_folder_path + keyword_filter - where to look and what counts as a match.
• weight_mode - the block-weight preset (or Direct Input). The genuinely new dial.
• lbw_input - only for Direct Input mode.
• num_loras, model_strength / clip_strength (fixed or "0.6-0.9" ranges), seed - standard, same as the pack's other nodes.

Outputs: MODEL, CLIP, positive/negative CONDITIONING, positive_text/negative_text strings, preview IMAGE. All the usual.

Install

ComfyUI Manager (search "RandomLoRALoader") or:

cd ComfyUI/custom_nodes
git clone https://github.com/shin131002/RandomLoRALoader

Restart ComfyUI. No heavy dependencies beyond ComfyUI's bundled libraries; optional pip install opencv-python only affects video previews. Important caveat from the repo: SD1.5 and SDXL only. LBW targets the SD1.5/SDXL U-Net block structure specifically - no Flux, SD3, or Pony, and the block weights are meaningless on architectures with a different U-Net layout.

Troubleshooting

• No LoRA found - filter too narrow, path wrong, or the tags are in metadata while search_in_metadata is off. Same checklist as the filtered node.
• "Weight count mismatch" warnings - expected when a preset designed for one architecture hits the other. The node auto-adjusts; the warning is informational, not fatal.
• Nothing seems to change - if weight_mode is a preset where the active blocks are weighted 0 for the effect you wanted, or strengths are 0, the LoRA silently does nothing. "Style Focused" won't give you composition changes, by design.
• The trap to avoid: don't reach for LBW on day one. "Normal (All 1.0)" is what 90% of LoRA work wants, and the whole point is not applying it everywhere. Use presets to diagnose which effect a LoRA is actually contributing before you build a three-node chain around it.
• No support promised - small MIT pack, explicit as-is/no-support policy, and this is its newest node (v1.2.0). It's a single readable Python file if you ever need to trace what a preset does.