LoRA Block Weight Group (SD3.5 Large)

Here's the finding that makes LoRA Block Weight Group (SD3.5 Large) worth having: SD3.5's 38 joint blocks carry a LoRA's signal so redundantly that knocking out any single block does almost nothing. The author's own demo measured a 47× spread between the highest- and lowest-MSE blocks - the most extreme of the three models this pack supports - and even zeroing the top-12 blocks together only softened the style, it didn't remove it. Single-block sweeps on SD3.5 get you noise. Group sweeps get you signal.

This node is the SD3.5 member of the pack's Group family. It takes the same sampler inputs as the Batch node (model, vae, positive, negative, latent_image, seed, steps, cfg, sampler, scheduler, denoise) and swaps the flat block list for two text fields: groups and modes.

What the fields do

• groups - one group per line. A range is J10-J19; commas combine ranges into one group (J00-J05,J20-J25). Ranges can't cross prefixes and unknown tags error loudly. The default is a naive even split (J00-J09, J10-J19, J20-J29, J30-J37, plus both halves) - the tooltip says it plainly: a starting point with no prior knowledge, edit it once your first sweep shows where the effect concentrates.
• modes - comma list of column types, each becoming one column of the output: knockout (group off, others on), solo (group on, others off), full (entire LoRA on - the reference), off (entire LoRA off - the no-LoRA baseline).

Outputs are images (batched IMAGE), info, and groups_used / modes_used - feed those strings into Save Grid so the labeled PNG always matches what actually ran.

When to reach for it

After a first-round scan. A full 38-block sweep is 190 images at 1024² - the USAGE.md notes that's roughly 50 minutes on an H-class GPU - so the sane path is a sparse first pass, then use this node to zoom into the neighbourhood that showed life. Because of the redundancy, the meaningful question on SD3.5 isn't "which single block matters" but "which range, knocked out together, changes the picture." That's literally what this node answers, one group at a time.

It's also the honest instrument for the classic stacking problem: if an SD3.5 LoRA fights another one you've piled on, group-knockout tells you which contiguous region is causing the interference, and you can zero just that band in the Custom node.

Install

Via ComfyUI Manager, search "LoraBlockWeight". Or:

cd <ComfyUI>/custom_nodes
git clone https://github.com/Baldwinzc/ComfyUI-LoraBlockWeight.git

Restart ComfyUI. No extra dependencies, no model downloads, no API key - just numpy, Pillow, torch, which ComfyUI ships.

Gotchas

The redundancy cuts both ways: SD3.5's robustness means a group knockout verdict is more prompt-dependent than on FLUX, because the same LoRA can distribute differently across prompts. Re-verify a group finding on a second prompt before you bake it into a permanent recipe. And remember conditioning is encoded upstream here, so the group sweep measures transformer-only contribution - the CLIP side of your LoRA isn't part of the measurement.