LoRA Block Weight Group (FLUX.1)

A full per-block sweep of a FLUX.1 LoRA is 57 blocks × 5 strengths = 285 images. That's a lot of sampling to answer one simple question: where does this LoRA actually live? LoRA Block Weight Group is the middle step - you've already seen a coarse sweep, you know the action concentrates in a few neighbourhoods, and now you sweep those neighbourhoods as groups instead of burning another 285 renders on single blocks.

FLUX.1's transformer is 19 double-stream blocks (D00–D18) followed by 38 single-stream blocks (S00–S37). Group mode treats any range of those as one unit, so you can ask "what happens if the whole D00–D06 region is knocked out at once?" - which is the question the README's demo actually answers. Knocking out the 7 highest-impact blocks stripped most of the LoRA's style; knocking out the 7 lowest did nothing visible. That's the kind of finding this node is built to produce.

The mechanism

The group sweep node is the whole pipeline in one box - it loads the LoRA, clones the model, patches per-group strengths, samples, and VAE-decodes - so it's not a loader like the plain single-block nodes. Every (group, mode) pair becomes one image in a batched images output, and the node returns info, groups_used and modes_used strings you can feed to the pack's Save Grid node for a labeled comparison.

The two inputs that actually matter

• groups - one group per line. D00-D06 is a range; D00-D06,S20-S25 combines ranges into one group. The default is an even 8-way split, and the tooltip is refreshingly honest about it: "naive even split with no prior knowledge - edit per LoRA once a first sweep shows where the effect concentrates."
• modes - a comma list of column types, each becoming one column of output: knockout (group off, everything else on), solo (group on, everything else off), full (whole LoRA on, the reference), and off (no LoRA, the baseline). Default is knockout,solo,full.

The rest - seed, steps, cfg, sampler_name, scheduler, denoise - are just the KSampler settings you already know, plus model, vae, lora_name, positive, negative and latent_image.

Installing it

Same pack, same install. ComfyUI Manager → search LoraBlockWeight, or:

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

Restart, and it shows up under the LoraBlockWeight category. No models to download, no extra pip installs.

Where people trip up

Because the batch/group nodes sample internally, the conditioning is already encoded before the LoRA is applied - the CLIP side of your LoRA is not included. The README says it plainly: if you need the LoRA's text-encoder effect, encode your prompts after a regular LoRA loader instead. If you type a block tag outside D00–S37, the node raises a clear error listing the valid range, so bad tags fail loudly rather than silently doing nothing. And remember a group sweep is still groups × modes images - with 8 groups and 3 modes that's 24 renders, cheap next to the full sweep, but not free.