Delta CWB LoRA Multi-Merge

A plain average of five LoRAs is how you end up with a model that looks like none of them. Delta CWB LoRA Multi-Merge is the maxed-out version of the pack's consensus merge: it takes anywhere from 2 to 8 equal-prior LoRAs, expands each into full per-layer weight changes, and merges the rows with Consensus-Weighted Blending - so a row only keeps influence if it agrees with the group consensus. The further you get from two inputs, the more that "only what agrees survives" property matters: with several LoRAs in the mix, CWB's mean-or-median consensus plus similarity weighting is closer to taking a vote than taking an average.

Like its two- and three-input siblings (Delta CWB Merge LoRAs (2) / (3)), this is a file-making node: it writes a merged .safetensors into your loras folder rather than patching anything live. LoRA 1 anchors the output layer set and metadata; every other input is an equal contributor whose actual influence is decided by the agreement math, row by row.

The inputs that matter

• lora_count (default 2, up to 8) - must match the number of slots you actually filled. Count says 5 but lora_6 is still None? It errors out with "unselected input(s)" rather than silently merging a partial set. The count is the first thing to check when a batch merge refuses to run.
• lora_1 through lora_8 - lora_1 and lora_2 are mandatory; the rest default to None.
• cwb_preset (default balanced_mean) - the six presets run from simple blends (balanced_mean, robust_medn, selective_mean) to the norm-aware, diversity-boosting ones (varied_mean_rn_softcb, diverse_medn_rn_dsc_softcb, strongdiv_medn_rn_dsc_softcb). For big multi-LoRA merges the long-name presets exist specifically to keep one dominant LoRA from flattening the rest - when you've got five inputs of very different strengths, diverse_medn_rn_dsc_softcb or strongdiv_medn_rn_dsc_softcb are worth testing first.
• cwb_config (optional) - connect the pack's CWB Custom Configuration node and it replaces the preset entirely.
• mismatch_mode (default skip) - with up to eight independently trained LoRAs, missing layers aren't a corner case. skip keeps LoRA 1's version, zeros treats absent inputs as zero contribution, error aborts.
• output_representation / output_rank - full_difference saves the exact merged dense deltas; factorized_lora re-factorizes each merged layer into compact lora_A/lora_B factors up to output_rank (default 384) if you want a conventional LoRA-shaped file instead.
• output_filename, save_dtype (default bf16), process_device (CUDA with per-layer CPU fallback), exclude_patterns/discard_patterns/glob_patterns, and force_clear_cache complete the panel.

Outputs are output_filename and cwb_report - the report lists merged vs. anchor-preserved layers, zero contributors, rejected unsupported layers, and CUDA→CPU fallbacks. With many inputs, that report is your only window into whether the merge did what you hoped; wire it to a text node and read it.

Honest expectations

This is the pack's most experimental corner - the whole Delta CWB family is marked experimental in the source, with no real community track record yet, so verify results against a plain stacked-LoRA render before you commit. It only merges canonical LoRA factor pairs; DoRA-scale and low-bit factors raise per layer. Streaming one layer at a time keeps memory proportional to a single layer rather than eight full LoRAs, and a layer that OOMs on CUDA retries on CPU. Output is atomic: a failed run leaves any existing same-named file untouched, but a successful rerun overwrites it - name your merges.

Installing it

ComfyUI Manager: search Model Utility Toolkit. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-ModelUtils

Restart ComfyUI. No API keys, no downloads - it reads only the LoRAs already in your models/loras folder. Find it under ModelUtils → LoRA → Merge → Delta CWB or search "Delta", and grab the Multi entry (the 2 and 3 variants are fixed-slot versions of identical machinery). If you want the same batch-expand idea without the consensus weighting, the Lodestone Multi merger in the same pack is the simpler, norm-based alternative.