PM Della

DARE prunes randomly and Breadcrumbs prunes both extremes. DELLA takes the smarter route: it looks at each row of the task vector and prunes adaptively, so bigger-magnitude parameters in a row are more likely to survive and small ones are more likely to get dropped. The result is a merge that keeps the significant changes and reduces interference, without the randomness of DARE or the two-sided carve-out of Breadcrumbs.

You reach for it when DARE's random pruning feels wasteful - when you have a strong sense that certain parameter directions genuinely matter and you don't want them to be lottery tickets. It needs two or more LoRAs plus a base model, and it extends DARE's rescaling.

How it works

Within each row of the delta parameters, DELLA assigns each parameter a keep-probability that's scaled between density − epsilon (for the row's smallest element) and density + epsilon (for its largest). So the top of each row has a higher chance of surviving the random draw, the bottom a lower one. Then it applies DARE-style rescaling to preserve magnitude, and optionally TIES sign consensus (the ties toggle). The epsilon knob sets how wide that probability spread is: bigger epsilon, more aggressive preference for large magnitudes.

The inputs that matter

• density - target fraction of weights retained (default 0.85).
• epsilon - half-width of the keep-probability range (default 0.1). Must satisfy density − epsilon > 0 and density + epsilon < 1, or the probabilities go out of bounds - the node will reject invalid combos.
• ties - add TIES sign consensus (default off).
• rescale_norm - rescaling strategy (l1, l2, linf, none, default).
• normalize - weight normalization for contributors (default true).

Output is a MergeMethod config for PM LoRA Merger.

Installing

Ships in the LoRA Power-Merger pack. ComfyUI Manager (search "LoRA Power-Merger") or:

cd ComfyUI/custom_nodes
git clone https://github.com/larsupb/LoRA-Merger-ComfyUI
cd LoRA-Merger-ComfyUI
pip install -r requirements.txt

Restart ComfyUI. Dependencies: PyTorch, lxml, mergekit.

Common issues

The parameter-boundary error is the classic first run: cranking density to 0.95 with epsilon at 0.1 breaks the density + epsilon < 1 rule. Keep epsilon modest (0.05–0.1) and density in the 0.7–0.9 range and you'll stay legal. If the merge barely differs from DARE, your epsilon is too small to matter - the adaptive pruning is what distinguishes them. And mismatched ranks between LoRAs fail at the merger: reconcile them first with PM LoRA Stack Decompose (rSVD). Like the rest of the family, the "(Mergekit)" label is legacy - current versions run DELLA on the pack's own delta-space code.