Ties Sum Extended

Ties Sum Extended is what you graduate to after the plain Ties Sum starts feeling limiting. Same core: trim each delta, majority-vote the sign, keep the agreement. But this node exposes the newer machinery from the TIES/mergekit lineage - geometric median aggregation, model-stock rescaling, and a wall of tolerances - so you can tune how the surviving deltas get combined instead of taking the plain average.

Inputs:

• models (delta) - the list of deltas to merge (via Mecha Recipe List).
• k (0.2) - trim fraction, top-magnitude values kept. 0.2 is the classic TIES setting; here it's the default rather than the afterthought.
• vote_sgn (False) - majority vote on sign only.
• apply_median (False) - if on, aggregate the surviving deltas with a geometric median instead of a mean. More robust to outliers, slower.
• apply_stock (False) - if on, apply model-stock (mergekit-style) magnitude rescaling instead of the plain t/param_counts normalization.
• cos_eps (1e-06) - cosine tolerance used by the stock rescaling.
• eps (1e-06), maxiter (100), ftol (1e-20) - convergence controls for the geometric median solver.

Output is a MECHA_RECIPE in delta space. Add it back to a base via Add Difference to get a model.

Why you'd reach for it

When your TIES merges keep getting dragged by one outlier LoRA. apply_median is the headline feature - the geometric median isn't pulled toward extreme values the way a mean is, which matters when you're merging a stack of LoRAs where one is wildly stronger than the rest. The apply_stock toggle is for people porting recipes over from mergekit who want matching rescaling behavior.

How it works

Trim and sign-vote exactly like plain TIES, then the difference: instead of sum / count, the survivors go through either a geometric-median optimization (an iterative solver governed by eps/maxiter/ftol) or model-stock scaling (which needs cos_eps). Everything streams per tensor key, so it's still memory-cheap even though the median solver is the slowest part of the pack's TIES family.

Install

Pack-level: ComfyUI Manager → search "mecha" → Mecha Merge Node Pack, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Restart ComfyUI. Only dependency is sd-mecha; no model downloads.

Gotchas

Don't turn on apply_median and apply_stock together out of curiosity - they're two different answers to the same problem (how to combine the survivors), and using both means apply_stock is ignored. And remember the defaults matter: this node's k is 0.2, unlike plain Ties Sum's 1.0, so the two nodes genuinely do different things out of the box. Output is a delta; add it back to a base, or the merge goes nowhere. merge_checkpointing works like the rest of the pack - fp16 CPU caching for stable branches, don't stack with cache units.