Ties Sum With Dropout

The DARE insight is a weird one: you can randomly drop the vast majority of a delta's values, and the model barely notices - the surviving few carry nearly all the information. Combine that with TIES' trim-and-vote and you get the merge technique people reach for when they want to pack a big stack of LoRAs into one without the whole pile degrading into noise. Ties Sum With Dropout is that combination, exposed as a single comfy-mecha node.

Inputs (all in delta space):

• deltas (delta) - the list of deltas, via Mecha Recipe List.
• probability (0.9) - dropout rate. Each delta value is kept with probability 1 − probability, so 0.9 means ~90% are dropped. Higher = more aggressive pruning.
• rescale (True) - divide the result by 1 − probability so magnitudes aren't shrunk by the dropout. Leave it on.
• della_eps (0.0) - a window on the rank distribution that biases which values get dropped (the "DELLA" variant). 0 disables it.
• seed (None) - makes the random dropout reproducible; set a non-negative value to lock it.
• k (0.2) - TIES trim fraction; 0.2 is the classic setting.
• vote_sgn (False), apply_median (False), apply_stock (False), cos_eps (1e-06), eps (1e-06), maxiter (100), ftol (1e-20) - the full TIES Extended surface for combining the survivors.

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

How it works

Per tensor, each delta gets its values randomly masked by a Bernoulli draw (seeded if you set seed), then the survivors run through the same trim → sign-vote → aggregate pipeline as TIES Extended. The rescale step is what stops the merged delta from ending up ~10× too small; drop it and your merge will look like it did nothing.

Why you'd reach for it

This is the community's answer to "I have fifteen style LoRAs and want one LoRA that keeps them all." The dropout acts as a regularizer - it stops similar LoRAs from reinforcing their overlaps into oversaturation while TIES' sign vote stops outright conflict. It's also the node where seed starts to matter: with dropout, the same merge run twice can give slightly different results, and reproducibility is nice when you're comparing recipes.

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

Three classic missteps. Inputs must be deltas, not checkpoints - the (delta) label is doing work. Output is a delta: add it to a base or it's a dead end. And if you raise probability toward 0.99, keep rescale on, because the rescaling math gets extreme and you'll chase faint results otherwise. If the run-to-run variation bothers you, set seed. merge_checkpointing behaves like the rest of the pack: fp16 CPU caching for stable branches, don't stack with cache units.