🧬 XS · DARE TIES Merge Models

Most of the checkpoints on CivitAI aren't trained - they're merged, which is exactly why so many merged models start to look like the same person. DARE TIES Merge Models is the "do it properly" version: it merges an SDXL or Illustrious-family stack against a reference using task arithmetic, with optional DARE dropout and TIES sign-alignment - the two language-model merging techniques that stop N-way blends from fighting each other. If you've ever combined two checkpoints and gotten something weaker than both parents, this node lets you debug why instead of crossing your fingers.

How it works

For each child in the stack, the node computes delta = child − reference - the direction that child pushes. Then, depending on merge_mode:

• weighted_delta - plain task arithmetic. Sum the deltas, done. Stable, but this is exactly the "average everything" approach that muddies both parents.
• dare_delta - DARE: each delta element is randomly dropped with probability drop_rate, survivors rescaled by 1/(1-drop_rate). Sounds destructive; it isn't - most of a fine-tune's delta is redundant noise, and DARE keeps the signal while killing cross-model conflict.
• ties_delta - TIES: each child delta is trimmed to its largest-magnitude ties_density fraction, then a two-pass sign election resolves conflicts where children push the same weight opposite ways.
• dare_ties_delta - both, in that order. The author's recommended starting recipe: drop_rate=0.10, ties_density=0.50.

The merged delta is scaled by delta_strength, then applied to a clone of the reference as a single diff patch per weight key. Children load one at a time and are discarded - RAM stays flat, but a merge is not instant. Grab a drink.

The inputs that matter

The first three inputs are a complete reference triplet from the checkpoint loader: reference_model, reference_clip, reference_vae. The VAE passes through untouched; the CLIP does too, unless you opt into merge_clip (leave it at off - SDXL text encoders are famously sensitive to merging, and a broken one makes every prompt feel wrong).

• model_stack - the child list, from one or more DARE TIES Model Stack nodes.
• merge_mode - the four-way enum above. The whole personality of the merge.
• drop_rate - DARE drop probability. 0.05–0.25 is the sane band; larger stacks tolerate less.
• ties_density - how much of each child's delta survives trimming. 0.4–0.7 is normal.
• delta_strength - overall influence. 1.0 = full merge, <1.0 stays near reference, >1.0 extrapolates, negative inverts.
• normalize_model_weights - on by default; rescales weights to sum to 1.0 so equal weights mean equal shares.
• seed - seeds DARE's random mask per key and child slot via blake2b; same seed + stack order = identical merge.
• merge_math_gpu - off by default; keep it off unless CPU is painfully slow. Faster on CUDA, but the TIES two-pass holds big fp32 tensors in VRAM and OOMs easily.

Outputs

model and clip feed straight into your KSampler; vae is your reference VAE back. The fourth output is the sleeper feature: merge_report, a plain-text rundown of the stack, weights, drop rate, density, strengths, and merged/skipped/mismatch key counts. If a child's keys don't match the reference (wrong architecture, mismatched text encoder), that's where you'll see it. Wire it to a text node and read it before blaming your prompt.

Install

Through ComfyUI Manager, search "ComfyUI XS DARE TIES". Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/x-stride/ComfyUI_XS_DARE_TIES

Restart ComfyUI. There's no requirements.txt and nothing to download - just ComfyUI's own torch/numpy/Pillow. What you do need is your own SDXL/Illustrious-family checkpoints in models/checkpoints. Nodes live under Advanced → model_merging → dare_ties; a two-child example workflow ships in examples/dare_ties_two_child_merge.json.

Where people get burned

• OOM during the merge, especially with ties_delta/dare_ties_delta or merge_math_gpu on. Flip GPU math off; the node warns it's the highest-VRAM path.
• Empty stack - ValueError: model_stack is empty. A disabled, "None", or zero-zero slot gets dropped silently; make sure at least one child survives.
• Order sensitivity. Unlike plain weighted averaging, the dare/ties modes are deterministic for a fixed stack order - reorder and the merge changes. By design (masks bind to slot index), but it surprises people.
• Slow. A multi-child TIES merge is minutes, not seconds. Run the Stack Visualizer first, or you'll redo it twice.

Quick recipe

Reference: your base Illustrious or SDXL. Two or three children, equal weights. merge_mode=dare_ties_delta, drop_rate=0.10, ties_density=0.50, delta_strength=1.0. If it feels closer to the base than you'd like, nudge delta_strength to 1.1 before touching anything else - and read the merge report once.