CWB Merge Checkpoints (3 Models)

The three-input version of the CWB checkpoint merger, and the place where Consensus-Weighted Blending actually gets to show off. With two models, CWB is really doing clever adaptive interpolation. With three, it's doing what it was designed for: each contributor votes on a consensus, and a layer where all three agree - or where one is an outlier - gets handled differently from a layer where they're split.

How it works

Model A is still the anchor: it supplies output metadata, anchors tensor names and shapes, and is the preservation reference when tensors go missing. Models B and C are equal-prior contributors. For each tensor, CWB computes an element-wise consensus, measures each of the three models' cosine similarity to it, and weights contributions by those similarities. That's where the third input changes the character of the merge - with three contributors, the median consensus becomes meaningful and genuinely more robust than the mean, because a single outlier can't drag a mean around. The robust_medn preset exists for exactly this situation.

Everything else behaves like the two-input version: output keys are the union of all inputs, tensors only present in a secondary input are copied or merged from the sources available, and mismatch_mode decides between preserving the anchor (skip), inserting zeros (zeros), or aborting (error) when tensors are missing or shape-incompatible.

Inputs that matter

• execution_mode - MERGE writes the file; DOCUMENTATION ONLY returns the CWB reference and opens nothing.
• model_a / model_b / model_c - your three checkpoints. Order matters only for A, which anchors metadata and naming.
• cwb_preset (default balanced_mean) - with three inputs, give robust_medn a look first. The name suffixes tell you what's inside: _medn = median consensus, _rn = norm rescaling, _dsc = dynamic similarity contrast, _softcb = soft comfort bandpass.
• output_filename (default cwb_merged_3_checkpoint) - written to models/checkpoints without extension.
• save_dtype / override_dtype - dtype for generated tensors; fp32 inputs stay fp32 unless you force it.
• exclude_patterns / discard_patterns / glob_patterns - regex (or glob) lists; excludes preserve matching layers from the anchor, discards omit them.
• lazy_load (default on) - UEL streaming, leave on so three multi-GB checkpoints don't stack in RAM.
• force_clear_cache (default on) - flush caches per layer; memory-safe but slower. Flip off if you've got headroom and want speed.
• process_device - CUDA with per-layer CPU retry on OOM, or CPU.

Outputs

output_filename (the written file), documentation (the CWB reference), and cwb_report (per-layer detail of what happened). Load the result with a normal CheckpointLoader to test it.

The realistic use case

Three-model merges are how a lot of checkpoints actually get made - take a base, a style specialist, and a character model, and CWB decides per layer which voice should dominate rather than you guessing a ratio. It's the same pack family as CheckpointModelAnalysis, so run that on pairs first if you're unsure what you're working with.

Install

From Model Utility Toolkit (silveroxides/ComfyUI-ModelUtils), installed once for the whole pack. ComfyUI Manager → search "Model Utility Toolkit", or:

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

Restart ComfyUI. Real dependency: unifiedefficientloader (UEL). Keep ComfyUI updated - the pack uses the newer extension API. The CWB nodes are marked experimental by the author, so expect the occasional edge case on oddball checkpoints.