CWB Merge Models (2 Models)

This is the CWB merger for standalone diffusion models - the UNet or DiT files in models/diffusion_models, without the text encoder and VAE that a full checkpoint bundles in. If you're merging the denoiser alone, this is the node; if you want the whole checkpoint including encoders, use CWBCheckpointTwoMerger instead.

How it works

Same Consensus-Weighted Blending engine as the rest of the pack's CWB nodes. Model A is the anchor - output metadata, tensor name/shape anchoring, preservation reference. Model B is the second equal-prior contributor. For each tensor, the merger computes an element-wise mean or median consensus, measures each model's cosine similarity to that consensus, rejects contributors below the threshold, raises survivors to a power, normalizes the weights, and produces the weighted sum - then optionally rescales the result's norm to the participating average. There's no global alpha to dial; per-layer similarity is the weighting.

Because these are standalone diffusion models, the merge output lands in models/diffusion_models and gets loaded with a UNet Loader (or the DiffusionModel Loader), not a CheckpointLoader. That's the practical difference from the checkpoint variant: you're building the denoiser, and you still need a text encoder and VAE elsewhere in your workflow.

Inputs that matter

• execution_mode - MERGE writes the file; DOCUMENTATION ONLY returns the CWB reference without loading or merging.
• model_a / model_b - the two files. A anchors metadata and naming.
• cwb_preset (default balanced_mean) - dense-model presets; robust_medn for median consensus, strongdiv_medn_rn_dsc_softcb to push diversity hard.
• mismatch_mode - skip preserves the anchor, zeros inserts zero contributions, error aborts for missing or incompatible tensors.
• output_filename (default cwb_merged_2_model) - written to models/diffusion_models.
• save_dtype / override_dtype - fp32/fp16/bf16 for generated tensors.
• exclude_patterns / discard_patterns / glob_patterns - regex (or glob) lists; excludes preserve matching layers from the anchor, discards drop them from the output.
• lazy_load (default on) - UEL streaming, keep it on for multi-GB DiTs.
• force_clear_cache (default on) - per-layer cache flushing; memory-safe, slower.
• process_device - CUDA with per-layer CPU retry on OOM.

Outputs

output_filename (the written file), documentation (the CWB reference), and cwb_report (per-layer merge detail). Load the result with the model loader for your architecture.

Why you'd reach for it

The common scenario: you have two fine-tuned UNets (say, a realism-tuned Flux and a detail-calibrated variant) and you want the strengths of both in one file. Or you're doing the trendy thing - merging a base DiT with a style specialist - and you want per-layer adaptation instead of a fixed 50/50. Run DiffusionModelAnalysis on the pair first to preview the consensus math and see whether the merge is likely to be a clean union.

Install

Part of Model Utility Toolkit (silveroxides/ComfyUI-ModelUtils). 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) for streaming. Keep ComfyUI current - the pack uses the newer extension API, and the CWB nodes are author-flagged as experimental.