Analyze Diffusion Model Similarity (2 Models)

Same idea as its sibling node for checkpoints, but pointed at a different folder and a slightly different thing. A checkpoint bundles everything - UNet/DiT, text encoder, often a VAE - into one file. A standalone diffusion model in models/diffusion_models is just the denoiser: the UNet or DiT that does the actual generating. If you want to compare two UNets - say, a base Flux and a finetune built from it - without the text-encoder and VAE layers muddying the numbers, this is the node.

How it works

It streams two safetensors files from models/diffusion_models in low-memory mode (via the pack's Unified Efficient Loader, same as everything else in Model Utility Toolkit), and compares them layer by layer in fp32. You get the standard metric set - MAE, MSE, RMSE, max absolute difference, relative L2, cosine similarity, Pearson correlation, exact equality, sign agreement, norm ratio - plus a per-tensor breakdown, plus a second report with CWB diagnostics: row-wise cosine statistics and each model's affinity to a mean and median consensus, computed with the same machinery the pack's CWB merger nodes use when they actually blend two diffusion models.

Same honest rules as the checkpoint version: missing keys and shape mismatches are listed, never padded or zero-filled. Quantized models carrying ComfyUI quantization metadata are rejected, because stored quantized values aren't comparable to raw weights. If you've been eyeing a GGUF UNet and want to compare it to fp16, you can't - you'll get an error instead of misleading numbers, which is the right behavior.

Inputs you'll actually set

• model_a / model_b - the two files. Model A is the anchor; keys unique to each input are reported separately.
• top_weight_differences (default 20) - how many of the largest individual weight differences get their own list entries. Zero turns the list off; up to 1000 if you want the full picture.
• exclude_patterns / glob_patterns - leave matching tensors out of all metrics and topology counts. Regex by default. This is genuinely useful on diffusion models: skip a known-identical conditioning branch or a norm layer so it doesn't flatten the aggregate stats.
• process_device - cuda or cpu for the per-work-unit math. A CUDA OOM only retries the affected layer on CPU rather than dying.
• force_clear_cache (default on) - garbage-collect and clear the CUDA allocator after every work unit. Saves retained memory, costs time. Leave it on for big DiTs; turn it off if you're on a big machine and want speed.
• execution_mode - ANALYZE does the work; DOCUMENTATION ONLY returns the reference doc and loads nothing.

Outputs and wiring

You get comparison_report, cwb_report, and documentation as plain text. Wire them into a Show Text node. No model comes out the other end - this is a read-only diagnostic.

Why you'd reach for it

The classic move: you downloaded a fine-tune of an architecture you already have, and you want to know how far it actually drifted before you decide whether a CWB merge between it and another fine-tune makes sense. This node tells you which blocks moved and how much. It's also the sane first step before running CWBModelTwoMerger or CWBModelThreeMerger from the same pack - the CWB report is literally a preview of the merging math.

Install

Same pack as all of these: Model Utility Toolkit (silveroxides/ComfyUI-ModelUtils). ComfyUI Manager → "Install Custom Nodes" → search the pack title, or:

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

Restart ComfyUI. The heavyweight dependency is unifiedefficientloader (UEL); requests, Pillow, mutagen, and av in requirements.txt belong to the pack's downloader nodes, not this one. These nodes use ComfyUI's newer extension API, so update ComfyUI if the node doesn't appear after install.