Analyze Embedding Similarity (2 Models)

Embeddings (textual-inversion vectors) are the smallest thing you can merge, and the least like the other model types. A checkpoint or a UNet has fixed layer coordinates - row 3 of tensor A is the same row in tensor B. An embedding is a bag of token vectors, and nothing guarantees that token #5 in one file means the same thing as token #5 in the other. That's why this node has a knob the checkpoint and diffusion-model versions don't: optional one-to-one alignment.

How it works

Like the rest of the pack's analysis nodes, it streams both files from models/embeddings in low-memory mode and produces two text reports. The standard one has the usual metrics - MAE, MSE, RMSE, relative L2, cosine similarity, Pearson correlation, and the exact-equality/sign-agreement checks - while missing keys and shape mismatches are reported separately rather than padded.

The second report is the CWB diagnostic, and here's where embeddings get special treatment. By default the node assumes tokens line up by index, which is the fast and often-wrong assumption. Flip cwb_similarity_alignment on and it runs a greedy one-to-one matching pass that pairs tokens by cosine similarity instead - reporting coverage, matched-score statistics, and how much better similarity alignment is than index alignment. The tooltip is honest about the cost: it's quadratic, and "can be slow for large embeddings." For a typical SDXL embedding with a few thousand vectors, that's a few seconds; for something huge it's a real wait. Enable it when you genuinely suspect two files were trained with tokens in different orders.

Inputs that matter

• model_a / model_b - the two files from your embeddings folder.
• cwb_similarity_alignment (default off) - the alignment pass described above. The one setting unique to this node.
• top_weight_differences (default 20) - largest individual differences to detail; zero disables the list.
• exclude_patterns / glob_patterns - drop matching tensors from every metric. Regex unless you toggle globs.
• process_device, force_clear_cache, execution_mode - same as the other analysis nodes: CUDA with CPU fallback on OOM, cache flushing on by default, and a DOCUMENTATION ONLY mode that loads nothing.

Outputs

comparison_report, cwb_report, documentation - plain text, wire to Show Text. Nothing is written or modified; this node exists purely to inform a decision, usually "should I merge these two embeddings with CWBEmbeddingTwoMerger?"

Why you'd reach for it

Two embeddings trained on the same concept by different people, and you want to see whether they learned the same thing or drifted apart. Or you want to check whether a new version of your own embedding is a conservative update or a rewrite. The alignment toggle is the reason to prefer this over comparing the raw files yourself - doing cosine-pair matching by hand is miserable.

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. The meaningful dependency is unifiedefficientloader; the other requirements.txt entries (requests, Pillow, mutagen, av) back the pack's downloader nodes. Keep ComfyUI current - this pack uses the newer extension API and won't appear on a stale core.