Merge Embeddings (2 Models)

Takes two embeddings and blends their learned vectors into a single new file. It's the merging node in silveroxides' ComfyUI-ModelUtils - a pack better known for its LoRA extraction and key-editing tools, but this one applies the same "combine two model files mathematically, no training" idea that the community has been doing to full checkpoints for years, scaled down to embedding size.

Why merge embeddings. Checkpoint merging is how a huge chunk of CivitAI's catalog gets made - it's fast, needs no GPU training run, and just interpolates weights. Embedding merging is the same trick at a tiny scale: if you've got two compatible embeddings that each capture part of what you want (say, two related style vectors), blending them can land somewhere in between without retraining either one from scratch.

How it works, and what to actually touch. This node exposes a lot of knobs - execution_mode, calc_mode, six greek-letter floats (alpha through zeta), mismatch_mode, alignment_mode - but most of them exist for edge cases you won't hit on your first run. calc_mode picks the blend formula; alpha and beta both default to 0.5, which reads as a straightforward 50/50 split between model_a and model_b, and that's the pair worth adjusting first if you want one embedding to dominate. gamma through zeta feed whichever formula calc_mode is set to - leave them at their defaults unless you already know why you're changing them.

The pair that actually matters more than the greek letters: mismatch_mode (default skip) and alignment_mode (default pad/crop). They exist because two embeddings don't automatically have matching shapes - different vector counts, or worse, different text encoders entirely - and the node needs a policy for what to do when model_a and model_b don't line up cleanly. skip drops keys that don't match rather than crashing the merge.

Inputs and outputs worth knowing. model_a/model_b are your two source embeddings. output_filename (default merged_2_embedding) names the result. seed matters if calc_mode involves any randomization. save_dtype and process_device control output precision and where the math runs. exclude_patterns/discard_patterns (with glob_patterns toggling glob vs regex syntax) let you skip specific keys from the merge entirely. Two outputs come back: output_filename and documentation - the latter is a text explanation of what the merge actually did, worth piping to a Show Text node if you want to sanity-check the run.

Installing it. ComfyUI Manager, search "Model Utility Toolkit" - or cd ComfyUI/custom_nodes && git clone https://github.com/silveroxides/ComfyUI-ModelUtils, then restart. No extra dependencies in the README, nothing to download.

Where the effort is actually worth spending, and where it isn't. lazy_load and force_clear_cache are there because this pack's LoRA and model merge nodes deal with files that can be gigabytes - but embeddings are 10-80KB. On a merge this small, those two toggles are functionally no-ops; don't spend time tuning them here the way you would on a full checkpoint or LoRA merge.

The failure mode that actually bites. Merging two embeddings trained against different text encoders doesn't error out cleanly - it produces a file that loads without complaint and then does nothing useful, because the resulting vectors don't correspond to anything meaningful in either encoder's space. Embeddings are architecture-locked: an SD 1.5 vector and an SDXL vector aren't secretly compatible just because they're both .safetensors files. Check that model_a and model_b were actually trained on the same base model family before merging - mismatch_mode will save you from a hard crash, but it won't save you from a merge that's mathematically meaningless.