Embedding Info Loader

Embeddings are the least self-explanatory files in your whole model folder. A checkpoint is multiple gigabytes and generates on its own - you can just try it. A LoRA at least has a name and usually a trigger word someone bothered to document. An embedding is 10–80KB of vectors with a filename and nothing else: no way to tell what it does short of loading it and burning a generation to find out. This node exists to close that gap - it pulls up the preview, the generating workflow, and the metadata for an embedding without you having to guess.

It's one of the Info nodes in ComfyUI-ModelUtils, silveroxides' toolkit for inspecting and manipulating model files. Worth knowing going in: embeddings (textual inversion) as a technique are in genuine decline. They're bound to whichever text encoder they were trained against, and every current architecture - Z-Image, Flux 2, Krea 2 - swapped CLIP for an LLM encoder that a CLIP-trained vector has nothing to bind to. New textual-inversion uploads on CivitAI fell roughly twentyfold from 2024 to the first half of 2026, and what's left is almost entirely SDXL/Illustrious pose-and-character files rather than the old "packaged negative prompt" use case. None of that makes this node less useful - if anything it makes it more useful, because the ecosystem is now a pile of old, sparsely documented files rather than a living, well-labeled one.

How it works. Same pattern as the pack's other Info loaders: it reads locally cached info rather than fetching anything live. That cache normally comes from running CheckpointInfoMetaDownloader-style siblings first (this pack ships one for checkpoints, VAEs, ControlNets, and diffusion models - no dedicated embedding downloader in this batch, so however that cache gets populated for embeddings, this node is strictly the read side).

Inputs that matter. embedding is a dropdown built from your local embeddings folder. workflow_index (default 0, 0–100) picks which cached example to show, for embeddings that have more than one on file.

Outputs. preview (IMAGE) is the example generation for that index - genuinely useful here, since it's often the only visual reference you'll ever have for what an embedding does. workflow_json (STRING) is the full workflow behind that preview, handy if you want to reproduce the exact settings someone used to show off a character embedding. metadata_json (STRING) is whatever description, base model, and trigger-word info got captured.

Installing it. Through ComfyUI Manager - search "Model Utility Toolkit" - or manually:

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

Restart ComfyUI. No extra models to fetch for this node specifically; it works off what's already in your embeddings folder.

Troubleshooting. Blank outputs almost always mean the cache was never populated for that file - the node won't go fetch it for you on the spot, it just reports empty. The other trap, specific to embeddings: since they're locked to a text encoder, an old SDXL-era embedding with a rich, well-documented cache entry will still do essentially nothing if you load it against a Z-Image or Flux 2 checkpoint. This node will happily show you a lovely preview and full metadata for a file that, loaded into your current workflow, changes nothing - check the base model in metadata_json before you assume it'll transfer.