Drop a .safetensors / .pt embedding here
or click to choose
processed in your browser · never uploaded
Drop or click to replace
Embedding
- Trigger word
- Format
- Vectors (tokens used)
- Likely base model
For Textual Inversion the trigger is the file name — rename the file to change it.
Each vector is one CLIP token in your prompt — a 4-vector embedding eats 4 of your 75.
Tensors
Metadata
Raw header JSON
A tool that opens a downloaded Textual Inversion (embedding) .safetensors in your browser and shows what's inside. Unlike a LoRA, an embedding doesn't record how to call it — in A1111 and similar tools the trigger word is simply the file name without its extension. This tool surfaces that trigger word first, then shows the number of vectors in the embedding (which equals the CLIP tokens it consumes in your prompt — a 4-vector embedding uses 4 of your 75) and the base model inferred from the embedding dimension (768 = SD1.x, 1024 = SD2.x, 768 + 1280 = SDXL). It's for when you grab an embedding and don't know whether it's for SD1.5 or SDXL, how many tokens it eats, or how to call it at all — even if the description on the download page is gone. It reads only the header at the start of the file (the JSON of tensor names and shapes) and never loads the weights, so it opens instantly. Because trigger words and training results are your own creative clues, the file is never uploaded — all parsing happens locally in your browser. A .pt (pickle) embedding can't be read from a header, but even then the trigger word (the file name) is shown.
How to use
- Drop in a Textual Inversion / embedding .safetensors (or .pt) or click to choose one (it parses the header only — the weights are never loaded).
- It automatically shows the trigger word (the file name), the vector count, the likely base model (SD1.x / SD2.x / SDXL), and the tensor list.
- Type the trigger word straight into your prompt to call it — just remember it costs one token per vector.
FAQ
Is the file uploaded to a server?
No. All parsing happens in your browser. The embedding file is never uploaded, stored, or sent anywhere — it is read only on your device.
Where does the trigger word come from?
An embedding doesn't store its own activation word, so the file name (without the extension) is shown as the trigger word. A1111 and similar tools also call it by file name — rename the file to change the trigger.
What does the vector count mean?
It's how many vectors the embedding holds, which equals the number of CLIP tokens it consumes in your prompt. A 4-vector embedding uses 4 of your 75-token budget — useful to check when a long prompt is running out of room.
How does it tell SD1.5 from SDXL?
From the embedding dimension: 768 = SD 1.x (CLIP ViT-L), 1024 = SD 2.x (OpenCLIP ViT-H), and both 768 and 1280 = SDXL (its two text encoders). An embedding won't work on a model with a different dimension, so use this as a compatibility hint.
Can it read .pt embeddings?
It can detect the format, but .pt is a pickle format whose vector count and base model can't be read from a header. The trigger word (file name) is still shown. Convert to .safetensors to inspect the details.
Nothing shows up for my LoRA.
This tool is for Textual Inversion / embeddings only. If no CLIP-sized embedding tensor (768/1024/1280) is found, the file may be a LoRA or full model and it says so — use the Safetensors Info tool for LoRAs.