Drop a .safetensors file here
or click to choose
processed in your browser · never uploaded
Drop or click to replace
Metadata (__metadata__)
An in-browser editor for the __metadata__ block embedded in a .safetensors LoRA or model — the write counterpart to the Safetensors metadata viewer. Edit, add or remove key/value pairs such as ss_output_name (name), ss_sd_model_name (base model), modelspec.title / modelspec.description, or your own trigger-word notes, then save a new file. Use it to give a LoRA a clear name and description before sharing, add trigger-word notes to a model whose metadata is empty, or fix a wrong base-model name. The key point: it only reads and writes the file's leading header (the JSON of tensor names and training metadata) and never loads or changes the weights themselves. Because data_offsets are relative to the start of the data section, the weights are concatenated unchanged even when the metadata length changes, so the model stays bit-for-bit intact. That also means even multi-gigabyte checkpoints edit instantly. Trigger words and training settings are your creative notes and don't need to leave your machine, so this tool never uploads the file — reading, editing and saving all happen entirely in your browser.
How to use
- Drop a .safetensors file or click to choose one (only the header is read, not the weights).
- The existing __metadata__ loads into a key/value list. Edit values, add fields with Quick add or the presets, or remove rows you don't need.
- Click "Save edited .safetensors" to download a new file with the metadata replaced and the weights left untouched.
FAQ
Is my file uploaded to a server?
No. Reading, editing and saving all happen inside your browser. The .safetensors file is never uploaded, stored or transmitted — it is processed only on your device, so it is safe to edit unreleased models or private training files.
Are the weights (the model itself) changed?
No. The tool only rewrites the __metadata__ block in the leading header and concatenates the tensor bytes unchanged. For peace of mind, open the saved file in the Safetensors metadata viewer to confirm the contents.
Can I add metadata to a LoRA that has none?
Yes. Even if a file has no __metadata__ block, adding fields and saving will create a new __metadata__ block from scratch.
Which keys should I use?
There are no strict rules, but kohya/sd-scripts keys like ss_output_name and ss_sd_model_name, or modelspec.title and modelspec.description, are common. Use the preset buttons to add frequent keys quickly. You can keep trigger words under any key (e.g. trigger_words) as your own note.
Can it handle multi-gigabyte models?
Yes. The weights are never loaded into memory — only the leading header is read and written, and the rest is concatenated as-is, so even large checkpoints edit and save quickly.
What formats are supported?
Any .safetensors file — LoRA, LyCORIS or checkpoint — can have its header __metadata__ edited. Other formats such as .ckpt are not supported.