Civitai backup: archive downloaded models locally and verify the restore

A Civitai download is easy to recognise today and surprisingly hard to identify six months later. Preserve the model file together with the version facts that make it usable, prove an exact restore, and only then decide whether a redundant working copy can go.

Where did Civitai put the model?

Civitai serves the download; your browser, extension or model manager chooses the destination. There is no universal “Civitai models” directory. A LoRA downloaded for ComfyUI may be in that runtime’s LoRA folder, while a checkpoint saved through the browser may still be in Downloads. Start from the workflow that uses the file, not from a guessed platform path.

Search for the filename shown on the Civitai version page and confirm it against a stable identifier. Civitai’s API exposes a model-version ID and file hashes, including SHA-256. Its by-hash endpoint can help reconnect an orphaned local file to the version record:

# Replace HASH with the SHA-256 of your downloaded file
https://civitai.com/api/v1/model-versions/by-hash/HASH

Record the facts the filename cannot carry

Before archiving, keep a small source note beside the file. Record the Civitai model and version URLs or IDs, creator, exact filename, SHA-256, intended base model, trained words or trigger terms when relevant, and the licence or creator terms you relied on when you downloaded it. Those details are far more useful than a folder called old-loras.

Then define a family that will still make sense later: one base and adapters intended for it, sequential releases of the same model, or one checkpoint plus its required configuration. Do not sweep an entire downloads drive into a single archive.

The safe local archive workflow

The critical rule is simple: do not delete a source just because an archive job finished. Prove recovery first.

1. Confirm the local file against its Civitai version record and save the provenance note with it.
2. Download Tensor Archive, choose Add model family, select Local folder, and point it at one related family.
3. Run Analyse source. Review the detected files and expected result before anything is written.
4. Name the family clearly and choose Archive folder. Tensor Archive creates a managed local copy; it does not move or delete the source folder.
5. Open the archived family and run Verify. This establishes that the stored version can be recovered exactly. It does not assess image quality, inference quality or training quality.
6. Run Restore to a separate output. Load or compare that recovered version in the local workflow that matters to you.
7. Only after the restored copy is verified, re-hashed and tested in the target runtime should you decide whether to remove an old source copy manually.

A local archive is not your only backup

An archive on the same machine can reduce working-disk pressure, but it does not protect against disk failure, theft or accidental removal of the application data. If a model library is irreplaceable, retain an independent backup on storage you control.

This matters for community downloads because a model page, version or account can change. Re-downloading is a convenience, not a recovery plan. The local archive protects the copy you already hold; the provenance note explains what that copy is.

What space saving should you expect?

Expect a measurement, not a slogan. In Tensor Archive’s published controlled benchmark, five self-contained model-and-adapter deployments occupied 907.3 MB independently and 261.6 MB as one family: 71.2% less physical storage. The gain came from retaining shared model data once within that family.

Analyse the family you actually have. If the files have no meaningful overlap, a compact local archive may save little; that is a result to respect, not a reason to delete the source anyway.

Start with the models you reach for most

Choose a base model plus the adapters you use with it, rather than attempting a heroic cleanup of everything at once. Archive it, verify it, restore it and test it. Once the recovery workflow is trusted, repeat it family by family.

Start with one related family, not your whole disk. If that family is mostly adapters, use the LoRA disk-space guide; for mixed runtimes and downloads, use the library retention checklist. Then download Tensor Archive and follow the first-family workflow in the product guide.