The curated LoRA list¶
The curated list is a hand-checked index of LoRAs on Hugging Face that ships inside the package. It lives in zvisiongenerator/utils/curated_loras.json, read by utils/curated_loras.py. The pictures on its tiles are not bundled: they load from Hugging Face. The list changes only with a release.
No hub offers safe-only results, so safety comes from someone looking at what is listed. An entry is curated only when all of these hold.
Criteria¶
Safe. Someone looked at the card and at the preview images. Hugging Face's own adult-content tag is not trusted: it is set by authors, and explicit uploads go untagged. A repository with explicit, hateful or otherwise unsuitable content is not listed, and neither is a gated one.
A working family match. The file's safetensors header, read with detect_lora_family, names the family the entry is listed under (and the size, for Klein). A repository whose tags say one family and whose weights say another is not listed. LTX-2.3 and LTX-2.5 share one detected family, so an LTX entry carries its version and is offered only for that version's model; utils/ltx_version.py reads a model's version without loading it.
Real previews. The repository shows what the LoRA does: two to four preview images that come from the LoRA itself, not a trainer's generic sample prompts. A repository whose previews are only videos lists two to four of them instead. A repository without previews is not listed.
A licence that was read. The licence is one of the ids in LICENCES. A non-commercial licence is listed, with non_commercial set, so the UI can badge it. A licence not in the table is read first, then added to LICENCES. A LoRA takes its base model's non-commercial badge only when that base makes the images non-commercial too, whatever the card says: the Ideogram 4 LoRAs are ideogram-4-non-commercial. A LoRA on FLUX.2 Klein 9B keeps its card's licence, because FLUX's non-commercial terms restrict use of the model while its section 2.d allows the images commercially.
Tested platforms. Content is looked at once. The file format (kohya, PEFT, ai-toolkit, LoKr, ComfyUI) is loaded on both macOS and Windows or Linux.
One per series. Per-step uploads, snapshots of one training run, near-identical sets and templated series keep their best entry. A repository with several files lists the one the author marks as final, or the one without a step number.
Tools only when the app can use them. A tool (it does something, as opposed to a style, which looks like something) is listed only when the app can apply it. Its workflow says which workflow it needs (txt2img, img2img, txt2vid, img2vid or extend; none when any works), and what_it_does says what it does in one line. A tool that needs nodes, masks or a pipeline the app does not have is not listed.
Rules the JSON does not show¶
An entry's fields are in curated_loras.json and the test below checks their shape. These are the rules behind them:
idis stable across releases, the repository path as a slug.nameis a readablesnake_caseslug, unique across the list: it becomes the file name on disk and what presets, Reuse and--lorarefer to, so it is never a generic file name such aspytorch_lora_weights.family,size_variant,versionname the family asutils/lora_family.pyand the video presets do, with a size (4b,9b) for Klein and a version (2.3,2.5) for LTX.descriptionis one short, factual line written from the card and the previews, without claims the card does not make.- The file is pinned.
repo,revision(a full 40-character commit),file,size_bytesandsha256(the Hub'slfs.sha256), and the preview URLs are pinned to the same commit. triggerholds the words the card names for the prompt, comma-separated, or none. A card that gives a sentence keeps the one word that triggers the LoRA. No word ends in a full stop and none repeats. It is never derived from the file name or from training metadata such asss_output_name.weightis 1.0 unless the card clearly recommends another value.downloadsis the repository's Hugging Face download count, shown on the cards.make check-curatedrefreshes it; a new entry starts at0.previewsandpreview_videosare two to four URLs in all, onhuggingface.co.
Adding or changing an entry¶
- Check the repository against the criteria above. Read the card, look at every preview, and load the file on each platform that applies.
-
Take the repository's current commit and the file's size and hash from the Hub:
The response holds
sha(the commit) and, for each file undersiblings, itssizeandlfs.sha256. Pin the full commit, and write the preview URLs against it (…/resolve/<commit>/<path>). 3. Add the entry tocurated_loras.jsonin the order the list is shown: best first within a family. 4. Runuv run pytest tests/contracts/test_curated_loras.py. It fails on a duplicateidorname, a family or size the app does not know, a malformed commit or hash, a file that is not.safetensors, a preview off the Hub, and a licence outside the allowed set. 5. Before a release, runmake check-curated(see Releasing). It also refreshes every entry'sdownloads.