Packs
Packs come to your server. Nothing about your library goes out, and nothing downloads until you choose it.
How packs work
A pack is a file your server can use to do more on its own: read picture subtitles in another language, name a folder whose name leaves the title unclear, search by what a title is about, know who is on screen. Each one runs on your server. None of them asks a service about your files, and none of them sends anything back.
Your server keeps a list of every pack, signed by Leyla, with each file's size and SHA-256. It downloads a pack only when you choose it in Server › Packs, and checks the file against the list before using it. Riff and Faces of 38,888 actors are Leyla's own and come from Leyla. Everything else comes straight from the people who made it, at one pinned version, and Server › Packs says where. A server without internet takes the same files by hand and checks them the same way.
This page is made from that list (number 5), so it says what your server will.
Models
A model is offered only to a server with the memory it needs. A row in grey is listed but not offered yet.
Reading names and text
| Pack | What it enables | Size | Licence | Comes from |
|---|---|---|---|---|
| Riff 0.1BLeyla's own model: fine-tuned by tropicalthink from SupersonicLabs/Julia-1 (Apache-2.0) on public Wikidata titles written as people name files, seed 2, Q8_0 with llama.cpp v0.7.0 | Helps name folders whose names leave the title unclear, and suggests an answer to the questions about them. Runs only while the server is idle. | 160 MB1 GB of memory | Apache-2.0 | From Leyla |
| Laya 0.4B | The same as Riff, answered better, for a server with room for it. | 449 MB4 GB of memory | Apache-2.0 | Not offered yet |
Looking at pictures
| Pack | What it enables | Size | Licence | Comes from |
|---|---|---|---|---|
| vllm-d3-edge 0.7B | Picks posters without text and clean backgrounds, and helps find intros and credits, after the first scan. | 806 MB4 GB of memory | Apache-2.0 | Not offered yet |
| Clef-Flash 9B | The same picture questions answered best, for a large server or one with a graphics card helper. | 6.5 GB16 GB, and a graphics card with 8 GB of memory | Apache-2.0 | Not offered yet |
Listening
| Pack | What it enables | Size | Licence | Comes from |
|---|---|---|---|---|
| d1-omni 0.4B | Helps tell intros, recaps and credits apart by their sound, after the first scan. | 407 MB4 GB of memory | lfm1.0 | Not offered yet |
Search that understands what you mean
| Pack | What it enables | Size | Licence | Comes from |
|---|---|---|---|---|
| Multilingual search 0.1Bibm-granite/granite-embedding-107m-multilingual (Q8_0) | Finds titles by what they are about, in any language, beside the usual search by name. | 115 MB1 GB of memory | Apache-2.0 | From Hugging Face |
| English search 22Msentence-transformers/all-MiniLM-L6-v2 (Q8_0) | The same search by what titles are about, in English only, in a fifth of the space. | 24 MB512 MB of memory | Apache-2.0 | From Hugging Face |
Who is on screen
| Pack | What it enables | Size | Licence | Comes from |
|---|---|---|---|---|
| Faces of 38,888 actorsWikimedia Commons portraits linked from Wikidata (P18) for the 50,000 most credited actors, 38,888 with a usable licence and one face: one 128-byte SFace embedding a person; attribution at faces-1.attribution.tsv beside the pack | Who is on screen while a title plays: the pass names the title's credited cast by face and voice, in the background, and the player shows them on pause. | 6 MB1 GB of memory | CC BY-SA 4.0 (the portraits' own licences, each kept in faces-1.attribution.tsv) | From Leyla |
Riff, Leyla's own model
Helps name folders whose names leave the title unclear, and suggests an answer to the questions about them. Runs only while the server is idle. Riff is Leyla's own model: fine-tuned by tropicalthink from SupersonicLabs/Julia-1 (Apache-2.0) on public Wikidata titles written as people name files, seed 2, Q8_0 with llama.cpp v0.7.0.
Measured on 1,337 folder groups the rules could not name and that it never saw in training: it answered 804 of them, and 96.3% of those answers were right, at a calibrated bar of 0.90. A group it is not sure of, it leaves to the rules and to you.
Subtitle languages
Picture subtitles (Blu-ray PGS, DVD VobSub) are read into text with Tesseract, so they can be restyled, searched and shown on any screen. English is in every Leyla; Server › Packs suggests the languages your libraries have.
| Pack | What it enables | Size | Licence | Comes from |
|---|---|---|---|---|
| Englishtesseract-ocr/tessdata_fast 4.1.0 | Reads English picture subtitles (PGS, VobSub) into text. Part of every Leyla. | Built in512 MB of memory | Apache-2.0 | In every Leyla |
| Frenchtesseract-ocr/tessdata_fast 4.1.0 | Reads French picture subtitles (PGS, VobSub) into text. | 1 MB512 MB of memory | Apache-2.0 | From GitHub |
| Germantesseract-ocr/tessdata_fast 4.1.0 | Reads German picture subtitles (PGS, VobSub) into text. | 1 MB512 MB of memory | Apache-2.0 | From GitHub |
| Spanishtesseract-ocr/tessdata_fast 4.1.0 | Reads Spanish picture subtitles (PGS, VobSub) into text. | 2 MB512 MB of memory | Apache-2.0 | From GitHub |
| Italiantesseract-ocr/tessdata_fast 4.1.0 | Reads Italian picture subtitles (PGS, VobSub) into text. | 3 MB512 MB of memory | Apache-2.0 | From GitHub |
| Portuguesetesseract-ocr/tessdata_fast 4.1.0 | Reads Portuguese picture subtitles (PGS, VobSub) into text. | 2 MB512 MB of memory | Apache-2.0 | From GitHub |
| Dutchtesseract-ocr/tessdata_fast 4.1.0 | Reads Dutch picture subtitles (PGS, VobSub) into text. | 6 MB512 MB of memory | Apache-2.0 | From GitHub |
| Swedishtesseract-ocr/tessdata_fast 4.1.0 | Reads Swedish picture subtitles (PGS, VobSub) into text. | 4 MB512 MB of memory | Apache-2.0 | From GitHub |
| Polishtesseract-ocr/tessdata_fast 4.1.0 | Reads Polish picture subtitles (PGS, VobSub) into text. | 5 MB512 MB of memory | Apache-2.0 | From GitHub |
| Russiantesseract-ocr/tessdata_fast 4.1.0 | Reads Russian picture subtitles (PGS, VobSub) into text. | 4 MB512 MB of memory | Apache-2.0 | From GitHub |
The identity pack
Naming a file is the first thing a media server does, and most do it by asking a service about every file. Leyla names them on your server, against a signed list of films and series built from Wikidata: 422,225 works with their titles in every language and their IMDb, TMDB or TVDB ids, 27 MB as a download and about 3 MB in memory. A typo, a missing year or a title in another script is looked up there, offline. Your file names are never sent anywhere.
The recommendation pack
Home suggests titles from your own library and your own watching, ranked on your server. Without a pack it compares titles by what they share (genre, people, studio, decade). A recommendation pack adds public neighbours: which titles are alike, built offline from approved public sources and signed, the same for every server. What you watched stays on your server; the pack only brings knowledge in. It is off until an administrator turns it on.