index

Types

TrainedLora

What a finished run produced.

/** What a finished run produced. */
export type TrainedLora = {
  path: string;
  steps: number;
  images: number;
  minutes: number;
  samples: string[]
}

(source)

LoraInfo

What an adapter's file says about how it was trained.

/** What an adapter's file says about how it was trained. */
export type LoraInfo = {
  base: string;
  trigger: string;
  rank: number;
  steps: number;
  sizeBytes: number
}

(source)

Effects

lora::train

@alwaysUnder(outDir, imagesDir)
effect lora::train {
  imagesDir: string;
  base: string;
  outDir: string;
  outFilename: string;
  steps: number;
  estimatedMinutes: number
}

(source)

lora::info

@alwaysUnder(dir)
effect lora::info {
  dir: string;
  filename: string
}

(source)

Constants

STYLE_TAGS

export static const STYLE_TAGS = [
  "monochrome",
  "greyscale",
  "white background",
  "simple background",
  "traditional media",
  "sketch",
  "lineart",
  "signature",
  "text focus",
  "no humans",
]

(source)

Functions

trainLora

trainLora(
  imagesDir: string,
  trigger: string,
  base: string,
  outPath: string,
  steps: number = 1000,
  rank: number = 16,
  learningRate: number = 0.0001,
  resolution: number = 1024,
  flip: boolean = false,
  seed: number = 1,
  samplePrompts: string[] = [],
  sampleEvery: number = 250,
): Result<TrainedLora> raises <lora::train>

Train a LoRA adapter for an SDXL image model from a folder of images, each with an optional caption file (a.txt beside a.png, comma-separated tags). The trigger word is put in front of every caption. Takes minutes on a Mac GPU and writes the adapter as a .safetensors file, plus a before-and-after sample grid at every sampleEvery steps beside it. Everything runs on this machine, with the Python agency local serve uses. Returns the adapter's path and the sample grids.

@param imagesDir - The folder of training images, PNG or JPEG, with optional .txt captions beside them @param trigger - The word the adapter answers to. A real phrase learns faster and keeps the base model's idea of it; a nonsense word owns the token @param base - The SDXL model to train on: a catalog name, a diffusers: URI, or a model directory, already downloaded @param outPath - The .safetensors file to write. It must not exist yet @param steps - How long to train. Too few and the style is faint; too many and every output is a training image. Judge by the sample grids @param rank - How much the adapter can hold: 8 for a style, 16 for a character, 32 for a character with a wardrobe. Doubling the rank doubles the file @param learningRate - Halve it if the grids get worse after getting better @param resolution - The training size. 768 trains twice as fast for a first look; 1024 for the real run @param flip - Also train on mirrored copies, which doubles a small set. Off for an asymmetric character @param seed - Fixes the randomness, so a run is repeatable @param samplePrompts - Prompts to render before and after at each checkpoint, to judge the run. Use the trigger word in them @param sampleEvery - Steps between sample grids. 0 for none

Parameters:

Name	Type	Default
imagesDir	string
trigger	string
base	string
outPath	string
steps	number	1000
rank	number	16
learningRate	number	0.0001
resolution	number	1024
flip	boolean	false
seed	number	1
samplePrompts	string[]	[]
sampleEvery	number	250

Returns: Result<TrainedLora>

Throws: lora::train

(source)

loraInfo

loraInfo(path: string): Result<LoraInfo> raises <lora::info>

What an adapter file says about itself: the base model it was trained for, its trigger word, its rank, the steps it trained, and its size. Reads the file's header only.

@param path - A .safetensors adapter written by trainLora

Parameters:

Name	Type	Default
path	string

Returns: Result<LoraInfo>

Throws: lora::info

(source)