edenartlab/sdxl-lora-trainer – Run with an API on Replicate

edenartlab / sdxl-lora-trainer

LoRa trainer for both SDXL and SD15

Cold

Public
7K runs
L40S
GitHub
License

Run with an API

Playground API Examples README Versions

Input

name

string

Shift + Return to add a new line

Name of new LORA concept

Default: "unnamed"

lora_training_urls

*string

Shift + Return to add a new line

Training images for new LORA concept (can be image urls or an url to a .zip file of images)

concept_mode

string

What are you trying to learn?

Default: "style"

sd_model_version

string

SDXL gives much better LoRa's if you just need static images. If you want to make AnimateDiff animations, train an SD15 lora.

Default: "sdxl"

max_train_steps

integer

Number of training steps. Increasing this usually leads to overfitting, only viable if you have > 100 training imgs. For faces you may want to reduce to eg 300

Default: 300

checkpointing_steps

integer

Save a checkpoint every n steps (The final checkpoint will always be saved)

Default: 10000

resolution

integer

Square pixel resolution which your images will be resized to for training, highly recommended: 512 or 768

Default: 512

unet_lr

number

final learning rate of unet (after warmup), increasing this usually leads to strong overfitting

Default: 0.0003

ti_lr

number

Learning rate for training textual inversion embeddings. Don't alter unless you know what you're doing.

Default: 0.001

lora_rank

integer

Rank of LoRA embeddings for the unet.

Default: 16

n_tokens

integer

(minimum: 1, maximum: 4)

How many new tokens to train (highly recommended to leave this at 2)

Default: 3

train_batch_size

integer

Batch size (per device) for training (dont increase unless running on a BIG GPU)

Default: 4

n_sample_imgs

integer

Number of sample images in validation grid

Default: 4

validation_img_size

integer

Resolution of sample images in validation grid

Default: 1024

sample_imgs_lora_scale

number

Scale factor for LoRa when generating sample images. If not provided, will be set automatically

seed

integer

Random seed for reproducible training. Leave empty to use a random seed

Run this model in Node.js with one line of code:

npx create-replicate --model=edenartlab/sdxl-lora-trainer

or set up a project from scratch

Install Replicate’s Node.js client library:

npm install replicate

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Import and set up the client:

import Replicate from "replicate";

const replicate = new Replicate({
  auth: process.env.REPLICATE_API_TOKEN,
});

Run edenartlab/sdxl-lora-trainer using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

const output = await replicate.run(
  "edenartlab/sdxl-lora-trainer:4767bababe6048535114863799de828c25ec5b935dab7f879d4fa29495118d22",
  {
    input: {
      name: "unnamed",
      ti_lr: 0.001,
      unet_lr: 0.0003,
      n_tokens: 3,
      lora_rank: 16,
      resolution: 512,
      concept_mode: "style",
      n_sample_imgs: 4,
      max_train_steps: 300,
      sd_model_version: "sdxl",
      train_batch_size: 4,
      checkpointing_steps: 10000,
      validation_img_size: 1024
    }
  }
);
console.log(output);

To learn more, take a look at the guide on getting started with Node.js.

Install Replicate’s Python client library:

pip install replicate

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Import the client:

import replicate

Run edenartlab/sdxl-lora-trainer using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

output = replicate.run(
    "edenartlab/sdxl-lora-trainer:4767bababe6048535114863799de828c25ec5b935dab7f879d4fa29495118d22",
    input={
        "name": "unnamed",
        "ti_lr": 0.001,
        "unet_lr": 0.0003,
        "n_tokens": 3,
        "lora_rank": 16,
        "resolution": 512,
        "concept_mode": "style",
        "n_sample_imgs": 4,
        "max_train_steps": 300,
        "sd_model_version": "sdxl",
        "train_batch_size": 4,
        "checkpointing_steps": 10000,
        "validation_img_size": 1024
    }
)

# The edenartlab/sdxl-lora-trainer model can stream output as it's running.
# The predict method returns an iterator, and you can iterate over that output.
for item in output:
    # https://replicate.com/edenartlab/sdxl-lora-trainer/api#output-schema
    print(item)

To learn more, take a look at the guide on getting started with Python.

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Run edenartlab/sdxl-lora-trainer using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

curl -s -X POST \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Prefer: wait" \
  -d $'{
    "version": "4767bababe6048535114863799de828c25ec5b935dab7f879d4fa29495118d22",
    "input": {
      "name": "unnamed",
      "ti_lr": 0.001,
      "unet_lr": 0.0003,
      "n_tokens": 3,
      "lora_rank": 16,
      "resolution": 512,
      "concept_mode": "style",
      "n_sample_imgs": 4,
      "max_train_steps": 300,
      "sd_model_version": "sdxl",
      "train_batch_size": 4,
      "checkpointing_steps": 10000,
      "validation_img_size": 1024
    }
  }' \
  https://api.replicate.com/v1/predictions

To learn more, take a look at Replicate’s HTTP API reference docs.

Output

No output yet! Press "Submit" to start a prediction.

Run time and cost

This model runs on Nvidia L40S GPU hardware. We don't yet have enough runs of this model to provide performance information.

Readme

This trainer uses a single training script that is compatible with both SDXL and SD15.

The trainer has the following capabilities: - automatic image captioning using BLIP - automatic segmentation using CLIPseg - textual_inversion training of a new token to represent the concept - 3 training modes: “style” / “face” / “object” - Full finetuning or LoRa or Dora training modes are supported in the code - LoRa modules are possible for both unet and txt-encoders

The generated checkpoint files are compatible with ComfyUI and AUTO111