Skip to main content

​
Overview

Train a custom model with your images (fast)
  • Output Type: lora
  • Base Model: sdxl
  • Estimated Cost: 0.2 * max_train_steps credits
  • Handler: replicate

​
Parameters

​
Required Parameters

​
name
string
required
Name of the LoRA (this will also be your trigger word)
  • Label: Model Name
​
lora_training_urls
array<image>
required
Images to train your visual model. (Use at least 5 imgs, 10-20 is even better.)
  • Label: Your training images
​
concept_mode
string
default:"style"
required
Type of model to train
  • Label: Training Mode
  • Options: face, object, style

​
Optional Parameters

​
sd_model_version
string
default:"sdxl"
Base Stable Diffusion model to finetune from
Almost always, we want to train and SDXL LoRa. Only when we want to make animatediff animations do we need an sd15 lora.
  • Label: Base SD Model
  • Options: sdxl, sd15
​
max_train_steps
integer
default:300
Number of training steps. Only increase this if you have lots of images (>50) otherwise the model wil overfit and will not look good.
This should be left at the default 300, unless the user specifically instructs you otherwise. If the base model is SD15, then you should change this to 800, again unless the user specifically requests something else.
  • Label: Training steps
  • Minimum: 100
  • Maximum: 1200
​
resolution
integer
default:512
Resolution your images will be resized to for training (512 is great and your generations will still look perfect when rendering at eg 1024)
Do not change this unless you are specifically instructed to. Highly recommended to train at 512 for optimal speed and at 768 for best quality.
  • Label: Resolution
  • Minimum: 512
  • Maximum: 768
​
seed
integer
default:"random"
Set random seed for reproducibility. If blank, will be set randomly.
You should only set this if you want to start from/copy the seed of a previous image. Unless one is specified, you should leave this blank!
  • Label: Seed
  • Minimum: 0
  • Maximum: 2147483647
⌘I