How to train SDXL LoRA models

Updated Categorized as Tutorial Tagged , 18 Comments on How to train SDXL LoRA models
train lora sdxl model

You can train your own SDXL LoRA model with the Google Colab Notebook created by this site. The following instruction trains a LoRA model for a person’s face.


You must have a Google Colab Plus subscription to use this training notebook.

Download the Easy LoRA Trainer SDXL and sample training images below. You can use the sample images to go through the whole process before using your own.

Option 1: Become a member

If you are a member of this site, access the notebook and training images below.

To read this content, become a member of this site

Already a member? Log in here.

Option 2: Purchase the notebook

Alternatively, you can purchase the training notebook and sample training images in the store here.

Structured Stable Diffusion courses
Become a Stable Diffusion Pro step-by-step.

Step 1: Prepare training images

For the best result, crop the images to 1024×1024 pixels. However, this is not strictly necessary because the training script supports bucketing.

See my training examples below.

Step 2: Come up with a good triggering keyword

For training SDXL LoRA models, it is better NOT to use a rare token for the triggering keyword but something that resembles your subject.

Since we are training a woman’s face, we need to find someone in the SDXL model who looks like her. That would increase your chance of success.

Let’s try prompting the SDXL Base 1.0 model.

photo of jane

Nah, Jane won’t be a good triggering keyword because it generates black-and-white images with a woman who doesn’t look like ours.

Let’s try another name.

photo of emma

The keyword “emma” generates color pictures of a young woman who resembles our subject. (Likely thanks to Emma Watson) Let’s use “emma” as our triggering keyword. Note you have to stick with a single word.

Step 3: Review the training settings

Open the Easy LoRA trainer SDXL notebook.

Project name

This is the folder that will be created in your Google Drive. Use a different project folder for each training.

Image Repeats

The Image repeats is how many times the training images are repeated in each training epoch. Keep this at 1 for the default training workflow.

Number of Epochs

The number of epochs is the number of training rounds. Increase this number to increase training cycles.

This is the main parameter to adjust how much want to train the model. Increase to train more. Decrease to train less.

Learning rate

How big a step for each model update.

A larger value trains faster and requires fewer training epochs.

But a learning rate too large may cause error or bad results.

Triggering keyword

The triggering keyword is the token associated with your subject. You need to use this keyword in the prompt.

Lora name

The Lora name is the name of your LoRA file. In AUTOMATIC1111, It looks like <lora:emma_XL:1> when you use the LoRA.

Lora output path

The LoRA file will be saved in this location in your Google Drive.

Skip Image upload

Select this option if you want to reuse the previously uploaded images. This is useful for retraining a LoRA model with a different setting or revising the captions.

Step 4: Start training

Start the training by clicking the play button on the left of the settings.

It will ask for your permission to access your Google Drive. You must accept the connection to save the final LoRA model in your Google Drive. (There’s no good way to download the model except by saving it in your Google Drive.)

A button will appear that allows you to upload the training images. Click Choose Files and select your training images. (The images, not the zip file.)

It will take a while to complete running. It will

  • Set up the training software
  • Generate captions for your images. You can find them in the project folder in your Google Drive. They are the .txt files with the same name as your images.
  • Train the LoRA model.

It may prompt for restarting the runtime. Click Cancel.

Monitor the training progress with the printouts. It is done when it shows 100% steps.

You can rerun the training cell without disconnecting and reconnecting the notebook.

Disconnect the notebook when you are done. Otherwise, it will continue to consume your compute credit.

Using the SDXL LoRA model

Your model is saved in your Google Drive in the AI_PICS > Lora folder. It is ready to use with the Stable Diffusion Colab Notebook.

Alternatively, download and install the LoRA model locally on your machine.

For AUTOMATIC1111, put the LoRA model in stable-diffusoin-webui > models > Lora.

To use the LoRA model in AUTOMATIC1111, you first need to select an SDXL checkpoint model.

Use the prompt with the LoRA:

photo of emma <Lora:emma_XL:1>

Don’t forget to set the image size to 1024×1024.

Now, we get our subject!

Note: You must apply the LoRA AND use the triggering keyword emma to get the effect.

Testing the LoRA weight (e.g. <lora:emma_XL:weight>) when using the LoRA. Sometimes, 1 is not the optimal value.

Tips for successful training

The default setting is good for training a realistic face. You may need to tweak the settings for your training.

Below are some tips for tweaking.

  • The quality of the training images is more important than the quantity. The training images should show clear faces if you are training a face. You can get good training with as few as five images.
  • The images should have a diverse background.
  • Try adjusting the number of epochs to increase or decrease the training. It is possible to overtrain a model.


Training a LoRA model requires patience and experimentation. You should treat the default parameters as a starting point.

Observe the result and change the settings one at a time. Observe the result with the same seeds. Generate multiple images to draw conclusion.

Just like a good old scientist would do.

A systematic approach may take longer. But in the end, the knowledge and intuition you gain will make you a better trainer.

Undercook and overcook

You should not undercook (training too little) or overcook (training too much) your model. You should aim at training the just right amount and stop.

A basic way to do that is to change the number of epochs.

Just right

Below are from a model trained with the default value of 50 epochs. Her face shows up nicely.


Training with 25 epochs is not enough. The face doesn’t show.


100 epochs overcook the model. The pose and the face lost diversity. Her face doesn’t look natural.


I got ModuleNotFoundError: No module named ‘torch._custom_ops’

You likely have restarted the runtime when prompted. Click cancel when prompted to restart the runtime.


By Andrew

Andrew is an experienced engineer with a specialization in Machine Learning and Artificial Intelligence. He is passionate about programming, art, photography, and education. He has a Ph.D. in engineering.


  1. I was using V100 high ram (also tried T4 high ram as well). I got the same error reported by other users. First I was asked to restart the runtime:

    WARNING: The following packages were previously imported in this runtime:
    You must restart the runtime in order to use newly installed versions.

    After restarting it will hit the the module not found error:

    ModuleNotFoundError: No module named ‘torch._custom_ops’”

    1. Hi, I can reproduce this error. It is because the runtime was restarted.
      You must not restart the runtime. Click “Cancel” and it will work.

  2. Any tips for training a face LoRA for use in Pony diffusion? I’m stuck at step 2 because that checkpoint is trained to the point that you can’t really find a celebrity lookalike (at least not in my experience).

    1. It could be difficult if the checkpoint is over-trained to produce a certain style.

      This likely not giving good results, but you can try training a lora on SDXL base and apply it to pony diffusion. Adjust the lora weight to see if you get what you want.

      Otherwise, you really need to pick a keyword and start training, even if it is not close.

  3. Hi in running “Easy_Lora_Trainer_SDA_SDXL_v1.1.ipynb”.

    I’m getting the following errors, after uploading the images:

    “Already installed.
    env: PYTHONPATH=/env/python:/content/kohya_ss
    python3: can’t open file ‘/content/kohya_ss/finetune/’: [Errno 2] No such file or directory
    Traceback (most recent call last):
    File “/usr/local/bin/accelerate”, line 5, in
    from accelerate.commands.accelerate_cli import main
    File “/usr/local/lib/python3.10/dist-packages/accelerate/commands/”, line 19, in
    from accelerate.commands.estimate import estimate_command_parser
    File “/usr/local/lib/python3.10/dist-packages/accelerate/commands/”, line 34, in
    import timm
    File “/usr/local/lib/python3.10/dist-packages/timm/”, line 2, in
    from .models import create_model, list_models, is_model, list_modules, model_entrypoint, \
    File “/usr/local/lib/python3.10/dist-packages/timm/models/”, line 1, in
    from .beit import *
    File “/usr/local/lib/python3.10/dist-packages/timm/models/”, line 49, in
    File “/usr/local/lib/python3.10/dist-packages/timm/data/”, line 5, in
    from .dataset import ImageDataset, IterableImageDataset, AugMixDataset
    File “/usr/local/lib/python3.10/dist-packages/timm/data/”, line 12, in
    from .parsers import create_parser
    File “/usr/local/lib/python3.10/dist-packages/timm/data/parsers/”, line 1, in
    from .parser_factory import create_parser
    File “/usr/local/lib/python3.10/dist-packages/timm/data/parsers/”, line 3, in
    from .parser_image_folder import ParserImageFolder
    File “/usr/local/lib/python3.10/dist-packages/timm/data/parsers/”, line 11, in
    from timm.utils.misc import natural_key
    File “/usr/local/lib/python3.10/dist-packages/timm/utils/”, line 2, in
    from .checkpoint_saver import CheckpointSaver
    File “/usr/local/lib/python3.10/dist-packages/timm/utils/”, line 15, in
    from .model import unwrap_model, get_state_dict
    File “/usr/local/lib/python3.10/dist-packages/timm/utils/”, line 8, in
    from torchvision.ops.misc import FrozenBatchNorm2d
    File “/usr/local/lib/python3.10/dist-packages/torchvision/”, line 6, in
    from torchvision import _meta_registrations, datasets, io, models, ops, transforms, utils
    File “/usr/local/lib/python3.10/dist-packages/torchvision/”, line 4, in
    import torch._custom_ops
    ModuleNotFoundError: No module named ‘torch._custom_ops'”

    1. I tried to ignore and continue
      “ERROR: pip’s dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
      torchaudio 2.2.1+cu121 requires torch==2.2.1, but you have torch 2.0.1 which is incompatible.
      torchtext 0.17.1 requires torch==2.2.1, but you have torch 2.0.1 which is incompatible.
      torchvision 0.17.1+cu121 requires torch==2.2.1, but you have torch 2.0.1 which is incompatible.
      Successfully installed GitPython-3.1.42 antlr4-python3-runtime-4.9.3 docker-pycreds-0.4.0 gitdb-4.0.11 lit-18.1.2 mypy-extensions-1.0.0 nvidia-cublas-cu11- nvidia-cuda-cupti-cu11-11.7.101 nvidia-cuda-nvrtc-cu11-11.7.99 nvidia-cuda-runtime-cu11-11.7.99 nvidia-cudnn-cu11- nvidia-cufft-cu11- nvidia-curand-cu11- nvidia-cusolver-cu11- nvidia-cusparse-cu11- nvidia-nccl-cu11-2.14.3 nvidia-nvtx-cu11-11.7.91 omegaconf-2.3.0 pathtools-0.1.2 pyre-extensions-0.0.29 sentry-sdk-1.44.0 setproctitle-1.3.3 smmap-5.0.1 tk-0.1.0 tokenizers-0.13.3 torch-2.0.1 transformers-4.30.2 triton-2.0.0 typing-inspect-0.9.0 voluptuous-0.13.1 wandb-0.15.0 xformers-0.0.20
      WARNING: The following packages were previously imported in this runtime:
      You must restart the runtime in order to use newly installed versions.

      1. Are you using the latest version of the notebook? Please follow the link in this article or Gumroad. I just tested the notebook and it is working correctly.

        You can also try disconnect and reconnect to start over.

        1. Thanks for the response. Yes I’m using the notebook for members.

          It seems to only be happening when using the V100.

          But I’m still facing issues that the LoRA isn’t actually making a difference. I’ve done it twice now and followed the instructions.

        2. Ok I figured out that in A1111 the lora’s don’t seem to work.

          But in Fooocus it works well. I’d update the instructions to show that. 🙂

  4. Your tutorials are great (learned how to install automatic1111 from here). So mucho thanks.
    This tutorial seems great for training a consistent face (although I guess Dreambooth is better for some reason), but what about if I want the character’s “figure” to be consistent?
    As an example, yes I want a strong Scandinavian face for my model that’s always the same (Valkyrie), but I also want the same powerful figure (wide shoulders, strong musculature, height, etc.) for all the generated figures (its the same person in a novel I’m writing).
    Can that also be achieved? Is that why Dreambooth might be better?

  5. >> You should not undercook (training too much) or overcook (training too little) your model. You should aim at training the just right amount and stop.

    Shouldn’t it be “undercook (training too little)” and “overcook (training too much)”?

    Another thing is, instead of finding someone look like the trained subject, I do the following:

    1) Annotate each image in the dataset with captions, using a rare token as before. However, better results might be achievable if the keyword could generate someone resembling the trained subject. I have not personally tested this theory. NOTE: The caption should start with the token + class. If “emma” is chosen as the keyword, the caption should begin with “emma woman.”

    2) Utilize the same captions to create regularization images using the DDIM sampler with the model that you’re training, and a “FIXED” seed (ex, 1234).

    3) Assign the same names to both the dataset images and the regularization images— for instance, if the first dataset image is named “00,” then the regularization image generated using that image’s caption should also be named “00.” This approach implies that the number of regularization images needed will be the same as the number of dataset images.

    4) Ensure that the caption files are not placed in the regularization directory.

    5) Proceed with training using the same seed employed in Step 2.

  6. It did not work (training any model has not worked for me yet).
    Here are the last few lines from the 30 odd minute run of the notebook
    diffusion_pytorch_model.fp16.safetensors: 100% 5.14G/5.14G [01:07<00:00, 76.3MB/s]
    Fetching 15 files: 100% 15/15 [01:07<00:00, 4.53s/it]
    Loading pipeline components…: 83% 5/6 [00:10<00:01, 1.94s/it]Traceback (most recent call last):
    File "/usr/local/bin/accelerate", line 8, in
    File “/usr/local/lib/python3.10/dist-packages/accelerate/commands/”, line 46, in main
    File “/usr/local/lib/python3.10/dist-packages/accelerate/commands/”, line 1057, in launch_command
    File “/usr/local/lib/python3.10/dist-packages/accelerate/commands/”, line 673, in simple_launcher
    raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)
    subprocess.CalledProcessError: Command ‘[‘/usr/bin/python3’, ‘./’, ‘–enable_bucket’, ‘–min_bucket_reso=256’, ‘–max_bucket_reso=2048’, ‘–pretrained_model_name_or_path=stabilityai/stable-diffusion-xl-base-1.0’, ‘–train_data_dir=/content/drive/MyDrive/AI_PICS/training/emma’, ‘–resolution=1024,1024’, ‘–output_dir=/content/drive/MyDrive/AI_PICS/Lora’, ‘–network_alpha=32’, ‘–save_model_as=safetensors’, ‘–network_module=networks.lora’, ‘–text_encoder_lr=3e-05’, ‘–unet_lr=3e-05’, ‘–network_dim=32’, ‘–output_name=emma_XL’, ‘–lr_scheduler_num_cycles=50’, ‘–no_half_vae’, ‘–learning_rate=3e-05’, ‘–lr_scheduler=constant’, ‘–train_batch_size=3’, ‘–max_train_steps=99999’, ‘–save_every_n_epochs=5000’, ‘–mixed_precision=fp16’, ‘–save_precision=fp16’, ‘–caption_extension=.txt’, ‘–cache_latents’, ‘–cache_latents_to_disk’, ‘–optimizer_type=AdamW’, ‘–max_train_epochs=50’, ‘–max_data_loader_n_workers=0’, ‘–caption_dropout_rate=0.05’, ‘–bucket_reso_steps=64’, ‘–min_snr_gamma=5’, ‘–gradient_checkpointing’, ‘–xformers’, ‘–noise_offset=0.0′]’ died with .

      1. No, do I need one for this exercise? Is it a monthly amount or one time deal, just to train the model

        1. Yes, training SDXL model is memory intensive. You need to enable the high ram setting to avoid out-of-memory issue.

          It is monthly $10 per month I think.

Leave a comment

Your email address will not be published. Required fields are marked *