Sign In
home
models
images
videos
posts
articles
bounties
events
shop

Celebrity LoRA Training Guide (Consolidated) - SD 1.5

__2_'s Avatar
__2_
Apr 20, 2024
training guide
lora trainingsd1.5
Celebrity LoRA Training Guide (Consolidated) - SD 1.5

Intro

The following article details all my personal findings and recommendations for training a celebrity LoRA for SD1.5. I am also aiming to make this article a good starting point for someone just starting out without any prior knowledge.

Take this guide as a first step in getting a good workflow going and experiment to personalize it and make it better for your usage.

There are 2 workflows documented on my other article but for the sake of simplicity, I will only explain the second one here, if you are interested in learning about the OG workflow, please visit here.

Getting Started

Install the following software on your device before doing anything else. Installation guides and troubleshooting guides are available in the provided links.

[Index]

  • Dataset/Training Data - This is the collection of images that we want to the model to train on.

  • Tags - Given an image in the dataset, tags are the way we describe all the elements present in the image that we want SD to look for (and associate) so that we can guide it during training( more on this later).

Software

  • Install kohya_ss from here.

    • This software runs the training process for you and can also help with tagging the specific images. I use kohya_ss for my training but other solutions like automatic1111 training, comfyui training, sd-scripts and others exist. If you're just starting out, stick with kohya.

  • Install BooruDatasetTagManager from here.

    • This software helps you manage tags for the given dataset.

Step 1 - Dataset Preparation

This is the most important step of this process and will decide how well your model turns out. First start collecting images of the celebrity we want to train on.

We will collect 2 kinds of images

  • Face closeups

  • Body images

This is important because we want the model to be able to accurately generate the person in all kinds of images - closeups, portraits, upper body shots, full body shots etc.

Common Tips

  • Do Not include images that are low res, blurry or with bad lighting (includes colored lighting) and shadowed faces

  • Do Not pick images that include anybody else in it - no faces/bodies nothing

  • The ratio of face images : body images should be 1.5x-2x

  • Avoid adding multiple images from the same photoshoot (the model will overfit to recreate the same lighting)

Face Images

  • Search google images/bing images/reddit for celebrity-name face

  • Download every single image where the face is clearly visible and the resolution is greater than 1k on each side

  • Pick and select at least 30 images - upto 50

  • Make sure you get multiple images of every hairstyle you want to train it for e.g. blonde curly, back and green, white, pink etc.

  • If the person wears glasses, get variations of those as well - Once the images are ready, go to presize.io and upload all the images

  • Crop all face images to a size of 512x512 - without losing chunks (no cut off face, chin, hair etc) - Crop until the person's neck/shoulder (nothing more)

  • Download this zip file

  • Extract and now you will have 30-50 images out of which you need to pick 25-30

  • In the final resultant set, include as many angles as possible (left facing, from top, from below, right facing, front) and as many hairstyles

  • As much as possible avoid duplicates of the same hairstyle (unless the person consistently has the same one/ you want the character to lean towards that hairstyle)

  • After deleting duplicates and culling the comparatively bad ones, you should be left with 25-30 really high quality images that showcase the person's face very well.

Body Images

  • All the above rules apply - collect 20-40 images

  • Crop until the waist or just below, no need to include feet (unless you want that to be replicated 😉)

  • Crop to a size of 512x768

  • Include different clothes shots and ensure at least 2-3 standing shots where 90% of the body is visible

  • If you want the model to replicate a specific body part accurately (include multiple images of the same in different angles)

  • No obstructions (nothing in front of the body like a desk, handbag etc.)

  • Resolution of the original image at least 1k on both sides

  • Include multiple zoom levels - but more of upper body (head to abdomen)

  • After the selection process, pick 15-20 good images

[Index]

Trigger word - The way SD works is that it looks at an image and it's tags and associates one with the other.
For example - given enough images of a tree with the tag tree it learns to associate the general shape of a tree whenever the word tree is used in the prompt.
So if we are training a LoRA of Emma Watson and we use the words emma watson in the prompt, SD will try and generate it's internal association of that word. We don't want that - we instead want to choose a word that SD probably hasn't encountered before like 3mm4w4ts0n.
Since this word probably doesn't mean anything to SD, we can isolate our model's training when generating images.
Other guides say this doesn't matter(even a reddit post by the creators of SD) but I use it anyway.

Repeats - This is the number of times we want a single image to be repeated during training. Different guides have different recommendations on picking this number - feel free to experiment.

Folder Structure

Now you need to pick a trigger for the model and create the folders necessary for kohya to train on them.

I suggest using something similar to the following examples

  • Emma Watson - 3mm4w4ts0n

  • Scarlett Johansson - sc4rj0

  • Joe Pesci - j03p3sc1

It just has to be a combination of numbers and letters that are different enough from commonly used words in the English language.

Once all your images are ready, create the following folders and place the images in them.

For number of repeats, I suggest going with 24 for face images and 16 for body images. (More details on this in the following section).

You can create this folder anywhere, just remember where you placed it. I have a training folder setup next to the SD installation folder for consistency. Inside this training folder you can keep adding new folders for every model you want to train.

- <trigger-word>
    - images
        - <repeats>_<trigger-word> (for face images)
        - <repeats>_<trigger-word> (for body images)
    - model (for kohya to place the trained models into)

This folder structure essentially tells kohya to train the model for the trigger-word and to train the images in each folder for that many repeats.

Tagging

Now we need to start tagging our images in order to drill down on exactly what we are training in this model.

  • Start kohya_ss by navigating to it's folder and clicking on gui.bat

  • Navigate to this URL in your browser http://127.0.0.1:7860/

  • Go to Utilities -> Captioning -> WD14 Captioning

  • In the Image Folder to caption select the face images folder i.e 24_<trigger-word>

  • Add the trigger-word, in the Prefix to add to WD14 caption field with a comma

  • The comma is used to separate one tag from the next

  • There are multiple models that we can use for tagging in the `Model` dropdown - we'll stick with the default for now.

  • Once you click on Caption Images it will download/load the model and tag each image in that folder.

  • Repeat this process for body images as well

Once this process is complete, you can see that in each folder a .txt file has been created for each image in the folder with the same name. These text files now contain all the tags corresponding to each image.

  • Start BooruDatasetTagManager by navigating to it's folder and clicking on BooruDatasetTagManager.exe

  • Once open - click on File -> Load Folder and pick the face images folder.

  • Once loaded - it will have 3 tabs Dataset - Image Tags - All Tags

  • For every image in the left tab - once selected, it will show you the tags that the model generated for you in the middle pane.

  • This is where we can do some cleanup

    • Go through each image and it's corresponding tags and try to find something wrong

    • For example, if it mistagged a female celebrity as 1boy or male focus - delete the tags using the cross button on the right and save your changes.

    • If there are any tags that don't exist in the image - get rid of those as well

  • These next few points are useful if you're training the person in a specific costume/hairstyle

    • In the following case, delete all mentions of the costume/style you want to train - For example, if the person has blue eyes and you want to make sure the all generations of this model should have blue eyes - remove that tag

    • Essentially we are baking in the blue eyes into the model by not adding it to the tag list. The same goes for costumes/hairstyles etc.

    • If the character has a specific costume that you want to train on while retaining the original appearance as well

      • Include up to 25% images of that character in both face and body images

      • For example, Margot Robbie/Harley Quinn - 20 face images of Margot and 10 face images as Harley | 15 body images of Margot and 5 body images as Harley

      • For the Harley images, add an additional tag to differentiate like h4rl3yq

      • This will allow the model to generate Harley images when it sees the h4rl3yq tag and trigger-word

  • Once you have repeated this process for body images as well, we are ready for training.

Step 2 - Training

This section can get pretty complicated pretty quickly, I will try and give explanations for each thing I think you should know but will be skipping over a lot of things. For detailed explanations on everything, please refer to some videos on YouTube where they will do a much better job than I can.

In this step, we will now train our model on the dataset we just prepared.

  • Go back to kohya in your browser and navigate to the LoRA tab.

    • FYI, we are training our model in such a way that it directly outputs the newly trained data

    • we can also use Dreambooth training to add onto our existing model and later extract it out. Many guides suggest that this approach is better but I haven't tried it - so idk. If you want to try the other method - YouTube.

  • I have included the settings you will use in kohya as an attachment that you can directly load into it by going to Configuration File -> Open but I suggest you do this manually the first time around to understand the options better.

  • Under Source Model -> Model Quick Pick select runwayml/stable-diffusion-v1-5 from the dropdown

    • This is where we are selecting the base model on top of which we want to train our dataset on.

    • Additional context for picking a model (ignore for new users)

      • I always train on the base SD model because it performs very well when used with different fine-tuned models. This provides flexibility during inference.

      • BUT you can also pick your favorite model in this step to get better/different results. This isn't a strict rule and experimentation is encouraged.

      • When training with a different model, you might have to change your repeats/epochs/model strength a bit but it should work okayish if you stick with these settings as well.

      • Model specific training is too varied to cover here.

  • Navigate to the Folders tab and select the following - Image folder

  • select the images folder within the parent trigger-word folder i.e 3mm4w4ts0n -> images. This will pick up all subfolders - both our face and body image folders.

  • Output Folder - select the model folder within the parent trigger-word folder i.e 3mm4w4ts0n -> model. This is where kohya will save the trained models.

  • Model ouput name - Add trigger-word here. This is just the name of the file but it's good practice to keep it the same name as trigger-word for organisation.

  • Navigate to the Parameters tab

    • Since we are training a regular LoRA ,we will leave LoRA type to Standard

    • There are other types we can train like LyCORIS and others - if interested - YouTube

    • Train batch size - This specifies how many images it trains at one time - we can leave it at 1 for now

      • this can be increased according to your computer's specifications

      • I don't recommend going higher than 4.

    • Epoch - One epoch is basically one round of training of all our images. For example, if we have a total of 10 images and 30 repeats - 300 steps of training = 1 epoch. We want to train our model enough for it to grasp the dataset's features.

      • Pick 10 for now

      • Picking epochs and repeats is a varied topic with many preferences

      • Detailed explanations for picking epochs can be found on YouTube

    • Save every N epochs we will leave this at 1. This tells kohya to save the training state after every epoch. This will be useful in picking the best version of the model (more on this later on)

    • Caption Extension - set this to .txt since this is the extension we used when generating tags.

    • Mixed precision/Save precision - set this to bf16 if you have a RTX 30 series card, if not - leave it at fp16

    • Optimizer - set this to AdamW8bit if you have a RTX 30 series card, if not pick - AdamW

      • (Ignore for new users)

        • Prodigy optimizer automatically finds the learning rate and is supposedly better

        • but I haven't had any luck with it (the models turn out too rigid).

    • Network Rank (Dimension) / Network Alpha - set these to 32 each.

      • Setting this higher/lower will affect how much data the model stores*

      • higher isn't always better

      • some guides recommend setting the Alpha to half or lower

        • no luck with these either

    • Max Resolution - set this to 512,512

    • Enable Buckets - uncheck this

      • This option is useful if we haven't cropped our dataset to have the same sizes

      • If you want to skip the cropping (don't), then enable this

  • Click on Start Training

General Tips

  • Try to keep total steps to ~1500 or lower.

  • Play around with the ratio of repeats of face vs body to get better results.

  • Experimentation is key here - altering the number of images and repeats will help you optimize and find the best model for your needs.

Like I mentioned before, I have skipped over a lot of details/variations that can be made to the parameters above. You can play around with these values to get drastically different results. Please refer to YouTube to see how other people train their models. Detailed Explanations and reasonings for each of these parameters can also be found there.

Step 3 - Picking a model

Once training is complete, kohya will have generated some files in our model folder.

Like so trigger-word-000001.safetensors, trigger-word-000002.safetensors, ..., trigger-word-000010.safetensors.

The numbers at the end of the file name is the epoch at which they were saved.

Higher isn't always better - we need to pick the model that works best for us.

Here is where we are introduced to 2 effects that take place during training

  • Under fitting - If the model hasn't trained long enough on the dataset, using this LoRA will not reproduce the likeness well enough. You can observe this by using that specific epoch and generating images. If the likeness just isn't there - it's a sign of underfitting. We have to move to a higher epoch.

  • Over fitting - If the model has trained for too long, using the LoRA will result in the images being burnt i.e it tries to reproduce the training set too much. The output will look messy and the hair/clothing from the original training set will seep into your generations. Changing hair color in the prompt has no effect, things like that. This is a sign of overfitting, we have to move to a lower epoch.

With this workflow, I have found that a strength of 0.7-0.8 is the ideal strength at which it works really well. Now that we know what to look for, we will now select the best epoch/version of the file that we want to use.

  • The basic idea is to generate the same image using each version and compare the output.

  • Look for signs of under/over fitting in each generation.

  • Pick the version that looks the best and select that.

  • When generating images for comparison, you might want to use the model you wish to use for inference i.e your favorite model that you use for all your images.

  • If you intend this model to be widely used, you can also test on multiple models and check it's performance.

  • I use serenityv2 my @malcomrey since it gives me the best looking likeness for this workflow.

  • ADetailer, FaceDetailer is a must for generating images that aren't portraits. They help bring back detail in the face after generation.

Now that you've found the best version of this model, you can either finalize that if you're happy with the results or go back and train with different parameters/dataset to get better results.

If the model doesn't turn out as well as you'd hoped - my first step would be to get a better dataset.

The dataset will have a stronger impact than changing a few parameters.

BUT as always, experimentation is key and finding the workflow that best match you preferences is half the fun.

Step 4 - Profit?

These are just some closing thoughts

  • There are many other creators on this platform with models much better than the ones you'll get with this workflow, but we all have to start somewhere. Use this as a starting point and try to make it better. Any suggestions/improvements you can make, please share it with the rest of us.

  • This article is a complete summary of everything I know about training LoRAs for SD1.5. It's certainly not the "best" but personally, I am unsure of what I can change to make it better.

Thank you guys for all the support you've shown me.
Please feel free to comment with any suggestions/feedback on the article.

116

Comments