Unlocking the Power of Custom Models with Leonardo AI

ChatPlayground AI | Chat and compare the best AI Models in one interface, including ChatGPT-4o, Google Gemini 1.5 Pro, Claude 3.5 Sonnet, Bing Copilot, Llama 3.1, Perplexity, and Mixtral Large!

Dive into the fascinating world of Leonardo AI as we explore how to train your own custom models! This guide will walk you through the entire process, from setting up your datasets to generating stunning images, all while maximizing the potential of this innovative platform.

Table of Contents

• Read the FAQ 📚

• Trained Dataset Preview & Setting Up Images 🖼️

• Create a New Dataset & Import Images 📂

• How to Use Your Trained Model 🎨

• Standard Model Examples 📸

• Try Different Schedulers 🗓️

• Leonardo Style & Prompt Magic On 🪄

• Final Thoughts 💭

• Exploring Limitations and Opportunities 🚀

• Engagement with the Community 🌐

• FAQ ❓

Read the FAQ 📚

I can't stress enough how important it is to read the FAQ before diving into training datasets on Leonardo AI. It contains invaluable information that will save you time and effort.

Getting Started

Head over to the FAQ section and click on the article about fine-tuning. This document is a quick read and offers essential insights for your training runs.

What Makes a Good Training Run?

A good sweet spot for images is around 8 to 15, depending on what you're training. The minimum is 5 and the maximum is 30 images. For styles, aim for 20-30 images.

Fine-tuning and tweaking are necessary to perfect your model. Training at a 512x512 aspect ratio is common, but you can adjust based on your needs.

Trained Dataset Preview & Setting Up Images 🖼️

Setting up your images is the first crucial step in training your custom model. Let's walk through the process.

Choosing Your Program

You can use any image editing program you're comfortable with. For a free and user-friendly option, I recommend Canva.

Creating Your Workspace

Go to Canva.com and create an account. Click on 'Create a Design' and select 'Custom Size'. For this example, I used 768x768 pixels.

Uploading Your Images

Once your workspace is ready, click on 'Uploads' on the left side. Upload your images and drag them into the workspace to set your composition.

Create a New Dataset & Import Images 📂

Now that your images are ready, it's time to create a new dataset and import them into Leonardo AI. This is where the magic begins.

Setting Up Your Dataset

First, click on "New Dataset". A prompt will appear asking you to name your dataset. I'm naming mine "Riza Rev One". You can also add a description, but I'll leave mine blank for now.

Importing Your Images

After creating the dataset, you can import your images. Either click to upload or drag and drop your images into the designated area. I chose to drag and drop, and you'll see the images get uploaded instantly.

Training Your Model

Once your images are uploaded, click on "Train Model". You will notice options for advanced training settings, but these might not be available yet. Choose between 512x512 or 768x768 resolution; I'll go with 768x768. Select the appropriate category for your model, such as "Photography". If your model contains NSFW content, make sure to check the NSFW option.

Setting the Instance Prompt

Next, choose the training method. I'll use Stable Diffusion 2.1. Here, you'll need to set an instance prompt. This is a trigger word that activates your model. I'll use "s k s woman". Now, click "Start Training". You will receive an email once the training is complete. You can also monitor the training status by clicking "View Status".

How to Use Your Trained Model 🎨

With your model trained, you can now start generating images using your custom model. Here's how to do it.

Incorporating Your Instance Prompt

When creating a prompt, include your instance prompt word. It doesn't matter where you place it in the prompt, but I prefer placing it near the beginning. In my example, the prompt is: "beautiful s k s woman warrior wearing Skyrim dragon scale armor".

Choosing Your Model

Make sure to select your specific model. If it's not visible, go to "Select Custom Model". By default, you might be on platform models, so switch to your custom models list. Hover over your model, click "View", and then select "Generate with this model".

ChatPlayground AI | Chat and compare the best AI Models in one interface, including ChatGPT-4o, Google Gemini 1.5 Pro, Claude 3.5 Sonnet, Bing Copilot, Llama 3.1, Perplexity, and Mixtral Large!

Generating Images

Leave the negative prompts on and turn off "Prompt Magic" for now. Stick with the default settings if you're happy with them. Click "Generate" to create four images. You can then adjust the dimensions for a portrait style photo if needed.

Standard Model Examples 📸

Seeing how standard models perform is crucial to understanding the capabilities of Leonardo AI. Let's dive into some examples.

Initial Results

Initially, I noticed some minor issues with the eyes and lips. These can be fixed in post-processing. However, the overall composition was impressive.

Adjusting Steps

To improve the images, I reduced the steps to 30. This made the images less contrasty and aggressive. It’s a small tweak, but it made a big difference.

Weapon and Armor Details

Although the model struggled with weapon details, it accurately captured the armor. This is a common limitation in image generation but manageable with adjustments.

Try Different Schedulers 🗓️

Experimenting with different schedulers can significantly impact your results. Here are my insights.

Using DPM

I prefer using DPM for generating images of people. It tends to produce smoother and more realistic results.

Trying DDIM

DDIM is another scheduler worth trying. It generates images quickly and often yields satisfying results. Switching between these two can help you find the best fit for your needs.

Final Adjustments

After trying different schedulers, I found that 30 steps with DPM worked best for my scenario. The hands and other minor details were not perfect but could be easily fixed in Photoshop.

Leonardo Style & Prompt Magic On 🪄

Leonardo AI offers a unique style and prompt magic feature to enhance your image generation experience. These additions can make a significant difference in the output quality.

Leonardo Style

The Leonardo style adds a distinctive touch to your images. This includes effects like bokeh in the background, which can elevate the aesthetics of your creations.

However, some minor distortions, especially in facial features, may occur. Despite this, the style is highly adaptable and can be fine-tuned.

Prompt Magic

Turning on prompt magic can help refine the details in your images. This feature works by enhancing the prompt's influence on the generated image.

For example, experimenting with different prompts can yield varied and often improved results. Don't hesitate to tweak your prompts to achieve the desired outcome.

Final Thoughts 💭

Training custom models with Leonardo AI has been an enlightening experience. While there are areas for improvement, the platform shows great promise.

Pros and Cons

One major advantage is the ability to create highly customized models. This allows for greater creative control over your projects.

• Customizable models

• Unique Leonardo style

• Prompt magic feature

On the downside, some outputs may require additional tweaking. This is particularly true for fine details like facial features.

Future Improvements

I believe further refinement of the training datasets and prompt adjustments will yield better results. Consulting with the Leonardo AI team can also provide valuable insights.

Overall, I'm excited to see how this platform evolves. It has the potential to become a powerful tool for AI-driven image generation.

Exploring Limitations and Opportunities 🚀

As with any technology, there are both limitations and opportunities when working with Leonardo AI. Let's dive into these aspects.

Dealing with Imperfect Results

In my experience, some generated images had minor issues, particularly with facial features and hands. These imperfections are fixable but require post-processing.

For instance, using Photoshop can correct janky hands and distorted faces, making the images more polished and presentable.

Optimizer Settings

Experimenting with different settings like steps and schedulers can significantly improve results. I found that reducing the steps to 30 made images less aggressive and more natural.

• Use DPM for smoother results

• Try DDIM for quicker generations

Future Prospects

Despite some limitations, Leonardo AI offers exciting opportunities. The ability to train custom models opens up endless creative possibilities.

Continued experimentation and tweaking can lead to even better results. Engaging with the Leonardo AI community can also provide valuable insights and tips.

Engagement with the Community 🌐

Engaging with the Leonardo AI community has been incredibly beneficial. Here's how it has helped me and can help you too.

ChatPlayground AI | Chat and compare the best AI Models in one interface, including ChatGPT-4o, Google Gemini 1.5 Pro, Claude 3.5 Sonnet, Bing Copilot, Llama 3.1, Perplexity, and Mixtral Large!

Community Support

The Leonardo AI community is active and supportive. They offer valuable feedback and insights that can help improve your models and results.

For example, the team at Leonardo AI commented on my previous video, clarifying features like outpainting and the unzoom feature.

Sharing Experiences

Sharing your experiences and challenges with the community can lead to collective problem-solving. Many users face similar issues, and discussing them can lead to effective solutions.

• Join forums and discussion groups

• Share your results and get feedback

Continuous Learning

Engaging with the community keeps you updated on the latest features and best practices. This continuous learning helps you make the most out of Leonardo AI.

For instance, trying different images and getting advice from other users can lead to discovering new techniques and improvements.

FAQ ❓

Here are some frequently asked questions to help you navigate Leonardo AI more effectively.

What is the ideal number of images for training?

A good range is between 8 to 15 images for general training. For styles, aim for 20-30 images. The minimum is 5, and the maximum is 30 images.

What resolution should I use for training?

Training at a 512x512 aspect ratio is common. However, you can adjust based on your needs. Options like 768x768 are also available.

What should I include in my negative prompts?

Negative prompts help filter out unwanted elements. Include terms like "ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands" to refine your results.

How do I set up my dataset?

Click on "New Dataset" and name it. Import your images by either clicking to upload or dragging and dropping them into the designated area.

What are advanced training settings?

Advanced training settings allow for more control over the training process. Choose between resolutions like 512x512 or 768x768, and select the appropriate category for your model.

Can I monitor the training status?

Yes, you can monitor the training status by clicking "View Status". You will also receive an email once the training is complete.

 ChatPlayground AI | Chat and compare the best AI Models in one interface, including ChatGPT-4o, Google Gemini 1.5 Pro, Claude 3.5 Sonnet, Bing Copilot, Llama 3.1, Perplexity, and Mixtral Large!