Overview
We are bringing the long awaited feature for creating your own Elements (LoRAs), offering greater flexibility and control to users when it comes to generating images that require specific content or styles.
Features
A handy feature for training Elements that can be used for certain situations such as applying a certain style or generating a specific character, Element Training will revamp how you create unique imagery.
Support for a variety of models: Element Training will support creation of Elements utilizing various XL models. You will be able to use the created Elements with other XL models, much like the platform Elements.
Higher resolution training: SDXL Elements are trained on a resolution of 1024×1024 pixels by default.
Ease of Use: Element Training has the best optimized default settings that works well for most training cases.
Advanced Settings: Settings such as Epoch count and Learning Rate can be set by the user if required.
Benefits
Element Training brings more flexibility to Image Generation, allowing users to generate images in certain styles or featuring a character or product by using their trained Elements.
Lighting or visual style: Generate images in your own visual style or apply an specific cinematic look to your image to enhance drama by training your own Element on your lighting or visual style.
Characters: Generate images featuring your character with high levels of consistency using Element Training, allowing you to feature your character down to details such as costumes and accessories with better consistency. You can even generate them in a completely different visual style by pairing the Character Element with another fine-tuned model or using additional Elements.
Product and Fashion: Feature your product or fashion item in visually appealing imagery with better consistency by training an Element using several images of the item.
How to use
1. Dataset Assembly
In order to start training your model you would first have to assemble your own dataset.
Datasets can be for specific things such as:
Concepts - pastel looking cheeseburgers, artificial planets made of metal etc.
Styles - Van Gogh's impressionist style, your very own artistic style etc.
Characters / Objects - Your very own character or a specific product/clothing etc.
Key Considerations
Start with a well-curated image dataset that mirrors the diversity of your chosen theme, sticking to a consistent size ratio of 1024×1024px. Images are optimally cropped behind the scenes during the training process if they are not in a square aspect ratio.
To avoid overfitting – incorporate a varied dataset with up to 40 high-quality, images to teach the model a wide array of scenarios. (Overfitting causes the Element to recall / recreate its trained data instead of being flexible enough to adhere to the given prompt).
Consistency in style, format, and aspect ratio are paramount for recognition efficiency while introducing variation within these constraints encourages the model to creatively reapply learned elements in novel contexts. (Striking the right balance between variation and consistency may require trial and error). For Element training there is more flexibility in regards to the aspect ratio, though it is best practice to ensure consistency if possible.
Ensure images in the dataset are free of unwanted visual elements as much as possible.
Consistency - character position, style and image composition.
Variation - characters themselves and their clothes.
Bad Dataset ❌
Good Dataset ✅
2. Dataset Creation:
From the home page, navigate to Training & Datasets:
Click on ‘Create New Dataset’ or 'New Dataset' to create your dataset.
Name Your Dataset and add an description if needed.
Add Images to Your Dataset. Images can be either uploaded or selected from previously generated images on the platform.
Double check the images align with your theme or subject of interest and remove any unwanted images that may cause conflict with the resulting model.
3. Training:
Training: Once you have finalized the selection of images for your Element within a dataset, click on Train Model. You will then be presented with a variety of options.
Model Type: Select Element (LoRA)
Description: Enter a description if required.
Resolution and Model Type: Select your preferred Training Resolution (Elements are 1024×1024px only) and the Base Model.
Category: Select a category, ensure this is adjusted when training an Element. Each category will utilize different methods of training for the best results.
Trigger On: Add an Trigger word. This word will be automatically included by default when the Element is active to trigger it. The Trigger word should ideally be abstract such as markxperson if you are training on something that is not in the base model to avoid any misinterpretation by the AI.
Training Start: Click on the purple Start Training button to initiate the training. An email notification will be sent to you upon completion and you may check on the job status though the Job Status tab within the Datasets & Training page.
Please note that the training process may take from several minutes to a few hours depending on the number of images and training settings.
⚠️ Important notice:
・Images should optimally be at 1024×1024 for best results. For images that are in other aspect ratios, smart cropping will be applied.
・It is highly recommended that you train an Element instead of a Finetuned model. Elements offer much more flexibility in terms of resolution and aesthetics in comparison to our legacy finetuned model training.
・It is important to note that making changes to a dataset or deleting it after having trained a model / Element on it will not affect the existing model / Element.
・Avoid using Lightning XL models as the base model for Element training.
4. Using your newly trained Element
Navigate to the Image Creation page.
Click the button in the left side of the prompt input then press View More under the Elements section.
Click on the Your Elements tab in the Elements popup and select the Element that you wish to use.
Adjust the Strength if the Element is not affecting the outputs enough.
Legacy Mode
Navigate to the Image Creation page.
On the top right corner, click on Legacy Mode to switch to the older Image Generation tool.
Click on the ⚛️ Add Elements button and then click Your Elements at the top of the Elements popup.
Adjust the Strength if the Element is not affecting the outputs enough.
Advanced Settings (Element Training only)
The default settings of our Element (LoRA) training are tweaked to the best optimal settings. However there are times that users may prefer tweaking the settings to their preferences. This should however be only done by users with experience in training LoRAs. We recommend strictly using the default settings as much as possible.
Epoch: Determines how many times each image of the dataset gets processed during training. Each epoch allows the model to minimize errors through learning and adjustments. Note that the higher the epoch amount, the longer the duration of training. The maximum possible epoch setting is 250.
Train Text Encoder: This setting involves training the part of the model that interprets your prompts. By training the text encoder, it helps the model better understand prompts and visual descriptions. This may not be necessary if a concept is already well understood by the text encoder.
Learning Rate: This determines how high or low the learning rate is. The higher the learning rate, the faster the training but this may lead to quality issues. With a very low learning rate, training is more stable but slower, but will likely lead to convergence issues which may lead to incoherent images such as people with malformed limbs or even just image noise. Basically, keep it as low as possible but not too low that quality issues start occurring.
Frequently Asked Questions
Will an Element automatically update if I add / remove image(s) from it's dataset?
Changes to a dataset will not affect any existing Elements trained on it. This would mean any modifications to the dataset would require an entirely new Element to be trained on it.