What is a “good training dataset”?A good training dataset for a finetuned custom model will vary greatly, depending on your desired output. To curate a good training dataset for a generator, there are a few key steps you can follow:
First, identify the specific style or subject that the generator will be trained with.
Next, gather a dataset that is relevant to the style or subject.
The dataset can have 5 images min and 100 images max.
The dataset should be large enough to allow the generator to learn the relevant patterPopular
What is the time duration for training a generator?Currently, it depends on the size of the training dataset - small datasets of under 10 images can take closer to 20 min, large datasets can take a few hours. We do not recommend datasets larger than 20, as they will take a long time and are more likely to overfit if you are not experienced with making custom generators.
You can learn about training parameters here.Some readers
What is the learning rate and how to use it?In machine learning, the learning rate is a hyperparameter that determines the step size at which the optimizer makes updates to the model parameters during training. A higher learning rate can result in faster training, but it can also make the model more prone to overfitting, as it may not fully capture the underlying patterns in the data. A lower learning rate, on the other hand, can help the model better capture the patterns in the data, but it may also result in slower training.
The numberSome readers
What are training steps?Training steps refer to to the number of steps the training program takes as it learns the new concepts you have provided in your dataset.
Although it may seem as though more training steps should be better, this is not always the case. Training steps should be relative to the size of your dataset (100-200 steps per image) and your learning rate. Typically, you will want a lower leSome readers
What is the text encoder?Text encoding is the process used to more accurately train the CLIP program to more accurately interpret the image dataset. The CLIP program was originally trained on billions of images and the tokens trained get their words and phrases from metatags.
Fine-tuning the text encoder seems to produce the best results, especially with faces.
It generates more true to data imagesSome readers
What is a "generator"?A generator is a finetune model using the Dreambooth program, which allows users to custom train style and object settings. This creates more heavily guided image outputs.
Most AI programs on the market allow their members to achieve a very nuanced level of detail and focus through prompting with words and images. However, they do not learn live from user input. When you train a geFew readers
What are training classes? How do they impact future image outputs?Regularization or Training images are a way to improve the accuracy of neural networks by introducing noise to a set of data points (such as pixel or vector images). This helps prevent overfitting, where a neural network becomes too precise with a specific training set, but unable to make accurate predictions on new data. Using regularization images during training can help a neural network detect more generalFew readers
How do I get my output images to look the same as the ones in my dataset?There are many reasons why your images may not look the same as your dataset. These can range from issues with the size of your dataset to issues with dataset curation. You may also need to add assistive prompts. It is also possible that a different regularization class would be a better fit for what you're trying to create.
Some good general rules of thumb are to make sure that whatever dataset you are using reflects your desired outcome. If you are attempting to create a more general model, tFew readers
How can I modify a generator, if I realize it’s not generating the outcome I’m expecting? Shall I remove/add pictures?If you find that your outputs are not what you expected, you may need to make adjustments to your training dataset. The best way to do this is to go to your existing generator through the Generator tab and Copy Generator. Now add or remove images as needed. You can also use this tab to adjust other training parameters.
To fix this issue and improve a training dataset, try remFew readers
Do training images always need to be 512 by 512 pixels?To properly train a dataset, all images must be resized to squared. They will be adjusted to 512x512 by our software. However, you can use images outside these dimensions, which need to be cropped/resized.
You can resize and crop your images directly within the Scenario web app, before initiating a training. If you have an image that is horizontal or vertical, simply adjust your square crop so that it is larger than the longest side of the image, and clip out the background using our backgrounFew readers
Can I download a ckpt file?Currently, during the alpha testing phase, CKPT files are not available for download. However, we will offer download options in the near future (Q1 2023).Few readers
How do I get as much detail as the images I used to build my dataset?If you are struggling to get the level of detail in your dataset, here are a few things you can check:
Guidance: Try increasing your guidance to 9 or 10.
Regularization Class: Perhaps the regularization class isn’t quite optimal, so consider retraining on a new one.
Sampling Steps: It is possible you are using too few sampling steps - or too many. Try to keep your steps between 30-50. In cases of finelinework, consider up to 80.
Dataset images: It is possible there are grainy images inFew readers
Should I remove the background of the images in the training dataset?It depends entirely on what you’re training. For broad concept style training we haven’t seen a benefit.
For specific subject assets, such as characters and objects, sometimes the output is better with a background removal, and it’s worth testing out. If all of your images have similar backgrounds, it worth removing them to avoid the AI learning your background as a part of the style.
Watch our walk-through video on how to remove backgrounds from your training images. (https://scenario.crispFew readers
How many images should I select to train a generator?It is important to have a variety of unique images in your dataset. You must have at least 5 and no more than 100 images. Here are some general guidelines:
If the subject or style you are training is fairly commonplace and likely already exists in the dataset, you do not need many images. 6-10 is fine.
For complex subjects like avatars, vehicles, and animals, try to provide between 15-20 images. For simple subjects like plants, garments, and weapons, aim for 8-20 images.
Style training uFew readers
What token should I pick for a unguided or custom class?A token refers to the unique word or phrase used to identify a concept in your generator and prompt. When training a new generator, it is best to pick a token that is not a real word, as the generator may already have an association with existing words. One example would be, if you were training a dog, you may choose to use the word dgg3.Few readers
What is an unguided generator?A generator which is "unguided" is not be trained on a set regularization class.
Unguided classes are recommended only for experienced users and specific use cases. In many cases, working without a class can lead to overfitting (https://scenario.crisp.help/en/article/video-tutorial-what-is-ovFew readers
What are custom classes?Scenario has a number of pre-established training classes for users to work with. However, there may be situations where a user finds that they cannot identify a suitable regularization class for their needs. (For example, humanoid characters that fall outside the normal expectations of a concept such as 'person')
In this case, advanced users may choose to use their own regularization class. When a customFew readers
Can I leave the training page or image generation page while it’s still working?Yes, you can close any active training or image generation page without losing progress. Your work will continue in the background.Few readers