• Why is the subject of my image repeating when I make it a larger size?

  During Alpha, images that are scaled beyond 512 pixels in both directions will repeat subjects. Please see the images below for an example. In the future, upscaling capabilities will resolve this issue. This image output is set to the standard size of 512x512 pixels. However, this image has been created at a larger size. The way the Stable Diffusion program functions causes it to fill in the space you create for it. With a custom generator in particular, it tries to give you the imagery you'vePopular

• I used five images and now my outputs look nothing like the original images. Why?

  While five images is the minimum, it may cause there to be too few training steps. It is recommended that you add more images to give the program enough time to train. If you want to learn more about creating a custom generator, take a walk through of one of your test generators!Few readers

• I want my image to look exactly the same every time - how do I achieve this?

  Diffusion model image generators are created with the intention of inferring and abstracting on a multitude of visual subjects. This means, by default, images are unique. However, there are situations where a user may not want as much variety in their output. They can achieve a more narrow range of output by using seeds, using img2img on a low setting, or intentionally overfitting. Intentional overfitting is the most risky approach, as overfit images have a tendency to create noise and poor outFew readers

• Is it possible to generate assets with a transparent background?

  Yes it is possible to generate images with a transparent background! Even if your training images have a transparent background, your output images will have a color added by the Stable Diffusion program. However, you can remove them using our in-app system for output images.Few readers

• What is a Seed?

  The “seed“ parameter in stable diffusion is a number used to initialize the generation process. The seed parameter can be set to a specific value or left to be randomly generated by the model. By controlling the seed parameter, it is possible to generate reproducible images, experiment with different model parameters, or create variations of a prompt. Using the same seed value and prompt can produce the same image, while changing the seed or prompt can result in different outputs. The text pFew readers

• What’s the best way to test my output images?

  After you have finished training a model, we recommend creating 10 instance images at 512x512 pixels and 50 steps. This will show you what the training retained and whether your model is overfit. You should then test your generator with a few simple prompts. If you notice that your images are including details from your original dataset that you did not want, you may be able to prompt them out. However, if testing the generator with prompts does not produce the desired results, you will need toFew readers

• How many output images is a good amount?

  It can range, but it’s best not to generate more than 10 at a time while the program is free. We recommend 4 at a time while testing prompts, as results will execute more quickly. Once you have identified an ideal prompt and settings, you may wish to execute more.Few readers

• My images started populating, but they just look like noise, is something wrong?

  There is nothing wrong with your images. They just take a little time to generate. The noise that you see at the beginning is the start of the diffusion process. Depending on the size and number of steps, it may take 15 seconds to a few minutes to complete. You can navigate away from the page and check back later.Few readers