General Settings

Modified on Tue, 24 Sep at 9:22 AM

General settings to control output images can be found below the prompt boxes on the left-hand side of the Generation tab.

Image settings

The number of iterations that the current settings will be used to execute a generation.
For example, when utilizing dynamic prompts to generate more than one prompt per generation, the number of images produced will be (# of dynamic prompts) x (# of iterations). Iterations started with the same prompt will be added to the queue as part of the same batch.
Iterations can be set by changing the number next to the "Invoke" button
Tip: Iterations can be utilized to easily generate multiple options with a given prompt and chose the best one to send to the Unified Canvas for Inpainting / Outpainting

The ratio of the generated image - `Free` indicates no aspect ratio lock has been applied. Models are best when generating certain image sizes, with each model having a specific ratio.:
- SD1.5 Models typically perform best at 512x512 images, unless a model creator has instructed otherwise.
- SDXL Models typically perform best 1024x1024 images
Generating outside of these image sizes can result in images that have duplications or are distorted
The aspect ratio can be Swapped (W/H switched) or Locked using the Swap/Lock icons.
The image settings can be optimized for the aspect ratio & model selected by hitting the Optimize button.

Seed

Controls the starting noise used for generation. Disable “Random Seed” to produce identical results with the same generation settings.

Denoising Strength

Denoising strength determines how much noise is added to an image during the diffusion process.
Setting denoising strength to 0 means no noise is introduced to the input image, while denoising strength of 1 means the input image is entirely replaced by noise.
The higher the denoising strength the more an image will change.

Scale Before Processing (Advanced)

Automatically enabled by default, the Scale Before Processing automatically uses a larger resolution for generating selected areas if supported by the selected model, in order to generate higher resolution images.
You can read more about Scale Before Processing here.

The Model used for the denoising steps
There are two base model architectures that are used: SD1.5, and SDXL - These base models are then trained to specialize in generating different content. Many new models have been created using these, with different models specialized in producing certain aesthetics and content. Each base model type is separated in the Model dropdown.
Adding custom models to InvokeAI:
- Indie tier users can add custom models to use with their account
- Enterprise & Premier users can contact their organization's administrator to add custom models to the account. Each model will then need to added to individual projects - This ensures that any proprietary or custom models are secured.

Number of steps that will be performed in each generation. Higher step counts will typically create better images but will require more generation time - A good place to start is about 30-35.

Classifier Free Guidance (CFG) Scale controls how much guidance your prompt terminology will impact the generation process. Lower values can offer more flexibility/freedom to the model, while higher models can better capture the spirit of your prompt. Too high, or too low, and you'll find that quality will deteriorate. A good place to start is typically around 6-7.

Scheduler defines how to iteratively add noise to an image or how to update a sample based on a model's output
There are many different Schedulers, each producing different results
Two popular ones to start with are Euler and DPM++ 2M Karras. Modifying the step count will affect the image result based on the scheduler used. With these two schedulers, Euler can generally be used with a lower step count, while DPM++ 2M Karras will require a higher step count.