Image control settings

Init image related:

init_image

Default: None

Path or URL of the initial image that is used to jump start the model. having an init_image can be very beneficial to the output: it can concentrate the model to a central focus, which it typically has difficulty in.

init_image

init_image_alpha and init_noise

Default: 200 and pixels

The two goes hand in hand. When using an init image, the noise is overlaid on the init image, making it noisier, and the color less intense, as shown in the last example, where the noise is none, the image becomes "brightened". Potentially, decreasing the init_noise may allow for the model to have a wider range of potential outcomes, and less prone to get in local minima, but this is just speculation. However, if repeatedly generating images and putting them back in, not setting init_image_alpha to 255 may cause the image to fade out.

init_noise can be one of pixels , gradient , snow , none. none cannot be used without an init_image and it would "brighten" the image if init_image_alpha isn't 255.

init_noise

Image Prompts and Target Images

image_prompts, image_prompt_weight, and image_prompt_shuffle

Default: [], None, and False

Image prompts work the same as text prompts: they are encoded as prompts and generate losses the same way a text prompt does.

Image prompts unfortunately still cannot use the prompt:weight:stop logic, despite having prompt in its name. it also cannot use commas to separate different images, it uses an odd package called braceexpand to generate a list of filenames. and example may be "a{1..3}b" is expanded as ['a1b', 'a2b', 'a3b'], while "a{1,3}b" is expanded as ['a1b', 'a3b']. This should be changed.

Image prompt shuffle is just for resetting the transformed images every time cutouts is made, if False every image prompt will have the same cutout, if True different image prompts would have different cutouts.

Usage

image_prompts = "pimgs{1..10}.png"
image_prompt_weight = 1.0
image_prompt_shuffle = True

image_prompts

target_images

Default: None

This is functionally similar to image_prompts. It differs in two places. target_images allows the prompt:weight:stop and | separation logic, and it is not made cutouts every time it is compared: it's embedding is generated during init, like text prompts, and not during ascent.

Usage

target_images = "fish.png:3:10 | anime.png:-1:-0.9"

Spot Prompts

spot_prompts, spot_prompts_off, and spot_file

Default: [], [], and None

Spot prompts is a misleading name: a more accurate name may be masked prompts. With this option, a mask image can be used to specify which parts of the image will be seen by which prompts. The black parts of the spot_file will be shown to the prompts specified in spot_prompts, while the white parts will be shown to spot_prompts_off. Below is a basic visual demonstration of how this is utilized.

Spot prompts also has the same input logic as normal prompts.

spot_prompts

Overlay

overlay_image

The image that is overlaid on top of the generated output, this is done every overlay_every iterations. When an init image is just not enough, and you need that power and control and specification… my impression is that this is best used for color location specification. Has big potential, but is kinda jank.

overlay_alpha

This is important because if it's too high, then the image is just going to "reset" every time it hits overlay_every. By default its behavior is same as setting it as 255, which is bad, so always remember to set this to something reasonable when using overlay. Somewhere around 50-100 is good for the default overlay_every=10.

overlay_every

The name says it all: every overlay_every number of iterations, it does an overlay.

overlay_offset

Seems not important. Not an image offset, but a iteration offset. For overlay_every=10, having offset at 5 makes it overlay at 5,15,25 etc. iterations.

overlay_image