Home Parameters Examples

easyDream command parameters

idea

What you want the bot to dream about.

You can separate distinct ideas with +. For example, prompt:new york times is very different than prompt:new york + times:

new york times new york + times
new york times new york + times

Be aware that typos can dramatically change the output.

preset

Use a predefined set of parameters.

Some presets require features that only patrons have access to.

size

The size (width and height) in pixels of the output image. Limited to 196 for non-patrons.

High size values often cause the output to “fall apart” and lose cohesion. To create a cohesive but large image, you may want to start with a small /imagine. Then you can work your way up to a higher resolution using /envision like so:

  1. /imagine idea:planets exploding -> https://discordapp.com/.../planets_exploding.png 🠞 ID ...1
  2. /envision image:https://discordapp.com/...1/planets_exploding.png idea:planets exploding size:256 🠞 ID ...2
  3. /envision image:https://discordapp.com/...2/planets_exploding.png idea:planets exploding size:384 🠞 ID ...3
  4. /envision image:https://discordapp.com/...3/planets_exploding.png idea:planets exploding size:512

rounds

How long to dream. Higher values produce more refined images, up to a plateau point depending on the optimizer and output size.

Here is how the prompt /imagine idea:sci-fi cave changes for increasing rounds::

5 10 30 40 50

step-size

How rapidly to think. Higher step-size values allow for quality results in fewer rounds. However, unreasonably high step-size values produce lower-quality images that are less like the prompt.

Changes to step-size result in totally different results, unlike changes to rounds which can be adjusted linearly.

optimizer

Which algorithm used to ‘zero-in’ on the output result. Different optimizers produce dramatically different results, and each optimizer has a different ideal step-size value for a given rounds value.

/imagine defaults to the AdamW optimizer. I have found it to be the best for from-scratch image generation.

/envision and /meld default to the RMSprop optimizer. I have found that this optimizer is great for morphing starter images quickly, but not so great for generating images from scratch.

See this Algorithmia blog post for better information about optimizers.

detail

How many times the system will attempt to add features at each round.

The default value is a function of the output image pixel count. 196x196 images will use a value of 24; 512x512 images will use a value of 64; with other resolutions using a linear extrapolation of these values.

Using a detail too low will result in a minimalistic image which doesn’t resemble the prompts. Using a detail too high will result in a chaotic image with some features from the prompts but jumbled and incohesive.

initial-image / image / start

An image to start from, as opposed to starting from a blank canvas. Must be a URL of an image uploaded to Discord.

For /envision, this parameter is named image. For /meld, this parameter is named start.

For safety reasons, easyDream can only see images that are hosted on Discord. Trying to give easyDream a link to an image not uploaded to Discord will not work. See Use images in commands for how to upload images to Discord.

idea-image / end

An image version of an idea, as opposed to a text idea. Must be URLs of images uploaded to Discord.

The bot will identify features in this image and integrate those features into the result.

You can specify multiple idea image URLs with +. The bot will identify features in all specified idea images.

For safety reasons, easyDream can only see images that are hosted on Discord. Trying to give easyDream a link to an image not uploaded to Discord will not work. See Use images in commands for how to upload images to Discord.

mask-image

An image with an alpha (transparency) channel used to constrain each pixel in the result. Must be a URL of an image uploaded to Discord.

By default, this will use a naive approach that makes a hard cut at each round. This is not a very intelligent way to mask images. For a more intelligent method, use the mask-interval method.

Partial transparency will result in partial masking.

mask-interval

A number specifying how many rounds to wait in between applying mask-image. Higher values will apply the mask-image less frequently.

A value higher than the value for rounds will cause the mask-image to never be applied; this would mean that the system would need to wait for more time than the dream will run.

A value too low will cause the system to lose its “momentum” by applying the mask-image too frequently, resulting in less cohesive results.

augments

Which filters to apply to image generated before each round. Separate multiple augments with +. The same augment can be used multiple times.

The default augments used to create every image are: Hf + Af + Pe + Ji.

Augment options

The following augments are available:

The order of augments matters significantly. For example, using Gr+Ji is very different than using Ji+Gr.

For better results with specific prompts, you may wish to omit the Ji augment, as it can mess with the colors of the result.

Use the None augment if you don’t want to use any augments (ideal for high-fidelity images).

See this Kornia docs page for visual examples of what these filters do.

Augment string format

! See the augment-strings channel in the Discord for up-to-date augment strings.

The default probability for each specified augment to be used per round is 50%. You can change this with a :x% specifier, like this: Ji:50%

You may also set the parameters for each augment like this: Af[type:key=value].

type may be one of:

Practical augment strings

Eleuther MSE ZQ:

HF:50% + AF[i:degrees=30, f:translate=0.1, s:padding_mode=border]:80% + PE[f:distortion_scale=0.2]:40% + JI[f:hue=0.01, f:saturation=0.01]:70%

Eleuther ST:

HF:50% + AF[i:degrees=30, f:translate=0.1, s:padding_mode=border]:80% + PE[f:distortion_scale=0.2]:40% + JI[f:hue=0.01, f:saturation=0.01]:70% + GR:10%

justinjohn0306 updated:

HF:50% + SH[f:sharpness=0.3]:40% + AF[i:degrees=30, f:translate=0.1, s:padding_mode=border]:80% + PE[f:distortion_scale=0.2]:40% + JI[f:hue=0.01, f:saturation=0.01]:70%

solid-fill

A hex color code (like #651fff) to fill the empty starter canvas with. This option may have little to no effect on the output, depending on the rounds, optimizer, step-size, and augments values used; high rounds and step-sizes will completely obliterate the starter canvas in most cases, and some augments like ColorJitter or Desaturate alter colors so significantly that the starter color is irrelevant.

Additionally, this option draws a small square in the center of the image, filled with the inverse of the solid-fill: color. This is to help the bot recognize the significance of the solid-fill: value with contrast, and to help it “latch on” to the center of the image in order to grow outwards from there.