delta-lock/ponynai3

The VAE to use

The mask for inpainting, white areas will be modified and black preserved (will be scaled then cropped to the set width and height)

The seed used when generating, set to -1 to use a random seed

The image for image to image or as the base for inpainting (will be scaled then cropped to the set width and height)

The LoRAs to use, must be either a string with format "URL:Strength,URL:Strength,..." (strength is optional, default to 1), or a JSON list dumped as a string containing key "url" (required), "strength" (optional, default to 1), and "civitai_token" (optional, for downloading from CivitAI) (NOTICE: Will download the weights, might take a while if the LoRAs are huge or the download is slow, WILL CHARGE WHEN DOWNLOADING)

The model to use

The steps when generating

The width of the image

The height of the image

The prompt, uses Compel weighting syntax

How much noise to add (for image to image and inpainting only, larger value indicates more noise added to the input image)

CFG scale defines how much attention the model pays to the prompt when generating, set to 1 to disable

How many CLIP layers to skip, 1 is actually no skip, this is the behavior in A1111 so it's aligned to it

PAG scale is similar to CFG but it literally makes the result better, it's compatible with CFG too, set to 0 to disable

The scheduler to use

Number of images to generate (1-4), note if you set this to 4, some high resolution gens might fail because of not enough VRAM

The factor to blur the inpainting mask for smoother transition between masked and unmasked

The negative prompt (for things you don't want), uses Compel weighting syntax

The amount to rescale CFG generated noise to avoid generating overexposed images, set to 0 or 1 to disable

Prepend preprompt (Prompt: "score_9, score_8_up, score_7_up, " | Negative Prompt: "score_4, score_3, score_2, score_1, worst quality, bad hands, bad feet, ")