2.1 KiB
Performs an advanced CLIP interrogation technique on the initial image known as Hard Prompts Made Easy.
Be aware that this technique is essentially a training routine and will significantly lengthen your inference time, at least on the default settings. On a Geforce 3090, it appears to take around 1-2 minutes.
By default, this shortcode is only compatible with SD 1.5 models. If you wish to use it with SD 2.1 or Midjourney, please set clip_model to ViT-H-14 and clip_pretrain to laion2b_s32b_b79k. It does work surprisingly well with Midjourney.
Supports the optional image_path argument. This is a path to file(s) or a directory to use as the initial image. If not provided, it will default to the initial image in your img2img tab. Note: you can supply multiple paths delimited by Unprompted.Config.syntax.delimiter, and img2pez will optimize a single prompt across all provided images. To use this with an entire directory, provide a wildcard as follows: c:/my/images/*.png
Supports the optional target_prompt kwarg. This is a prompt to optimize with img2pez that can be used on its own or in addition to image_path.
Supports the optional prompt_length argument, which is the length of the resulting prompt in tokens. Default to 16.
Supports the optional iterations argument, which is the total number of training steps to perform. Defaults to 200.
Supports the optional learning_rate argument. Defaults to 0.1.
Supports the optional weight_decay argument. Defaults to 0.1.
Supports the amusingly-named prompt_bs argument, which is described by the technique's authors as "number of intializations." Defaults to 1.
Supports the optional clip_model argument. Defaults to ViT-L-14.
Supports the optional pretrain_clip argument. Defaults to openai.
Supports the optional clip_preset kwarg. This provides an easy way to select a compatible pair of clip_model and pretrain_clip settings.
Supports the optional free_memory argument which attempts to free the CLIP model from memory as soon as the img2pez operation is finished. This isn't recommended unless you are running into OOM issues.