TOML support for sample prompt

[Linaqruf]

Hi, I want to propose small changes for loading args.sample_prompt. I added .toml support so we can easily assign the sample prompt parameter to each key instead of putting all parameters in one line (using .txt). The sample prompt parameter is also more readable and easy to maintain. Not only that, we don't need to set the same parameter again because I borrowed the dataset_config approach by using [[prompt.subset]] to store additional parameters.

An example of sample_prompt.toml is as follows:

[prompt]
negative_prompt = "(worst quality, low quality:1.4)"
scale = 7
sample_steps = 20
width = 512
height = 768

[[prompt.subset]]
prompt = "masterpiece, best quality, masterpiece, best quality, solo, looking at viewer, short hair, bangs, red eyes, gloves, 1boy, holding, hair between eyes, closed mouth, upper body, weapon, male focus, red hair, black gloves, belt, sword, cape, holding weapon, armor, fur trim, holding sword, red background, red cape"

[[prompt.subset]]
prompt = "masterpiece, best quality, 1girl, solo, long hair, looking at viewer, smile, bangs, blue eyes, dress, ribbon, holding, bare shoulders, blue hair, flower, frills, teeth, sleeveless, hand up, signature, white dress, grin, red ribbon, blue sky, petals, neck ribbon, floating hair, sleeveless dress, blue background, light particles, blue theme, multicolored eyes, yellow flower, holding flower, orange flower, orange ribbon"

[[prompt.subset]]
prompt = "masterpiece, best quality, 1girl, solo, long hair, breasts, looking at viewer, smile, bangs, blue eyes, skirt, red eyes, thighhighs, holding, jewelry, medium breasts, very long hair, blue hair, weapon, braid, red hair, multicolored hair, sword, cape, holding weapon, armor, two-tone hair, thigh strap, heterochromia, holding sword, tiara, black background, crossed bangs, crown braid, split-color hair"

We can also set prompt into different resolution or using a different sampler for each prompt by putting important parameters to [[prompt.subset]].

Example:

[prompt]
negative_prompt = "(worst quality, low quality:1.4)"
scale = 7
sample_steps = 20

[[prompt.subset]]
prompt = "a dog"
width = 512
height = 768
seed = 1337

[[prompt.subset]]
prompt = "a cat"
width = 1024
height = 768

I also added .json support in the latest commit.

Thank you

[kohya-ss]

Thank you for this! The .txt file is good enough for me, but this would make it easier to understand the prompt settings.

[Linaqruf]

Thank you! I use it for Colab notebook so end user can easily add or edit the prompt config without needing further explanation about prompt args. It also to makes merging easier in the future. 😅