Advanced Settings - Configure Tokens and Other LLM and Stable Diffusion Settings

Advanced Settings

Learn about the advanced LLM and SD settings.

Advanced Settings

Touchy allows for a ton of customization and this includes various Stable Diffusion settings to give our character creators all the tools they need for unique inspired results. These settings are not for beginners and are in no way required to produce great characters. For newbies to AI, it's always best to stick to the defaults. Below we will explore the character creation settings and their pros and cons.

Max New Tokens

At the core of a language model's response generation lies the Max New Tokens field. Tokens, essentially words, are capped by this setting, ensuring a controlled response length. It's important to note that this doesn't guarantee reaching max_tokens due to other potential stopping conditions, but it certainly halts at max_tokens.

Uses: Prevents overly lengthy and chaotic responses from your character.

Temperature:

The Temperature setting plays a pivotal role in controlling the randomness of the language model's output.

Uses: Higher values increase output randomness and lower values enhance focus and determinism.

Risks: Higher values may compromise coherence. Lower values may lead to dull and repetitive responses.

Top K:

An alternative to Temperature or Top P, Top K sampling involves exclusively selecting the top K tokens with the highest probability.

Uses: Effective in cases of narrow band distributions, eliminating the possibility of choosing a completely random token. Higher values increase output randomness, while lower values enhance focus and determinism.

Risks: Higher values may decrease coherence. Lower values may result in dull and repetitive outcomes.

Top P:

Similar to Top K, Top P is an alternative to Temperature, influencing the randomness of the language model.

Uses: Higher values increase output randomness. Lower values enhance focus and determinism.

Risks: Higher values may make the output less coherent, while lower values may make it dull and repetitive.

Typical P:

Yet another alternative, Typical P influences token sampling based on the expected information content given the previous partial text.

Uses: Lower values increase randomness and diversity. Higher values allow for more determinism, maintaining response quality while reducing degenerate repetition.

Risks: Lower values may generate less coherent text.

Stable Diffusion Advanced Settings:

Num Inference Steps:

Determining the number of denoising steps from random noise towards a prompt, a higher value implies a more detailed image.

Uses: Higher values result in a higher level of detail.

Risks: Higher values may introduce unnecessary details. Lower values may lead to an underdeveloped image.

Guidance Scale:

This setting controls how closely the generated image adheres to the provided prompt. A higher guidance scale enforces strict adherence, while a lower scale encourages creativity.

Uses: Higher guidance scale enforces stricter adherence. Lower guidance scale fosters more creative outputs.

Risks: Higher guidance scale may compromise image quality.

Clip Skip:

A metric influencing the accuracy of the text model during the diffusion process by "skipping layers" of the CLIP model.

Uses: Higher values correspond to faster generation speed and increased creativity.

Risks: Higher values run the risk of inaccurate generations.

Negative Prompts:

Use this setting to specify what you do not want to see, preventing undesired elements in generated content.

Multilingual:

Enables multilingual prompts to generate images with English as the default.

Self Attention:

Improves image quality at the expense of speed.

Upscale:

Increases the generated image resolution by two times (2x) if set to "yes."

Use Tomesd:

Significantly speeds up token diffusion by merging redundant tokens.

Use Karras Sigmas:

Improves image quality by using the Karras Sigmas scheduler, adjusting timesteps to minimize truncation errors in denoising.

LORA Training

Overview