Image style
Image style is the description provided to the Image AI to describe how the generated pictures should look. There are several presets you may choose from. If you would like to tinker to create your own, unique style, select Manticore and then "Custom".
The instructions given to the AI for each element of an image can be found in Variable names. The contents of Image Style do not influence the way in which the storyteller AI writes image instructions for each turn.
Image Model Types[edit]
The are currently two image AI model types available on Infinite Worlds: Natural language and tag-based. Natural language models are receptive to both tag-like keywords and descriptive language.
Natural Language[edit]
The natural language image models currently available are Flux.1 Schnell, Manticore, and Wyvern. These models are fast and effective, utilizing two fields each for the image subject and the environment of the image: "Prompt Beginning" and "Prompt Ending". Each of the fields can use a combination of keywords and statements, such as "masterpiece", "professional", and "style of Pablo Picasso".
Flux also accepts style fusion, i.e. "a mixture of Vincent van Gogh and Banksy". The important thing to remember about these fields is that they are wrapped around the contents of the final image instructions. These do not influence how the storyteller AI writes the image instructions.
Because the image instructions are written by the storyteller AI, main instructions can be used to give more detailed instructions that take advantage of Flux's natural language interpretation. Flux's style guide includes the use of "layer" instructions, i.e. "layer 1 (foreground): description", "layer 2 (midground): description", and "layer 3 (background): description", and the AI can be instructed to use this approach and more when it writes the image instructions. These same layering methods can apply similarly to Manticore and Wyvern. One can also include elements like "lighting notes".
Limitations[edit]
Flux can reasonably only support about 300 words in a prompt before dropping elements at the end of the illustration prompt.
Manticore can reasonably support about 400 words in a prompt before dropping elements at the end of the illustration prompt.
IW Flux Styles[edit]
There are a few keywords which trigger Infinite Worlds to use a specific LoRA (LoRAs are add-ons which change the behaviour of the image model):
- IWDefault
- IWClassic
- IWAnime
- IWRemoveNudityWordsWhenNoNudity
If Infinite Worlds finds one of these keywords it will turn on the associated LoRA, and then remove the keyword from the prompt before sending it to the image model.
IW Manticore Styles[edit]
- IWUpsaleFace attempts to improve face quality in photorealistic and near photorealistic image styles with upscaling
- IWUpsaleFaceSmooth attempts to improve face quality in photorealistic and near photorealistic image styles with upscaling while reducing "flaws" (i.e., smoothing)
Tag-based[edit]
All other image models are tag-based, with limited space for tags regarding the subject and environment of the generated image. Both the subject and environment receive "high priority" and "low priority" tags which influence the content and focus of the image generation. These tags are then added to the text created by the storyteller AI.
As with Flux, the storyteller AI can be told in main instructions how to enhance its effectiveness with these models, such as by telling it to use "tag-only" descriptions.