Modes and Parameters

Mode

Mist provides three modes for different anti-imitation scenarios. Users can decide which mode to use according to their needs.

Textural mode: By injecting confusing texture information into the watermark to achieve the effect of anti-AI imitation; mainly against Img2Img; requires less GPU memory.

Semantic mode: By interfering with the semantic information of the original image with the watermark; mainly against subject-driven generation (Textual Inversion, Dreambooth, etc.) scenes; requires more GPU memory.

Fused mode: By mixing Textural and Semantic modes in a certain ratio; requires more GPU memory.

The following table demonstrates the performance of the three modes in the four scenarios: Textual Inversion, NovelAI Img2Img, Dreambooth, and Scenario.gg. Among them, Textual Inversion and Dreambooth are both based on Stable Diffusion.

	Textual Inversion	NovelAI Img2Img	Dreambooth	Scenario.gg
Textural	○	◎	○	◎
Semantic	◎	△	╳	○
Fused(Fusion weight=1)	◎	○	△	○
◎:very strong ○:strong △:medium ╳:weak

Parameters

Mist allows users to customize several parameters that would characterize the watermark. A brief introduction to these parameters are given in the UI of Mist. The following table further gives some details of these parameters.

Parameter	Recommended value	Note
Strength	16	16 promises comprehensive performance in Fused Mode. Relatively, 8 would provide strong performance against certain applications in other two modes.
Steps	100	100 is a good tradeoff of time cost and performance.
Output size	512
Fused weight	1	Fused weight balances the performance tradeoff of Fused mode between Semantic and Textual mode.
Low VRAM Mode	False	Low VRAM Mode greatly lows down the VRAM requirements with subtly reduced performance.