This is the v0.32 - BF16 version. Created by baking in Lora trained by using the dataset of [`Archetype: Oversoul - v0.32`](https://moescape.ai/models/b67ce0fe-e214-49a4-8751-dd0706e79a47?index=0&modelversion=47dfc3c1-ef10-4a85-86a2-9978d0f96529) with modification to the captions. This model is aimed to create `digital anime-style illustration`, other styles are not tested thoroughly. --- Recommended Parameters --- Canvas Size: 1K Panorama (1216 × 832) Vertical Panorama (832 × 1216) Square (1024 x 1024) For a more detailed image, or a full body shot, try raising the resolution to up to 2K, as follows: 2K - CUSTOM - use a 2:3 resolution calculator for other custom resolutions Panorama (2496 × 1664) Vertical Panorama (1664 × 2496) Square (2048 × 2048) -> Note: It might be worth it to just use 2K resolution because it generates sharper images with the same mochibean cost as of 2026-07-16. However, understand that resolution choice heavily impacts the resulting image being generated. --- Settings Sampling Method: Unchangeable in this platform (the default is Euler, however) Steps: 8 - 25 (steps above 8 seems to provide diminishing returns, but the cost is the same as of 2026-07-17) CFG Scale: 1 --- You can use any combination of prompt style. NATURAL LANGUAGE works best, however. If writing prompt proves difficult, try using any free LLM in order to create a natural language prompt. Krea 2 is trained by using Qwen as its tagger assistance, meaning that Qwen would probably provide a better, more effective prompt for Krea 2. It is highly capable of rendering text. Try various combination of concepts, such as color combinations, pattern forms, et cetera. I find Krea 2 to be far more capable than other open-weight models in terms of more abstract yet legible concepts. Try random prompts and build it from there. Krea 2 needs a fine-tuning to get itself open and more accessible, as such, the capabilities to produce an "*interesting*" images would be very limited in this regard.