YooKooTu Anime Test Version

Before we begin: Strongly recommend using Chinese natural language generation, the prompt effect of English words is very poor. The conversation with the model has no preset parameters, the effect is very bad, don't use it...

If hinting the whole body (such as writing into shoes, feet, etc.) will require adjusting the appropriate length-to-width ratio, otherwise, the image might not fit the whole body and strange limb errors may occur.

Recommended resolution (portrait or landscape): 864*1152*2, 864*1536*2, 1024*1024*2, 1280*1280*2

CFG: 3.5 (fine lines, low saturation) or 4.0 (thicker lines, high saturation)

Sampling method: DPM++ 3M SDE Karras

HD restoration model: 4x-AnimeSharp

HD repetition rate: 0.35

Vae: Built-in Vae, select automatic

Negative trigger words: Leave it blank (no need for flattering words) If there is something you don't want to appear, you can write it separately

Style trigger words:Two-dimensional anime style (if there are many realistic-related words, it will activate the idea of Kantu officials' paintings. If you still need to maintain a two-dimensional style, you can add this style trigger word), A girl, young man (young age and appearance) A woman, man (mature age and appearance). For children, write young boy or young girl.

Future Kantu | Preferred Kantu Second Dimension Model is the first model of the Kantu series, in accordance with the Kantu Apache License 2.0 open source agreement, codenamed Youkengi Anime Base Colors. In order to better and faster support the development of the Chinese model ecosystem, this model is completely open source, and any reprinting, fine-tuning/fusing work based on this model only needs to indicate the source.

Model performance evaluation:

1. Capable of producing images in a stable and exquisite two-dimensional style: Basic fine two-dimensional style, moderate in detail. Although tags such as realistic, 3D rendering remain effective, under low weight, a strong two-dimensional style will still be present;

2. Naturally good at hands and good with limbs: Similar to rock, paper, scissors, love, holding hands, etc., the hand shapes are well represented. When not specifying hand shapes, the performance is slightly worse, but still much stronger than the Kantu official model; with the appropriate aspect ratio, the limbs are better, otherwise strange limb errors may occur.

3. Extremely strong text comprehension and good understanding of Chinese local concepts: Capable of understanding high-difficulty prompts that SDXL cannot grasp, with a better understanding of Chinese local concepts and ancient poetry compared to many foreign models;

4. Supports Chinese, easy to use: Directly inputting prompts in colloquial Chinese, moms no longer need to worry about encountering unfamiliar words, no negative prompts needed;

5. Strong ability to combine lora: Based on the powerful generalization of the Kantu model, while refining to control pollution, it has been extensively tested and combines well with most lora styles. Because the model itself has strong exposure, the only thing it may not be suitable for is lora with its own light pollution.

6. Strong natural composition ability: Strengthened the natural composition ability at the expense of slightly sacrificing hand and limb performance, characters drawn under normal circumstances without specified actions will not stand still in a rigid manner, hands hanging, but will randomly perform some actions to make the scene more lively.

Kantu model (kolors) introduction: With good Chinese prompt word support, lower training computation compared to SDXL during training (only training unet), it is a model architecture that is more promising in expanding a complete Chinese ecosystem at the moment. The basic model of Kantu itself also has strong generalization, the training results can be well reflected in the model, the model has built-in multiple image styles, and has excellent comprehensive strength.

Since the main language mode carried by the glm3 model is Chinese natural language, using Chinese natural language sentences can achieve a good experience. And because the Kantu model does not carry the clip model, the performance of word-by-word popping is not as good as other models that carry the clip language.

Postscript: (debugging record)

V0.1 Adjusted basic two-dimensional drawing style;

V0.2 Optimized hand performance, corrected the issue of the lower body being difficult to appear in usual circumstances;

V0.3 Optimized natural composition performance, optimized based on text comprehension ability;

V0.4 Further adjustment of natural composition performance, softening and fine-tuning the drawing style;

V0.5 Adjusted the common limb errors below the waist, sacrificing a small amount of natural composition performance;

V0.6 Adjusted the image precision performance, reducing the occurrence of facial collapse due to insufficient pixels;

V0.7 Enhanced the model's two-dimensional illustration texture, further optimized hand performance, but this led to overexposure and cluttered detail issues;

V0.8 Corrected cluttered detail issues;

V0.9 Balanced the overall drawing style, corrected overexposure issues that occurred during the aforementioned adjustments;

V1.0 Adjusted the head-to-body ratio, improved limb performance, increased basic clarity, corrected occasional muscle error representations. Hidden trigger words for the traceable model were embedded.