Kolors|youkengi anime base V1.0

中文自然语言生成を強く推奨します。英語のプロンプトは効果が非常に劣ります。

全身を示唆する場合（たとえば「靴」や「足」などの言葉を含む場合）、適切な縦横比を調整する必要があります。そうでないと、全身が映りきらず、不自然な肢体のエラーが発生しやすくなります。

推奨解像度（縦向き・横向きどちらでも可）：864*1152*2、864*1536*2、1024*1024*2、1280*1280*2

CFG：3.5（細い線で彩度低め）または4.0（太い線で彩度高め）

サンプリング方法：DPM++ 3M SDE Karras

高解像度修復モデル：4x-AnimeSharp

高解像度修復強度：0.35

VAE： 内蔵VAEを使用しているため、「自動」を選択してください

否定プロンプト：空白のままにしてください（賛美語は不要です）。表示したくない要素がある場合は、個別に記述してください

スタイルトリガー語：通常は必要ありません。必要に応じて追加してください：二次元アニメ風（写実的な関連語が多すぎると、可図公式モデルの概念が活性化されるため、二次元スタイルを維持したい場合はこのスタイルトリガー語を追加してください）、1girl、young man（若々しい外見）、a woman、man（成熟した外見）。子供を描きたい場合は、「小男孩」または「小女孩」と明記してください。

It is strongly recommended to use Chinese natural language generation, as English word prompts may be have poor effects.

If the full body is implied (for example, by including shoes or feet in the description), it's necessary to adjust the aspect ratio appropriately; otherwise, the image may not be able to fit the entire figure, which could lead to strange limb errors.

Suggested resolutions (both vertical and horizontal images are acceptable): 864*1152*2、864*1536*2、1024*1024*2、1280*1280*2

CFG (Classifier-Free Guidance Scale): 3.5 (for fine lines and low saturation) or 4.0 (for thicker lines and higher saturation)

Sampling Method: DPM++ 3M SDE Karras

Upscaling Model: 4x-AnimeSharp

High-Resolution Denoising Strength: 0.35

VAE : A built-in VAE is available, so selecting 'Auto' is sufficient

Negative Prompt: Leave blank (no need for any quality words); if there's something you don't want to appear, it can be specified separately

Style Trigger Words: In most cases, style trigger words are not needed; If it is necessary, add Anime-style (When there are many realistic-related terms, this will activate the concept of the official model; if you still want to maintain an anime style, you can add these style trigger words), 1girl, young man (young appearance), a woman, man (mature appearance).

未来可图|优可可图二次元模型是优可可图系列模型的第一个模型，遵从可图Apache License 2.0开源协议，代号Youkengi Anime Base Kolors 。为了更好更快地助力中文模型生态发展，本模型完全开源，基于本模型做出的转载、微调/融合只需注明出处。

Kolors|Youkengi Anime Base is the first model in the Youkengi series of models in Kolors, adhering to the Apache License 2.0 open-source agreement. To better and more quickly promote the development of the Chinese model ecosystem, this model is fully open source. Any redistribution, fine-tuning, or fusion based on this model only requires crediting the source.

模型能力评价：

1. 能够以较稳定且精美的二次元风格出图：基本画风精细的二次元风格，细节适中。虽然realistic,3D rendering等tag依然有效，但低权重下仍然会带有强烈的二次元风格；

2. 天然好手和较好的肢体：类似于石头、剪刀、布、比爱心、握持等手型表现较好。不指定手型时表现略差，但已大幅强于可图官模；在合适的长宽比下，肢体比较好，反之容易产生奇怪的肢体错误。

3. 极强的文本理解力、中国本土概念较好：可以理解SDXL无法理解的高难度prompt，有较多国外模型没有的的中国本土概念和古诗词的理解能力；

4. 支持中文，容易上手：可以直接用中文白话输入提示词，妈妈再也不用担心老是遇见不认识的单词了，不需要负面prompt；

5. 极强的lora结合能力：基于可图模型很强的泛化性，同时炼制的时候注意控制了污染，经测试与大多数lora风格的结合较好。因为模型本身曝光较强，唯一不太适合的可能是自带光污染的lora。

6. 较强的自然构图能力：以牺牲少量手和肢体表现为代价加强了自然构图能力，在通常没指定动作的情况下画的角色不会十分呆板的站在原地，双手下垂，而是会随机出现一些动作，使画面更加生动。

Model Capability Evaluation:

1. Capable of producing images in a stable and exquisite anime style: The basic art style is detailed in its anime aesthetics, with moderate detail. While tags like realistic and 3D rendering are still effective, they maintain a strong anime style even at low weights.

2. Naturally good hands and decent limbs: Hand poses such as rock-paper-scissors, heart gestures, and holding objects are well-represented. Performance is slightly worse when hand poses are not specified, but it's significantly better than the official Youkengi models. Limbs perform well under appropriate aspect ratios; otherwise, there can be odd limb errors.

3. Excellent text comprehension and good understanding of Chinese local concepts: It can understand high-difficulty prompts that SDXL cannot comprehend, and it has a grasp of many Chinese local concepts and classical poetry that are absent in foreign models.

4. Supports Chinese and easy to use: Prompts can be input directly in colloquial Chinese, eliminating the worry of encountering unfamiliar words and negating the need for negative prompts.

5. Strong LoRA integration capability: Due to the robust generalization of the Youkengi model and careful control during training to prevent contamination, it integrates well with most LoRA styles. The only potential incompatibility might be with LoRAs that introduce their own light pollution due to the model's inherent exposure strength.

6. Strong natural composition ability: This is achieved at the cost of slightly reduced hand and limb performance, enhancing the natural composition so that characters drawn without specific action instructions do not stand rigidly in place with their arms hanging down. Instead, they will randomly adopt various poses, making the scene more lively.

可图模型（kolors）简介：有较好的中文提示词支持，在训练时的算力相较SDXL更低（仅训练unet），是目前较有希望扩展出完整中文生态的模型架构。可图官方的基础模型本身也具有很强的泛化性，训练结果可以很好地反映到模型上，模型内置了多种图像风格，本身具有很好的综合实力。

Introduction to the Kolors Model: It offers better support for Chinese prompts and required less computational power during training compared to SDXL (training only the UNet component). It is currently one of the moSt promising model architectures for developing a complete Chinese ecosystem. The base model provided by Kolors officially also possesses strong generalization capabilities; the training outcomes are well-reflected in the model. It comes equipped with various image styles and demonstrates excellent overall capabilities.

后记：（调试记录）

V0.1 调整基本二次元画风；

V0.2 优化手部表现，修正通常情况难以出现下半身的问题；

V0.3 优化自然构图表现，基于文本理解能力优化；

V0.4 进一步调整自然构图表现，画风柔和微调；

V0.5 调整腰部以下的部位容易出现的肢体错误，损失少量自然构图表现；

V0.6 调整画面精细度表现，降低像素不足是出现的脸部崩坏；

V0.7 增强模型二次元插画质感，进一步优化手部表现，但由此产生了过曝和细节杂乱的问题；

V0.8 修正细节杂乱的问题；

V0.9 平衡整体画风，修正前述调整过程中出现的过曝问题；

V1.0 调整头身比，提升肢体表现，提升基本清晰度，修正偶发的肌肉错误表现。埋了可追溯模型的隐藏触发词。

Postscript: (Debugging Record)

V0.1 Adjusted basic anime art style;

V0.2 Optimized hand representation, fixed the issue of lower body parts;

V0.3 Improved natural composition, optimized based on text understanding capabilities;

V0.4 Further adjusted natural composition, softened the art style with minor tweaks;

V0.5 Addressed limb errors frequently occurring below the waist, at the cost of some natural composition;

V0.6 Improved detail fidelity, reduced facial distortion due to insufficient pixels;

V0.7 Enhanced the anime illustration texture of the model, further optimized hand representation, but introduced overexposure and cluttered detail issues;

V0.8 Fixed the cluttered detail issues;

V0.9 Balanced the overall art style, corrected overexposure issues from previous adjustments;

V1.0 Adjusted head-to-body ratio, improved limb representation, enhanced basic clarity, corrected occasional erroneous muscle representation. Implanted hidden trigger words for model traceability.