<h3 id="0.8">权重推荐0.8左右。</h3>一般来说，不使用触发词也能有很好的效果。如果效果不明显则可以使用触发词：MellowStyle<h3 id="weight:-around-0.8">weight: around 0.8</h3>general, it works well even without trigger words.If the effect is not obvious, you can use the trigger word: MellowStyle<h3 id="mellow">这是 MELLOW 社的画师 まうめん 的画风：</h3>This is the art style of Maumen, an artist from MELLOW Studio:<ul><li>有较强的环境光影表现（例如 黄昏、星空、边缘辉光 等）。It features strong environmental lighting and shadow rendering (e.g., dusk, starry skies, edge glow, etc.).</li><li>脸型和头发轮廓比较固定（因为画师画风就是如此），当指定角色或加入角色lora时，脸型和头发轮廓会受到一定影响。Facial features and hair silhouettes are relatively consistent (as this is inherent to the artist's style), though they will be influenced to some extent when specifying a character or incorporating a character LoRA.</li><li>角色的皮肤阴影比较浓厚，带有些许的立体感。The character's skin shading is relatively pronounced, lending a subtle sense of three-dimensional depth.</li><li>整体效果趋向于galgame游戏cg的观感。The overall effect leans toward the aesthetic of galgame-style game CGs.</li><li>因为训练图片中有接近一半的nsfw图片，在不描写动作的情况下很大概率出现nsfw内容，请注意写出动作提示词或者添加负面提示词。Since nearly half of the training images are NSFW, there is a high probability of NSFW content being generated even without describing actions. Please remember to include action prompts or add negative prompts.</li></ul><h3 id="loralora">这是我训练的第一个风格lora，可能存在一些问题，如果有什么训练画风lora的建议请评论告诉我。</h3>This is the first style LoRA I've trained, so there may be some issues. If you have any advice on training style LoRAs, please let me know in the comments.Training Parameters (on Civitai)：<ul><li>190 images (1980*1114 / 1280*1810)</li><li>Epochs: 25 (selected 23)</li><li>Num Repeats: 1</li><li>Train Batch Size: 4</li><li>Shuffle Tags: Checked</li><li>Keep Tokens: 1</li><li>Unet LR: 1</li><li>Text Encoder LR: 0</li><li>LR Scheduler: cosine</li><li>Network Dim = Network Alpha = 32</li><li>Noise Offset: 0.04</li><li>Optimizer: Prodigy</li></ul>

Files with SHA256: abf61d1aaf751fe8aa4df6d4f7b08d66c3bedd676b94ff4baba769cdef31cc94