关于klein 4b版本：强度：1-1.3我在网盘里翻了半天终于翻出了这个素材，感觉这套素材比较干净，适合用在新模型上做尝试。已知问题：<ul><li>因为训练用的图片不多（89），所以尽量不要用这个lora生成面部特写或者一些古怪的表情，容易将画风还原成底模的写实风格。</li></ul>碎碎念：klein这个模型的编辑能力不错，配置要求低，速度也快。但单纯的文生图能力，我觉得是有点糟糕的，除开几乎没有的美学调整外，就连肢体结构也是一塌糊涂（无论是4b模型还是9b模型）：三条腿、四根手指、糊成一团的手指等等，在文生图中大量出现，甚至在生成打伞的人时，伞柄极大概率会插进角色的脑袋里......简直是噩梦，我真不知道为什么2026年的新模型在结构上的表现会这么糟糕，不说比不上flux1，甚至和初版sdxl差不多。---About klein 4b version：Strength: 1-1.3I searched through my cloud drive for a while and finally found this material. It feels quite clean and suitable for trying out on new models.Known issues：<ul><li>Since there aren't many training images (89), try not to use this LoRA for facial close-ups or weird expressions—it can easily revert the style back to the base model's realistic look.</li></ul>PS: The klein model isn’t bad for editing—it has low system requirements and runs fast. But when it comes to pure text-to-image generation, I find it bit awful. Aside from barely any aesthetic tuning, even limb structures are a mess (whether in the 4b or 9b version): three legs, four fingers, blurred hands, etc. appear constantly in text-to-image results. Even when generating someone holding an umbrella, the handle often pierces the character's head… It’s a nightmare.I really don’t understand why a new model from 2026 performs so poorly in structure—not only worse than flux1, but even on par with the early SDXL.<hr />强度：0.5-1.5<ul><li>可爱系画风，主要特点是柔和、低饱和度的色彩，稍显简约的线条以及完全平面的风格，适合可爱系的肖像特写。</li><li>受基础模型的影响较大， 会在不同的基础模型上表现出差异化的风格。</li></ul>---Strength: 0.5-1.5<ul><li>A cute art style, featuring soft, low-saturation colors, slightly simple lines, and a completely flat illustration style. Great for cute portrait close-ups.</li><li>It is influenced by the base model and may show different styles depending on the model used.</li></ul>

Files with SHA256: 4c29ececfedeb4f59b8fba4392be89fdc22be94860fd989b0340c7a99600f8e8