Try size: 512 x 768 (SD 1.5) / 960 x 1280 (Illustrious)Weights: 1Clip skip: 2Pictures from CalicoMix 8.5 and Illustrious-XL SmoothFT-SOLIDRecommend negative: ((book, table, border, grayscale, monochrome)), &lt;= SD 1.5Note: 這是一個訓練失敗的lora，它原本是灰階的，幾乎不受prompt控制的。我混合幾個lora後，勉強提高顏色跟操控性，它的特性是在某些模型上描述黑暗主題時會產生大量細節，但在某些模型幾乎無效，例如知名的anylora，雖然我就是用它訓練的，但我看不出套上的效果，表示anylora很乾淨吧？不會干擾，也不會加分。目前我只在CalicoMix8.5跟7.5上得到了較佳的結果。假如你希望一套上去，就達到一樣的效果，那可能你要失望了。我猜這個lora可能還要搭配一些強調美學細節風格的lora，或者已融合進基礎模型裡了。 (SD 1.5)Note: This is a lora that failed to training, it was originally grayscale and almost uncontrollable by prompts. I mixed several LoRAs to somewhat improve color and controllability. Its characteristic is that it generates a lot of detail when describing dark themes on certain models, but it is almost ineffective on others, like the well-known Anylora. Even though I used it for training, I can't see its effect, which means Anylora is quite clean, right? It neither interferes nor adds any value. Currently, I've only obtained better results on CalicoMix 8.5 and 7.5. If you're hoping to just apply one and achieve the same effect, you might be disappointed. I guess this LoRA might need to be paired with others that emphasize aesthetic details or has already been integrated into the base models. (SD 1.5)Note: 素材不可能拿一張靜香戴竹蜻蜓的圖片訓練任何SDXL版本，那是複印法偶然間產生細節爆炸的失敗品，所以只好拿SD 1.5 生成的圖片來訓練LORA。這次我沒有像往常一樣訓練Pony版本，而是第一次試著訓練illustrious，也好奇它對美學的表現如何，似乎看到了以前類似Anything V4.5的畫風影子(關於它的爭議性我不想多談)。反正，未來將高難度姿勢、復刻動漫角色跟需要大量細節的畫風，主要交給illustrious來訓練看看。至於Pony，在選擇優化器方面，有時是Prodigy訓練比Adafactor好(需要低輪數提前擬合時)，有時是Adafactor訓練比Prodigy好(需要高輪數處理背景時)，未來可能會以Prodigy為主，因為Prodigy需要的訓練時間較Adafactor短，比較適合先探路，不適合的話再靠Adafactor訓練。Pony訓練素材時，有時也很依賴不同的單張次數，也不擅長遠距離素材(需要裁剪改善)，想押對寶要靠些推測跟運氣。至於Anything V5是我很熟悉的訓練模型，訓練時間也較快，所以我都是先用Anything V5探路的，視情況修改或增刪素材，而Illustrious也與Anything V5一樣很快就讓我找到熟悉感，很容易測試出圖時找到放觸發詞的最佳順序，Pony我時常需要多摸索一陣子才能找到適合的順序。但平心而論，Pony在搞笑跟NSFW上就真的是數一數二的強了，我沒打算放棄它，它跟CalicoMix的風格很合拍，Illustrious也不見得能占上風。(Illustrious)Note: Omitted translation... because both Google and chatgpt machine translations are weird. (Illustrious)

random dark style

randomdarkstyle_illustrious.safetensors

Illustrious

SD 1.5

Try size: 512 x 768 (SD 1.5) / 960 x 1280 (Illustrious)Weights: 1Clip skip: 2Pictures from CalicoMix 8.5 and Illustrious-XL SmoothFT-SOLIDRecommend negative: ((book, table, border, grayscale, monochrome)), &lt;= SD 1.5Note: 這是一個訓練失敗的lora，它原本是灰階的，幾乎不受prompt控制的。我混合幾個lora後，勉強提高顏色跟操控性，它的特性是在某些模型上描述黑暗主題時會產生大量細節，但在某些模型幾乎無效，例如知名的anylora，雖然我就是用它訓練的，但我看不出套上的效果，表示anylora很乾淨吧？不會干擾，也不會加分。目前我只在CalicoMix8.5跟7.5上得到了較佳的結果。假如你希望一套上去，就達到一樣的效果，那可能你要失望了。我猜這個lora可能還要搭配一些強調美學細節風格的lora，或者已融合進基礎模型裡了。 (SD 1.5)Note: This is a lora that failed to training, it was originally grayscale and almost uncontrollable by prompts. I mixed several LoRAs to somewhat improve color and controllability. Its characteristic is that it generates a lot of detail when describing dark themes on certain models, but it is almost ineffective on others, like the well-known Anylora. Even though I used it for training, I can't see its effect, which means Anylora is quite clean, right? It neither interferes nor adds any value. Currently, I've only obtained better results on CalicoMix 8.5 and 7.5. If you're hoping to just apply one and achieve the same effect, you might be disappointed. I guess this LoRA might need to be paired with others that emphasize aesthetic details or has already been integrated into the base models. (SD 1.5)Note: 素材不可能拿一張靜香戴竹蜻蜓的圖片訓練SDXL版本，那是複印法偶然間產生細節爆炸的失敗品，所以只好拿SD 1.5 生成的圖片來訓練LORA。這次我沒有像往常一樣訓練Pony版本，而是第一次試著訓練illustrious，也好奇它對美學的表現如何，似乎看到了以前類似Anything V4.5的畫風影子(關於它的爭議性我不想多談)。反正，未來將高難度姿勢、復刻動漫角色跟需要大量細節的畫風，主要交給illustrious來訓練看看。至於Pony，在選擇優化器方面，有時是Adafactor訓練比Prodigy好(需要高輪數處理背景時)，有時是Prodigy訓練比Adafactor好(需要低輪數提前擬合時)，Pony訓練素材時，有時也很依賴不同的單張次數，想押對寶要靠些推測跟運氣，我個人反而覺得Anything V5很少讓我誤判，所以我都是先用Anything V5探路的，而Illustrious的訓練穩定度也與Anything V5旗鼓相當。(Illustrious)Note: Omitted translation...Both Google and chatgpt machine translations are weird. (Illustrious)

Files with SHA256: 9e348fd39a20d8d8e0658f1bc82ca49b2612517c623e4cf6a56250ea03360749