CivArchive
    CalicoMix EroILL - v2.5
    NSFW
    Preview 52409870
    Preview 52409545
    Preview 52409568
    Preview 52409753
    Preview 52414995
    Preview 52415136
    Preview 52419598
    Preview 52413007

    CalicoMix EroILL is a versatile Illustrious SDXL model offering enhanced capabilities over the classic version. It generates diverse anime content, from SFW to NSFW.
    It is a model focused on cuteness and sexiness.

    It's the best and most lewd model I have made until now.

    CalicoMix EroILL is Available on TensorArt

    "If you like my work, feel free to support me on Ko-fi!"

    CalicoMix EroILL is a crazy model

    recommends using rating prompts:
    general > sensitive > questionable > explicit

    use 'general' for SFW.

    can add 'questionable' and 'explicit' at the same time to unlock max crazy lewdness.

    can add 'questionable' and 'explicit' in the negative prompt to force more SFW.

    Breast size control:
    flat chest > small breasts > medium breasts > large breasts > huge breasts > gigantic breasts
    can add breast size to the negative prompt.
    can add multiple breast sizes for share effects between multiple tags.

    can use weight to tone down breast sizes. eg. (large breasts:0.7)

    Sampler: DPM++ 2M SDE (dpmpp_2m_sde_gpu), Euler, Euler A, DDIM

    Scheduler: Karras, DDIM Uniform, AYS SDXL, Beta, Simple

    Steps: 20-30

    CFG Scale: 5.0 - 8.0

    Clip Skip: 1 - 2

    VAE: set to Auto to use included VAE

    Resolution: various

    Hires. Fix Settings (if necessary)

    Upscaler: 4x UltraSharp, 4x AnimeSharp

    Hires step: 10-15 (around half of the Steps, increase if it is too low details, decrease if it starts to generate duplicate objects like navel)

    Denoising: 0.3-0.5 (More Denoising = More Makeup)

    Hires scale: 1.5+

    Recommended Prompt
    masterpiece, best quality, very aesthetic, absurdres, newest,

    Recommended Negative Prompt

    worst quality, bad quality, low quality, lowres, scan artifacts, jpeg artifacts, sketch, signature, depth of field, blurry, blurry background, simple background, fog, (white background:1.1), pasties, censored

    ***Beware of spelling wrong English on prompt

    Thanks, TechnoByte, for testing and showcasing.

    Thanks, NovaZone, for your support and testing.

    Thanks, all supporters.

    Description

    Improve color, anatomy, hand drawing, and prompt effectiveness

    FAQ

    Comments (13)

    Kybalico
    Author
    Jan 18, 2025· 13 reactions
    CivitAI

    Sorry I'm late, but Happy New Year!

    ghostpaintJan 18, 2025· 1 reaction

    新年快樂!2.5看起來好像是克服了1.0太亮太白的問題了,不過這個版本太完美了,ilxl的danbooru內建數量跟精緻度又遠遠超過pony,訓練lora陷入一個難題,會不會不小心劣化它...可能我會盡量朝著低擬合訓練吧。

    Kybalico
    Author
    Jan 18, 2025· 1 reaction

    @ghostpaint 你可以對基礎模型進(ILL)行 LoRA 訓練,這將有助於防止過度擬合。

    ghostpaintJan 18, 2025· 1 reaction

    @Kybalico 我之前的ilxl lora都是用ill v0.1訓練的,並沒有太過度擬合的情況,只有噪點偏移太亮的情形,現在關掉了。不過如果出圖的基礎模型底模足夠好看的話,可以降低覆蓋率去混合畫風,可能要放棄prodigy,或在超參數裏調低學習率...Pony版就還好,參數可以維持不變,1.0的底模跟我的lora的水平接近。SD1.5版通常是我的Lora比8.5底模好看一些…EroILL 2.5就是剛好相反有點尷尬,我的ilxl lora能配得上它嗎?還是不小心劣化它了。(哈)

    ghostpaintMar 4, 2025· 1 reaction

    v4.0可能有受到噪點偏移影響,淺色的地方更白,深色的地方更黑。所以我還是回去用v3.0好了,有些看不習慣。(笑)

    ghostpaintMar 4, 2025· 1 reaction

    我在kohya的Additional parameters是用--noise_offset="-0.05"去抑制,動漫比寫實更怕噪點偏移影響。所以有些人是用AnyIllustrious-XL1.1去訓練lora,不過這個模型不太適合我,練出來的Lora背景結構常出錯,所以我還是用Illustrious-XL0.1去訓練lora。

    ghostpaintMar 4, 2025

    好奇怪…我的電腦螢幕明暗對比好明顯,但手機上看xyz圖就差異不明顯了…

    ghostpaintMar 24, 2025

    我晚點來試...

    ghostpaintMar 25, 2025· 1 reaction

    5.0版通常淺色不會亮到很誇張,深色也不會暗到像黑色(偶爾才會有所以可以忽略無視,也可能是我Lora的問題)。5.0版看起來是很耐看的,而且素材的還原度是較好的,更近似於SD1.5版CalicoMix7.5誇張的表情。提詞的服從度有下滑一些,所以不像以前一定要寫出來,它才願意主動做出那些變化,即使looking at viewer或looking away沒寫出來,女孩視線也很自然,視線落點很平均地出現。如果寫出兩種髮型它會很平均出現,較少偏重在其中一個髮型。若寫出smile也不會頻繁造成兩個兩個女孩同時笑,因為實際上素材大多是一個笑一個不笑的,這也是為何一般都建議文本過濾器不要跟Unet一樣數值(或者N輪或N%丟棄文本Dropout Caption),SD1.5是較怕文本先過擬合才設低。如果專攻NSFW或單一姿勢,確實服從度高一些是比較好,讓danbooru tag較有效率感。但對於非NSFW的主題就會顯得很呆版,圖片生成較缺乏變化性。所以我覺得5.0這個平衡性,算是最好的,較不用一直更換提詞來提高變化性。

    Kybalico
    Author
    Mar 25, 2025· 1 reaction

    评测写得太详细了,真的很感谢!

    ghostpaintMay 15, 2025

    由於目前評級變得更加嚴格,Suggested Resources這欄位是完全看不到你精彩的v5.0。就算我選任何瀏覽等級也都是看不到的,它很笨也不會跳到第二張安全的圖。如果您還會升級v6.0,建議不要讓封面圖片被標註 R、X、XXX。(目前v5.0這封面太經典不一定要更換順序)

    LazarotJan 19, 2025· 1 reaction
    CivitAI

    Me gusta mucho lo detallado de tu descripción, super claro. Ojala todos hicieran lo mismo <3

    Kybalico
    Author
    Jan 19, 2025· 1 reaction

    ¡Graciass! Dime si quieres que te explique mejor.