Heart of Apple XL V3 is trained with a curated dataset of 300,000 images from Kohaku-XL Epsilon. It may be the best model for artist style representation yet, supporting approximately 700 or more artist prompts from Danbooru (possibly more).
Heart of Apple XL V1 and Heart of Apple XL V2 were based on the Animagine XL V3 model, showing decent performance capabilities. V3, V2, and V1 all utilized similar training methods as Animagine XL V3, such as tag ordering. Compared to Animagine XL V3 and Kohaku-XL Epsilon, Heart of Apple XL offers a more diversified representation of styles, especially for some game characters and artists.
Heart of Apple XL V3 has been training on an A100 40G server for 14 days and will continue to be updated. I'd like to thank CagliostroLab,琥珀青叶, and the community for their contributions, as well as 阿萨姆奶茶，小苹果，青秋，zwh20081，alens,原物质,灾厄之轮，hans,Chenkin and others for their help with technology or datasets, and those who encouraged me.
Thanks to the following sponsors for their financial support: 阿萨姆奶茶,yZz, 原物质 ,离歌 ,阿星, dp,星河主炮向前，4421,1lch2,b74, StyxHua
Training Details:
A manually curated dataset of 300,000 images from Danbooru and Pixiv, trained using SD-Trainer with tag reordering.
Resolution: 1024x1024
Hardware: A100 40G Batch
Size: 18
Minimum Bucket : 256
Maximum Bucket Resolution: 3072
Full FP16
Epochs: 10
Optimizer: Adamw 8bit
Learning Rate: 5e-1
Gradient Accumulation Steps: 12
Usage Guide
Similar to Animagine XL V3, you need to order the tags, with a recommendation to use artist prompts.
Quality prompts Quality Tags: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality, quality bad, hands bad, eyes bad, face bad Rating Tags: safe, sensitive, nsfw, explicit Date Tags: newest, recent, mid, early, old It's recommended to place quality prompts at the end or use artist prompts only.
*Recommended artist prompts (These artist prompts usually produce better results than others)
Effect prompts Flipanime, god light, mmd style, sketch
QQ Group: 244078420 This is a new QQ group I created for users to discuss and provide feedback to improve the model's usage, and also where other models I've developed are discussed.
Heart of Apple XL V3是基于Kohaku-XL Epsilon，人工挑出来的30万张数据集进行训练。Heart of Apple XL V3可能是艺术家风格表现最好一个模型,实际测试支持将近大约700个以上来自danbooru的艺术家prompt （可能更多）。
Heart of Apple XL V1 和Heart of Apple XL V2是一款基于Animagine XL V3的模型,拥有较为不错的表现能力。V3 V2 V1都使用和Animagine XL V3大致相同的方法进行训练，如对tag进行排列。相比于Animagine XL 3, Kohaku-XL Epsilon ，Heart of Apple XL在风格表现方面更加多元化，对一些游戏角色和画师进行了专向训练。
Heart of Apple XL3目前在a100 40g的服务器上训练了14天。未来还会继续更新的，再次感谢CagliostroLab，琥珀青叶，对社区的付出，同样我也要感谢阿萨姆奶茶，小苹果，青秋，zwh20081，alens,原物质,灾厄之轮，hans,Chenkin，等人在技术或数据集帮助，还有那些鼓励我的伙伴们。 感谢以下赞助者的资金支持：阿萨姆奶茶,yZz, 原物质 ,离歌 ,阿星, dp,星河主炮向前，4421,1lch2,b74, StyxHua
Batch Size:18 Min Bucket Resolution: 256 Max Bucket Resolution: 3072 Full FP16 Epoch:10 优化器:Adamw 8bit 学习率：5e-1
gradient accumulation steps:12
使用指南 和Animagine XL V3一样你需要对标签进行排列，建议使用艺术家prompt。
质量prompt
Quality tags: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality,quality bad, hands bad, eyes bad, face bad Rating tags: safe, sensitive, nsfw, explicit Date tags: newest, recent, mid, early, old 建议正面质量prompt放在后面，或者仅使用艺术家prompt
Introduction
Heart of Apple XL V3 is trained with a curated dataset of 300,000 images from Kohaku-XL Epsilon. It may be the best model for artist style representation yet, supporting approximately 700 or more artist prompts from Danbooru (possibly more).
Heart of Apple XL V1 and Heart of Apple XL V2 were based on the Animagine XL V3 model, showing decent performance capabilities. V3, V2, and V1 all utilized similar training methods as Animagine XL V3, such as tag ordering. Compared to Animagine XL V3 and Kohaku-XL Epsilon, Heart of Apple XL offers a more diversified representation of styles, especially for some game characters and artists.
Heart of Apple XL V3 has been training on an A100 40G server for 14 days and will continue to be updated. I'd like to thank CagliostroLab,琥珀青叶, and the community for their contributions, as well as 阿萨姆奶茶，小苹果，青秋，zwh20081，alens,原物质,灾厄之轮，hans,Chenkin and others for their help with technology or datasets, and those who encouraged me.
Thanks to the following sponsors for their financial support: 阿萨姆奶茶,yZz, 原物质 ,离歌 ,阿星, dp,星河主炮向前，4421,1lch2,b74, StyxHua
Training Details:
A manually curated dataset of 300,000 images from Danbooru and Pixiv, trained using SD-Trainer with tag reordering.
Resolution: 1024x1024
Hardware: A100 40G Batch
Size: 18
Minimum Bucket : 256
Maximum Bucket Resolution: 3072
Full FP16
Epochs: 10
Optimizer: Adamw 8bit
Learning Rate: 5e-1
Gradient Accumulation Steps: 12
Usage Guide
Similar to Animagine XL V3, you need to order the tags, with a recommendation to use artist prompts.
Quality prompts Quality Tags: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality, quality bad, hands bad, eyes bad, face bad Rating Tags: safe, sensitive, nsfw, explicit Date Tags: newest, recent, mid, early, old It's recommended to place quality prompts at the end or use artist prompts only.
*Recommended artist prompts (These artist prompts usually produce better results than others)
ask\(askzy\), omoné hokoma agm, fajyobore, chen bin, kani biimu, mikozin, kedama milk, rurudo, tab head, ru zhai, kgt(pixiv12957613), ogi pote, wlop, anmi, tianliang duohe fangdongye, ningen mame, mignon, nai3, konya karasue, ciloranko, sho_(sho_lwlw), kamarukogedago, kanechi, oldnuu/(nu-nyu/), nuu/(nu-nyu/), irodori warabi, hagimorijia, himitsu/(hi_mi_tsu_2/), abara heiki, as109, darkdolphrokun, dolphrokun, freng, ibara_dance, jyt, Mrito00, novelance, rella, reoen, tyakomes, liduke, betabeet, deadnooodles, Indol, jima, kanda_done, kasuga_haruhi, modare, Nekojira, shiratama(shiratamaco), terada_tera, toosaka_asagi, torino_aqua, yanyan_(shinken_gomi)
AI artist prompts
willy style, oilwillystyle, breadman style, nsfwResearcher style, alens style
Effect prompts Flipanime, god light, mmd style, sketch
QQ Group: 244078420 This is a new QQ group I created for users to discuss and provide feedback to improve the model's usage, and also where other models I've developed are discussed.
Patreon:https://www.patreon.com/HeartofAppleXL/collections
介绍
Heart of Apple XL V3是基于Kohaku-XL Epsilon，人工挑出来的30万张数据集进行训练。Heart of Apple XL V3可能是艺术家风格表现最好一个模型,实际测试支持将近大约700个以上来自danbooru的艺术家prompt （可能更多）。
Heart of Apple XL V1 和Heart of Apple XL V2是一款基于Animagine XL V3的模型,拥有较为不错的表现能力。V3 V2 V1都使用和Animagine XL V3大致相同的方法进行训练，如对tag进行排列。相比于Animagine XL 3, Kohaku-XL Epsilon ，Heart of Apple XL在风格表现方面更加多元化，对一些游戏角色和画师进行了专向训练。
Heart of Apple XL3目前在a100 40g的服务器上训练了14天。未来还会继续更新的，再次感谢CagliostroLab，琥珀青叶，对社区的付出，同样我也要感谢阿萨姆奶茶，小苹果，青秋，zwh20081，alens,原物质,灾厄之轮，hans,Chenkin，等人在技术或数据集帮助，还有那些鼓励我的伙伴们。
感谢以下赞助者的资金支持：阿萨姆奶茶,yZz, 原物质 ,离歌 ,阿星, dp,星河主炮向前，4421,1lch2,b74, StyxHua
训练细节：
人工挑出来的30万张danbooru和pixiv数据集，基于用SD-Trainer训练，并对tag进行重训排列。
分辨率：1024x1024
设备：A100 40G
Batch Size:18
Min Bucket Resolution: 256
Max Bucket Resolution: 3072
Full FP16
Epoch:10
优化器:Adamw 8bit
学习率：5e-1
gradient accumulation steps:12
使用指南
和Animagine XL V3一样你需要对标签进行排列，建议使用艺术家prompt。
质量prompt
Quality tags: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality,quality bad, hands bad, eyes bad, face bad
Rating tags: safe, sensitive, nsfw, explicit
Date tags: newest, recent, mid, early, old
建议正面质量prompt放在后面，或者仅使用艺术家prompt
*推荐的艺术家prompt (这些艺术家prompt通常比其他艺术家prompt效果明显)
ask (askzy),omone hokoma agm, fajyobore, chen bin, kani biimu, mikozin,kedama milk,rurudo,tab head ,ru zhai, kgt\(pixiv12957613\), ogi pote, wlop, anmi, tianliang duohe fangdongye,ningen mame,mignon,nai3,konya karasue ,ciloranko,sho_(sho_lwlw）, kamarukogedago , kanechi, oldnuu/(nu-nyu/) , nuu/(nu-nyu/), irodori warabi，hagimorijia，himitsu/(hi_mi_tsu_2/), abara_heiki, as109 , darkdolphrokun , dolphrokun, freng, ibara_dance , jyt, Mrito00, novelance，rella，reoen，tyakomes，liduke，betabeet， deadnooodles， Indol，jima，kanda_done，kasuga_haruhi，modare，Nekojira，shiratama\(shiratamaco\), terada_tera, toosaka_asagi, torino_aqua,yanyan_(shinken_gomi)
ai艺术家prompt
willy style, oilwillystyle, breadman style,nsfwResearcher style,alens style
效果类prompt
Flipanime, god light, mmd style, sketch,
qq群：244078420
这是我新建的qq群，用于使用者讨论，和提出意见来改善模型的使用，并且这里有我炼的其他模型。
爱发电： https://afdian.net/a/willy23333123
Reference list&Resource list
animeloverbooru2022 · Datasets at Hugging Face 2023, huggingface.co, <https://huggingface.co/datasets/animelover/danbooru2022>.
cagliostrolab/animagine-xl-3.0 · Hugging Face 2024, huggingface.co, <https://huggingface.co/cagliostrolab/animagine-xl-3.0>.
Danbooru: Anime Image Board n.d., Danbooru，<https://danbooru.donmai.us/>.
KBlueLeaf/Kohaku-XL-Epsilon · Hugging Face 2024, huggingface.co, <https://huggingface.co/KBlueLeaf/Kohaku-XL-Epsilon>.
イラストコミュニケーションサービス [pixiv] 2019, pixiv, pixiv, <https://www.pixiv.net//>.
秋葉杏 2024, Akegarasu/lora-scripts, GitHub, <https://github.com/Akegarasu/lora-scripts>.