Donation: https://ko-fi.com/heart_of_apple_xl
All funds will be used to expand the dataset and continue training.
Heart of Apple XL will remain completely free!
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 locally for artist style representation yet, supporting approximately 700+ artist tags from Danbooru/Pixiv (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 V3 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 Resolution: 256
Maximum Bucket Resolution: 3072
Full FP16
Epochs: 10
Optimizer: Adamw 8bit
Learning Rate: 1e-5
Gradient Accumulation Steps: 12
Usage Guide
Similar to Animagine XL V3, tag ordering is recommended but not required, with a recommendation to use artist tags.
*Recommended Sampler/Steps: Euler A with a step count between 20-28.
Quality Prompting
Quality Tags: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality, bad quality
Rating Tags: safe, sensitive,