LEOSAM's HelloWorld XL

Many improvements and attempts have been made in the production process of this version. The main improvements are listed one by one for reference:<ol><li>Further selections were made to the training materials, but the total volume is still maintained at the scale of 500 training sets + 1500 regular sets. The proportion of full-body photos, male photos, high-definition texture photos, and photos of different races has been increased.</li><li>The word library used for clip labeling originally contains about 110,000 phrases, but there are a large number of errors, garbled codes, and repeated phrases. With the help of GPT4 batch modifications and multiple rounds of test labeling and manual addition and deletion, this word library has been reduced to 40,000 words, and a large number of phrases related to photography, portraits, and China have been added.</li><li>A large number of comparative tests have been conducted. Including a. The difference between training the SDXL under the same training set with dreambooth and first training with SDXL lora and then merging into the large model; b. The training effect differences under adafactor, adamW8bit, prodigy three optimizers, different LR schedulers, different learning rates, different batch sizes; c. The effect differences when different training set data enhancement methods are used and not used; d. The training effect under different SDXL base models.</li><li>Batch image processing was performed before the training set was bucketed, compressing and cropping the training set and putting it into the target resolution groups of (768, 1360),(832, 1248),(864, 1184),(1024, 1024),(1184, 864),(1248, 832),(1360, 768). This improves the subsequent large batch size training effect (but the improvement seems limited).</li></ol>Above are the main updates for HelloWorld 2.0. There were quite a few challenges when updating this version, but the good news is that I've figured out the way to train the SDXL large model, so future updates should be much smoother.该版本在制作过程中在多方面进行了改进尝试。主要改进逐一列举如下：<ol><li>对训练素材进行了进一步的增减精选，但总量仍维持500训练集+1500正则集的规模。增大了全身照、男性照、高清质感照片以及不同人种照片的比例。</li><li>clip打标所用的词库本身有约11万词组，但其中存在大量错误、乱码与重复词组。借助GPT4批量修改，以及多轮测试性打标人工增减，将该词库缩减至4万词规模，并大量补充了与摄影、人像、中国相关的词组。</li><li>进行了大量的对比测试。包括a.同训练集下dreambooth训练SDXL大模型与先SDXL lora训练再合并入大模型的效果差异；b.adafactor、adamW8bit、prodigy三个优化器、不同LR scheduler、不同学习率、不同batch size下的训练效果差异；c.不同训练集数据增强方法使用与未使用时的效果差异；d.不同sdxl底模下的训练效果。</li><li>对训练集进行了分桶前的批量图像处理，将训练集压缩裁剪并归入 (768, 1360),(832, 1248),(864, 1184),(1024, 1024),(1184, 864),(1248, 832),(1360, 768)这7个目标分辨率组别。以提高后续大batch size下的训练效果（但感觉提升有限）。</li></ol>以上就是HelloWorld 2.0版本的主要更新内容，这个版本在更新过程中属实踩了太多坑，好处是摸到了sdxl大模型训练的门道，以后的更新应该会顺利很多。

leosamsHelloworldXL_helloworldXL20.safetensors

HelloWorld XL 2.0

HelloWorld XL 7.0

HelloWorld XL 6.0

HW5.0_Euler_a_Lightning

HelloWorld XL 5.0 GPT4V

HelloWorld XL 4.0

HelloWorld XL 3.2 DPO

HelloWorld XL 3.0

HelloWorld XL 1.0

Film Grain 2.0

Film Grain 1.0

Utopia 3.0

Utopia 2.0

Utopia

Reality 3.0

Fantasy2.0

Reality2.0

Reality

Fantasy

Files with SHA256: 713cf528911e5ccd259c7d6871abefb067f28ef713cf37ece1989927c0720be3