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

LEOSAM HelloWorld | SDXL Realism Checkpoint

2023.11.10 HelloWorld 2.0 UpdateEveryone has been waiting for this. After going through various challenges and setbacks, HelloWorld 2.0 can finally be presented in a satisfactory state. The major differences between HelloWorld 2.0 and 1.0 versions are as follows:1. HelloWorld 2.0 performs equally well as the 1.0 version with trigger words when no* ***ng trigger words. The trigger word leogirl in version 1.0 was strongly associated with East Asian ethnicity. After removing trigger words, although high-frequency words like 1girl still often generated East Asian figures without specifying ethnicity, you can nowuse nationality, skin color, and other cues to specify ethnicity. For example, triggering effects for words like Chinese, Russian, Iranian, Jamaican, Kenyan, dark-skinned, and pale-skinned are listed below.<img src="https://liblibai-online.liblib.cloud/web/image/cd9afc9474c1ad9f74f7afd7c8f1ec0501a39c54718f209ad06dba8a2e06cc1f.jpg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="web/image/cd9afc9474c1ad9f74f7afd7c8f1ec0501a39c54718f209ad06dba8a2e06cc1f.jpg" data-href="https://liblibai-online.liblib.cloud/web/image/cd9afc9474c1ad9f74f7afd7c8f1ec0501a39c54718f209ad06dba8a2e06cc1f.jpg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style="width: 435.14px;height: 192.94px;"/>You can alsouse names from different countries and genders to obtain characters with different styles, such as Han Meimei (China), Sophie Martin (France), Priya Patel (India), Fatima Al-Hassan (Arab), and Wanjiru Mwangi (Kenya). These are just examples. There are many more cues and ways to play around with. Feel free to explore and share.<img src="https://liblibai-online.liblib.cloud/web/image/14806c020787ad3c72585fa63a8579dd77a686cb0884e3ac0aaedf84773c2fe1.jpg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="web/image/14806c020787ad3c72585fa63a8579dd77a686cb0884e3ac0aaedf84773c2fe1.jpg" data-href="https://liblibai-online.liblib.cloud/web/image/14806c020787ad3c72585fa63a8579dd77a686cb0884e3ac0aaedf84773c2fe1.jpg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style="width: 443.1**x;height: 161.44px;"/>2. HelloWorld 2.0 balances image quality/color, offering more diverse style options. When used with leogirl in version 1.0, images often had a strong vintage feel. HelloWorld 2.0 no longer enforces this vintage feel and allows for customization through quality-related cues. Some words that have shown good triggering effects include:high-end fashion photoshoot,product introduction photo,popular Korean makeup,aegyo sal,Sharp High-Quality Photo,studio light,medium format photo,Mamiya photography,analog film,Medium Portrait with Soft Light,real-life image,refined editorial photograph,raw photo,real photo,Scanned Photo,film stillThese words have different color effects as shown below:<img src="https://liblibai-online.liblib.cloud/web/image/025fe7f55600e1912bda445087725b3fcc1408ce3d771701f78ba13bc4adc40a.jpg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="web/image/025fe7f55600e1912bda445087725b3fcc1408ce3d771701f78ba13bc4adc40a.jpg" data-href="https://liblibai-online.liblib.cloud/web/image/025fe7f55600e1912bda445087725b3fcc1408ce3d771701f78ba13bc4adc40a.jpg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style="width: 445.1**x;height: 876.81px;"/>3. HelloWorld 2.0's training set has significantly increased the proportion of full-body shots, aiming to improve the results when generating full-body or long-distance portraits. Although there has been improvement compared to version 1.0, there is still a strong suggestionto avoid using adetailer during full-body image generation. For players with sufficient memory (24g), it issuggested to enhance image resolution by 1.5 times, which significantly improves facial details. During the development of this version, improvements and experiments were made in various aspects. While some improvements may not be obvious, the main enhancements are listed for reference:1. The training materials were further refined, with a total of 500 training sets + 1500 regular sets. The proportion of full-body shots, male portraits, high-quality images, and images of different ethnicities has been increased.2. The word library used for trigger words originally contained about 110,000 phrases, but many errors, garbled text, and duplicate phrases were present. By utilizing GPT4 for batch modifications and multiple rounds of testing with manual addition and removal of tags, the word library was reduced to about 40,000 phrases. It also added a large number of phrases related to photography, portraits, and China.3. Extensive comparison tests were conducted, including a. different training effects between training the SDXL large model with dreambooth under the same training set and training SDXL lora first before merging into the large model; b. differences in training effects under different optimization methods such as adafactor, adamW8bit, prodigy, different LR schedulers, learning rates, and batch sizes; c. differences in training effects with and without different data augmentation methods; d. training effects with different SDXL base models.4. The training set underwent batch image processing before binning, compressing and cropping images into the following 7 target resolution groups to improve training effects under large batch sizes (although improvements were limited): (768, 1360), (832, 1248), (8**, 1184), (1024, 1024), (1184, 8**), (1248, 832), (1360, 768).These are the main updates in HelloWorld 2.0. While this version encountered many challenges during the update process, the upside is the insight gained into training the SDXL large model, making future updates smoother.The new version has made progress compared to 1.0, but there are still areas for improvement.Feedback and suggestions are welcomed for continuous improvement, as they are crucial for the model's progress! Thank you for your long-standing support!-------------Welcome to the SDXL large model HelloWorld new world! After two weeks of trial and error and debugging, I am finally able to release this version that I am quite satisfied with! I believe HelloWorld new world can greatly improve the portrait capabilities of SDXL.Compared to SD1.5, the SDXL model has made a qualitative leap in semantic understanding of trigger words and image information. This brings four differences in usage from the traditional SD1.5 model:1. HelloWorld supports direct output of 1024*1024 resolution images without the need for high-resolution enlargement. Currently, it surpasses the SD1.5 version in quality for close-up portraits, but flaws still exist in distant portraits. Therefore, it is recommended to use the ADetailer function in lib online image generation to correct distant facial issues.2. Unlike the SD1.5 large model, which usually does not include trigger words,make sure to use the trigger word leogirl when using Hello World to ensure the SDXL model triggers the training set effects more stably.3. You can now use simpler cue words for image output, and it is also recommended to trynatural language prompts, as they lead to better results when producing realistic AI photos. You can combine usage with large language models like chatGPT, enabling GPT to compose fluent descriptions from scattered label-style cues or to describe desired scenes in English.4. SDXL greatly expands the scope of stable diffusion in virtual photography, with the large amount of information in the base model supporting the output of many movie poster-level AI images. I will showcase some sample images in the sharing area. I hope that while continuing the tradition of generating beautiful images with SD1.5, you can also explore and experiment with SDXL's application in more virtual photography domains. Moreover, liblib is currently hosting a virtual photography competition. I can confidently say that in the field of lighting and shadows, the SDXL model deals a heavy blow to the SD1.5 model, therefore, I encourage all drawing enthusiasts to actively participate using this model! The making of HelloWorld new world encountered many challenges and received help from many experts. In-depth collaboration and communication with the author of XXMix, YeyaoXiao, were a crucial step for the model's success. Dragon expert's development of the DB training script based on Kohya's training package significantly reduced my testing time. I also gained valuable insights through exchanges with xiaolxl, the author of the national wind 4 model. I am grateful for all the help!There are still many areas for improvement in HelloWorld. I welcome your feedback and suggestions in the comment section, as they are essential for the model's future improvement! The model is now available for download, and the 2.0 version is currently in intense production, so stay tuned!In conclusion, welcome to Hello World new world, and I wish you a pleasant journey in the virtual universe created by SDXL!! 🤝Lastly, a little advertisement: If you have model customization or other collaboration needs, please contact me via WeChat: 13439222668. Please briefly state your purpose as time is valuable and no chitchat, thank you for your understanding!

Files with SHA256: 713cf528911e5ccd259c7d6871abefb067f28ef713cf37ece1989927c0720be3