HelloWorld 5.0 Lightning Extreme Edition has been released. Equipped with the Eular-A sampler, it can generate images in 6-8 steps, triple the speed of the original version, with the same VRAM usage as the original. Welcome to try it out for friends with limited computing power.

Copyright Statement:

The HelloWorld series models (hereinafter referred to as "this model") are developed by me (hereinafter referred to as the "owner") with the support of the LiblibAI platform and exclusively launched on the LiblibAI platform as the XL large model. Except for the LiblibAI platform and Civitai platform, any individual or entity that releases this model on other platforms does so without the owner's authorization.

The owner authorizes individuals or organizations to use the images generated by this model for non-commercial educational or informational purposes for free, and:

- Comply with relevant laws and regulations, and do not infringe upon the legal rights of this model or any third party.

- When using images, please indicate the image source as "Generated by LEOSAM's HelloWorld large model".

For commercial use, a commercial authorization agreement must be signed with the owner first. For commercial authorization and model customization, please contact the owner through the contact information on the LiblibAI platform homepage.

The training and open-source sharing of SDXL models is a time-consuming, challenging task. The owner will continue to provide updates to the HelloWorld model for individual players for free to express support and gratitude to the community's open-source contributors. Commercial cooperation from business users is an important driving force for the development and continuous improvement of this model. Thank you for every user's understanding and support.

Please note that any unauthorized use may violate relevant laws and regulations and lead to legal liability. The final interpretation of this statement belongs to the owner and is subject to the relevant laws and regulations of the People's Republic of China.

2024.2.4 HelloWorld 5.0 Update

HelloWorld 5.0 is the largest update among all versions of HelloWorld, using GPT4V for labeling, and making significant adjustments in science fiction, animals, architecture, and illustration fields.

Through comparative testing, the advancements of this version include:

1. Richer, more dynamic, and visually appealing character poses and compositions;

2. The film dataset has been thoroughly trained. Versions 2.0 to 4.0 had weak film textures, but many users have been nostalgic for the Leogirl style from version 1.0. Therefore, the film texture has been significantly enhanced while trying not to affect other photographic textures. Users can trigger the film texture by using related phrases like "film grain texture" and "analog photography aesthetic";

3. Enhanced portrayal in science fiction, horror, animals, etc., with more artistry in mecha designs and more realistic depictions of animals like Pallas's cat, snow leopard, red panda, giant panda, tiger, domestic cats and dogs, etc.;

4. With the help of GPT labeling, prompt adherence and conceptual accuracy have been further improved.

The drawbacks of this version include:

1. Since it is a significantly tuned version, errors in body parts may slightly increase. This is a normal phenomenon when stepping out of comfort zones and entering new, locally superior points. Earlier versions underwent extensive limb testing improvements in previous locally superior points, while the new version had limited improvement time. Nonetheless, the limb accuracy of this version is at least higher than the 1.0 version, and I will continue to improve and optimize in future versions.

2. Due to the reinforced film texture, although GPT marking is as accurate as possible, default warm-toned images may still occur. However, by using terms like "studio light" and "sharp focus", users can output high-definition studio-quality images. With proper prompt usage, the studio-quality images output will have better skin tones and appearance compared to previous versions.

3. This version includes more full-body images to enhance the effects of full-body shots. Therefore, when no specific character composition is specified, the model's output will have a wider view compared to previous versions. Currently, the facial details of full-body shots at 1024 outputs will have some differences compared to half-body shots and close-ups. However, users can improve facial details by using "adetailer" and a 1.5x 0.3 intensity upscale, or avoid generating full-body shots by specifying compositions and related prompt terms.

4. Due to the inclusion of a small amount of high-quality illustration training data, there is a probability that prompts related to the anime-style will generate anime images. If this concerns you, please pay attention to adjusting the prompt terms.

These are the main contents of the version update. Training the SDXL large model is not easy. When the training dataset approaches ten thousand images, the cost of single labeling and computing power for each version exceeds 2000 yuan. Feel free to use the model and provide feedback! If you find this model satisfactory, please help spread the word about it, and I would be grateful!

Special Thanks:

Special thanks to Old Wang_Big Yiyeba Lang for sharing technical experiences in training multiple concept and large quantity image sets, which was crucial for the successful breakthrough of this version. Old Wang is an outstanding model author in the field of architectural photography, and I also invite everyone to try out his models.