Dream Mirror： CG Master

This is an attempt at Chinese-themed content. Currently, the font Lora, used for traditional Chinese painting and other traditional Chinese styles, has been studied by many friends. However, it seems that there haven't been many strong narrative CG-style Chinese themes. Due to the needs of my own project, I have created this Chinese style CG master Dan. The training materials are mostly top CG works in China, but I won't list the artists due to the current hostility toward the AIGC community in the CG industry. Their works are mostly based on Eastern fantasy, mythology, and other themes, mainly focusing on scenes with few characters.

This model is suitable to be used in conjunction with mature models like revAnimated and chilloutmix to present a good narrative effect with a focus on characters. It also works well with more comprehensive base models like Anything_v5 and final-pruned for scene types.

I like to call it YaoDan.

Due to my personal interest in Chinese mythology themes, I often have continuous thoughts about Buddhas, demons, and mythical creatures. Currently, there is serious homogenization on Midjourney, lacking some personalized aesthetic expressions. For example, the support for the most imaginative and expressive Chinese mythological themes is very low, so the privately trained model is a necessity. This model still has many shortcomings, such as the high probability of the details such as faces and hands breaking down, partly because the training of this model is focused on strengthening Eastern narrative scenes, which makes characters secondary. Furthermore, when characters appear in scenes, they are often not the main narrative elements, resulting in insufficient pixels to represent them adequately. It is well-known that this leads to a lack of detail in character modeling.

Therefore, adjusting the camera to close-ups or mid-range shots can greatly improve this issue. For narrative purposes, you can also improve this weakness through the smart mode of controlnet, using tile + Ultimate SD upscale to enlarge while re-drawing details, which can basically solve this weakness.

Its advantages are also quite notable, as it finally provides strong narrative abilities for Chinese-themed content while also balancing the Eastern aesthetic's unique imagination. At the moment, it can indeed solve many of my work-related issues. At least it is a good middle optimization step.

The above is just a bit of personal experience sharing. Welcome feedback from everyone and discuss more solutions.

Note:

The range I usually use for CFG is 5 to 7.5, and the Lora weight is usually between 0.6 and 0.9.When Clip Skip is set to 1 or 2, the generated results will be slightly different. You can try and see which style you prefer.The training set of this Lora model mainly consists of works by Chinese CG masters, so generating effects like European and American portraits and buildings may be impacted.

Note:

Principally, this model is only for educational purposes. Please do not use it for commercial purposes.All content generated using this model is the individual creation judgment of the user. Please comply with relevant laws and regulations.Any intellectual property issues arising from the use of this model are the responsibility of the user.Without the written consent of the author, this model's services may not be integrated into any third-party software or service. Furthermore, you may not disclose the association between the generated content and this service to any third party, directly or indirectly, without authorization. Any losses incurred to the author due to your violation of this agreement will entitle the author to compensation from you.

Disclaimer:

This model is essentially designed for learning and communication only and should not be used for commercial purposes.All content generated by this model is the result of individual creativity and should be used according to your discretion and conform with relevant laws and regulations.Users are solely responsible for any intellectual property issues arising from the use of this model.Without the explicit written consent of the author, this model should not be integrated into any third-party software or services, nor should the association between the generated content and this service be disclosed to any third party, directly or indirectly, without authorization.