If you find our articles informative, please follow me to receive updates. It would be even better if you could also follow our ko-fi, where there are many more articles and tutorials that I believe would be very beneficial for you!
如果你觉得我们的文章有料，请关注我获得更新通知，如果能同时关注我们的 ko-fi 就更好了，那里有多得多的文章和教程! 相信能使您获益良多.
GRID stands for Game Roles Instant Design. This series aims to assist in game character design. Its training method, image size, and usage are different from what we usually use in normal. Please keep this in mind. I recommend using it with a batch, which is also the implied mean of GRID.
Chibi is the first work in the Game Dev Tools/GRID series, and it's a 3-in-1 concepts LoRa. First, let's understand what "3-in-1" means. Normally, a LoRa one main concept, such as a character, costume, or style. But in this LoRa, there are three sub-concepts under one big concept, which is chibi. The three sub-concepts are chibi, chibi monster, and chibi mecha. These are also the three trigger words for this LoRa. They can be used separately or in combination (I'll explain this advanced usage later).
As always, at AIGARLIC, our goal is to integrate and optimize workflows in actual production environments by incorporating AI into traditional workflows. Therefore, the training goals and usage of this LoRa are different from regular. If you only need to draw some characters, many base models or LoRas can do that, including Midjourney, which is good at it. However, finding and trying out models, trying different prompts, and inconsistent styles can be time-consuming and energy-consuming. Sometimes, it's difficult to achieve the desired results or the quality is unstable (with Midjourney). These are often unacceptable for actual production processes. Therefore, the design and training of this series follow the following principles:
Speed: On a 3060 (12G) graphics card, it takes 13-18 seconds to generate 9 small samples, with an average of 1.5-2 seconds per image. This allows artists to focus more on choosing and painting, rather than fiddling with AI.
Simplicity: Usually, adding one prompt can achieve the desired features without too much trial and error.
Consistent Style: Using the same checkpoint produces a consistent style.
Variability: This is a big topic, but we have tested the following types of variability:
Variability based on the base models: checkpoint + LoRa = stable style
Variability based on the prompt: trigger words + prompt = character with LoRa features + prompt features
Variability based on the character: trigger words + character prompt + description prompts = variable character
This LoRa aims to integrate AI into the artistic design workflow, serving as a reference and assisting with some drawing work (when paired with ControlNet, it can produce line drawings) and producing color thumbnails, helping artists reduce some of the repetitive daily work.
The complete version (18 pages) of this manual has been uploaded to our ko-fi. We are currently in the process of moving from Fanbox, so please follow our new page. Thank you.