Recap
2023-07-01 Hangzhou is cloudy on Saturday
Note: mmp, I ate grilled gluten last week and accidentally stuck a toothpick in my mouth. As a result, I feel uncomfortable eating anything these days, but I still have to make a plan. I will sort out the current AI applications in the second half of the year. I still like it. First use it and then learn the relevant principles, and then if you have time, learn to swim.
Mission statement
1 Origin
During the Dragon Boat Festival holiday, I went to Jinhua and Yiwu. Suddenly I felt lonely when traveling alone. Studying does not always make people happy. I have a bit of an adventurous spirit and am always a bit curious. I am willing to try anything. However, when traveling alone, you don’t have to compromise. This is the only advantage. But what is the meaning of life? It is to continue to increase yourself. The process of bonding. My definition of happiness is that the process of love is not so tiring, and that is enough; during the journey, I looked at the looming hills in the distance, and the rain stirred up the mist in the mountains, which was emerald green and hazy, with a bronze texture. Since After this, I will train a bronze lora and see what problems there will be in actual operation. After all, those who talk nonsense are those who have high vision but low hand;
2 Task arrangement
** Task**
- Prepare training set
- Training set labeling
- lora model training
- style test
process practice
1. Prepare training set
Material selection: Bronze
Time consuming: 5-30 min
Materials: 30-100 sheets
Instructions: Generally, you can prepare the material according to your own ideas. My hometown belonged to the Yin and Shang areas in ancient times, so the bronzes were the characteristics of that time, so I might as well just use them. This creates a style for the theme. In fact, I have thought that the design of icons in the UI can be completed through AI stylization, so that a unified style can be trained for product auxiliary design (there are too many bastards in the world, and I feel that I am amazing) , this world will eliminate you, you will never be polite);
2. Training set labeling
Explanation: This part is a very annoying process. First, the picture selection must be appropriate. Second, in order to make the lora you train have a clear theme, the selection of prompt words must match clearly, and you must check the selection yourself. The work we have done is consistent with It has absolutely nothing to do with painting, and all our efforts are outside the scope of painting;
3. lora model training
Note: There is a reference standard for parameters, but most of them still need to be adjusted and trained by themselves. This is alchemy. You still need a benchmark reference. Only by fine-tuning on a better basis can you achieve better training results.
4. Style Test
After the training is completed, the trained model will be retained under the project, and the trained model will be placed under its own category for easy call;
**Prototype architecture diagram**
** Vincent picture**
5. Training evaluation
With certain training results, it can be seen that there is a certain bronze style in architecture. I originally wanted to incorporate the bronze style into the architectural design, but it actually did not meet my expectations very well. Generally, the lines of house buildings are very Strong, but the bronze itself does not have such a strong sense of lines. If it is forcibly nested, it will feel a bit strange. It did not meet my psychological expectations, but it was also a good experience. Below are some photos of bronzes that have not been discovered by archaeology. Bar!
Summarize
After testing, it can be seen that if I use water cups, utensils, etc., the actual trained bronze model is still good. If I use the design drawings of buildings or modern vehicles as architectural input, it will lead to a situation where the head is wrong and the horse is wrong. Now many people like Training star characters, because as long as the character is assigned a name, there will not be much difference in appearance, and the posture transformation can also be random, as long as the characteristics are human. But bronze is different. The main training model of bronze is a vessel, so it cannot render characters very well. Therefore, training characters is the simplest among lora models. On the contrary, cross-style migration requires more data cleaning efforts.