Leosam's Clothing +/- Adjuster Clothing Increase/Decrease LoRa

🤝 If you have model customization or other cooperation or consultation needs, please contact WeChat: 13439222668. Please simply indicate your intention

. Time is precious and don't chat casually. Thank you for understanding!

This is a functional LoRa that adjusts how much of the object's clothing is being drawn. By adjusting the LoRa weight from 1.0 to -1.0, it is possible to gradually increase the amount of clothing to be drawn. This model works best when generating vertical images

.

This model is the first method introduced by the Blue Dragon boss in this video, “Learn by copying Inspired by “law”. I've made some improvements to enable batch training with multiple graphs. The specific steps are as follows: Step

1: Select comparison images of state A and state B of different characters in N groups, and form a state A image training set and a state B image training set. Ensure that the file names of the two images in State A and State B of the same person are the same

.

Step 2: Add a txt tag to the training set in state A. Each image only has a special vocabulary tag that can distinguish different characters. For example, if there are 10 characters, then each character is assigned their own label from jinitaimei1 to jinitaimei10. Then copy and paste all the tags of the training set in the state A training set into the state B image training set

.

Step 3: Select base model C, which is similar in style to the training set, and use the state A image training set to perform Lora training until the model fits. After entering the corresponding label of character N, only a state A photo of character N can be generated.

Step 4: Blend the trained overfitted LoRa model into bottom model C at a ratio of 1.0 (update: after further testing, checking the same to strength will work better), then use the state B image training set to perform Lora training based on the new bottom model. The training process does not have to be seriously overfitted. LoRa process files can be selected for AI drawing tests, as long as the transition from state A to state B can be achieved by adjusting weights. (Update: After further testing, the degree of fit is still higher, and Lora's effect is

better.)

Step 5: If the LoRa training scenario is complicated, fitting will occur under high weights. There are two suggestions for improvements: one is to perform LoRa hierarchical adjustment to reduce the weight of the number of layers unrelated to A/B state switching in LoRa; the second is to compress the dimensions of LoRa

, such as compression from 64 to 4.

Improved skin texture at weight 1

through layered fusion Improved skin texture at weight 1

with layered fusion