Gu Simiao, a splendid rabbit

Essential trigger words for models: s4s the pallas's cat. Other suggested hints include: manul, 4K HD hi-res photo, realistic Hasselblad photography

Model Recommended weight: 0.6~0.8

Model recommended parameters: Sampler: DPM++ 2M Karras or Restart, CFG scale: 7~10, size≥1024x1024

, steps≥30

Heartfelt words: Lu

Simiao (2015-2022.10.10) is a male

rabbit rescued from the wild and used to live in Xining Wildlife Park. It is the most famous rabbit in China.

Thanks to Nishino Yuanzhang's continuous popularization of science day by day, and the great charm of Lu Si-yu and her family, rabbits have gradually become a star species in recent years. Along with species such as snow leopards and desert cats, they have greatly increased people's attention to wildlife conservation and ecological construction on the Qinghai-Tibet Plateau.

The name comes from the fact that its first observed mating behavior lasted only 4 seconds, which gave it the homonym the name. Its life began with a ridiculous name and ended with another method of ridicule. On October 10 of last year, he died because he was too anxious to eat chicken and suffocated by chicken bones stuck in his throat. Normally, the life span of captive rabbits is around 12-15 years old. At 7 years old, I am still in adulthood. My friends who love rabbits were all shocked and saddened at the time

.

In February of this year, I first came into contact with the LoRa model, and the first model I made was the rabbit LoRa. As an unpopular species, whether it's Midjourney or the official SD model, it hasn't been possible to accurately output the image of rabbit rabbit, and LoRa is an excellent way to fine-tune this niche concept. My first edition Lora the rabbit was a generic model similar to FilmGirl. It used 300 different individual pictures of rabbit puppies for training. After I finished this model, my second LoRa model was to create an exclusive image LoRa. For this reason, I collected a total of 26* *hotos of from my friends, 70% of which are from @yspenny, 30% from @西宁野生动物园 , @圆掌, and @天音文创馆.

However, the results were not good. There are three main reasons: 1. The vast majority of photos are taken by mobile phones, cropped from a distance, and generally have blurry image quality and automatic color filters.

2. The glass of Nishino Kitten House is greenish and highly reflective, further deteriorating the image quality.

3. Lu Si-yu has lived in a small exhibition room in the Kitty Cat Hall all her life, and the scene is solidified in a single place.

Due

to the above reasons, the model rendering rate is low, and a large number of roll graphs are required. As a result, I did not publish the version of I also posted on Weibo to ridicule that this large amount of screening was the reverse fine-tuning and brainwashing of me by AI, distorting my mind's memory of what Yu Si-yu really looked like, so I had

to sift through the actual photos for a while to wash my eyes.

But now, there are new ways to solve these problems. Although it still faces many shortcomings, the SDXL model is unquestionably more powerful and has a higher ceiling. Just before the first anniversary of's death, I decided to make another SDXL LoRa model with the exclusive image of Si-yu.

Faced with the shortcomings of the SD1.5 model, I tried the following improvements: 1. Considering that the SDXL model was extremely sensitive to the quality of the training set, but the quality of the original training set was uneven and the image quality was generally blurry, I re-selected the training set to 224. While eliminating a batch of images, AI enhanced the image quality of some salvable training maps.

2. Added 1006 selected individual photos of rabbits from around the world to the regularized collection. The training used a mixed natural language+label marking method to test the triggering effects of different trigger words in order to properly distinguish and correlate the training set with other rabbit regularized sets. While trying to avoid polluting individual appearance

, use regularized sets to enhance hair details and generalize background environments and facial expressions.

3. The four optimizers, adam8bit, DadaptAdam, Prodigy, and AdaFactor, were used separately for training. The output models were compared and fused, and the final version with the best balance in

the three dimensions of color detail, generality, and similarity was selected as the official version of LoRa.

I was satisfied with the final completion results. Although there are still common problems with SDXL's blurred vision, the generality of the model in terms of color, attitude, and environment has improved significantly compared to the model.

The above is the complete journey of the Lou Si Miu LoRa model. This is bound to be a niche model that only a few people will use regularly, but its value in my eyes is no less than any other model I have made, because this model supports me and the photographers who provided the training maps in memory of Lu Si-yu. The field is short, and the living area is brilliant. Being in a poor Northwest Zoo didn't take many good pictures, but that didn't prevent you from surviving in the game forever. It's been a year and we miss you so much. I wish you to eat slowly in Xing, and everything

is fine.