Hello everyone, I'm from China. "Creamy Mami" was a beloved animation of mine from my childhood. Due to its release many years ago, there are very few fan arts, and no artists creating them. Thanks to AI intelligence, someone like me who has no drawing skills at all can use AI to create.The currently released Lora is the first model I've learned to create on my own. There are many areas that need improvement, but I can't wait to share it with everyone. I'll continue to optimize the model in the future, but it may not happen very quickly.
The underlying model paired with Lora is AnythingV4.5. Another style with good results is Scmix. My suggestion is to set Lora's weight to 0.65.
I captured most of the costumes from the animation as training materials, but because it's my first Lora that I learned to create, I'm still not very familiar with many detailed adjustments, such as the quality of the materials, the quantity, and labeling the model to prevent overflow. So, Lora's control over clothing is still relatively weak, and triggering it is quite difficult. The following trigger words are the ones I experimented with and found to be more effective. Everyone can try out other costumes, exchange experiences, and I will continue to improve the model."
there are some trigger keywords for clothing worn by the main characters in the animations.（for SD1.5）
japanese clothes, floral print kimono, flower, hair ornament, hair flower,
hat, red beret, green sunglasses,pink dress, dress,turtleneck, belt, long sleeves,
gloves, dress, red dress,flower,puffy sleeves, puffy short sleeves,legs, bare legs,high heels,
school uniform,serafuku,sailor collar,yellow skirt, ,shoes, socks, red footwear,
swimsuit, bikini, red bikini, hairband
leotard, boots, choker, pink footwear, knee boots, pink leotard,
leotard, boots, choker, yellow footwear,yellow knee boots, yellow leotard,
dress, pink dress,sleeveless, bare shoulders,wrist corsage,ribbon,frills,red footwear,tiara,
coat,red coat,fur trim,
flower,wedding dress, rose, gloves, jewelry, pearl necklace, bridal veil, frills, hair flower,bow, hair ornament,bilateral symmetry rose, bilateral symmetry pink flower, bilateral symmetry hair flower,
retro_artstyle, green_pajamas,pajamas,
Description
FAQ
Comments (11)
"Thank you for the reminder. I just found out that someone had already created this model before. The search function on this website is really poor, but it has its benefits. It forced me to learn how to create and improve models, gaining experience. Ha ha, most of the time, I enjoy everyone's contributions, and this time I can also contribute my own effort."
@creamyws Well, in this case Is due to japanese creator refusing to use english in the title.
May be you can create it also a SD XL Pony version?
@settima_ai Currently, my main focus is still on optimizing the existing LoRa in my spare time, because I'm not very satisfied with it. It's only a semi-finished product, lacking a sufficient variety of clothing, and having weak control over clothing and facial expressions. The version I've currently released is the fifth version. The first four versions had many flaws. Although the second version didn't have as rich clothing options, it produced great results when drawing kimono images. Surprisingly, the currently published fifth version, despite the addition of clothing materials and more precise prompts, doesn't have such good quality. This has left me puzzled and is the direction I aim to learn and improve upon next.
@settima_ai I am aware of the SDXL version, but currently, the SDXL ecosystem is not quite mature yet. I haven't figured out SD1.5 yet, so for now, I'll focus on optimizing LoRa 1.5. Once the SDXL model and ecosystem become more complete, I'll consider porting it to SDXL. After all, SDXL has significant advantages such as higher output resolution, better finger and posture movements, and better accuracy in generating images. Creating SDXL is just a matter of time. Once I've optimized the current LoRa, my next step might be to delve into SDXL. Next, I also want to create a long-haired version of Ranma 1/2's Shampoo. It seems that Civital only has a short-haired version.
@creamyws Yeah, I was a great user of SD 1.5, with hundreds of checkpoint uploaded in my Google Colab account. Unfortunately, recently I cancelled my "Google Gemini AI" subscription, they got confused and cancellled also my Google Colab! As result, at the moment, I can use only the CivitAI online generator. I posted an image but there is no Hires Fix, no Face detailer, no upscale... with SD 1.5 this is very basic and poor, sorry...
@settima_ai I have always been using SD1.5 in a local deployment manner, including making LoRa and creating images. The hardware I use is 4070ti. Apart from the VRAM, I am satisfied with everything else. I believe that in the future, there might be more restrictions due to laws and copyrights. Local deployment and robust hardware might become necessary for us; otherwise, we might suddenly face restrictions at any time. I have seen your work; the images are great, especially the facial reconstructions. Controlling the clothes is a bit challenging due to LoRa. "Creamy Mami" is originally an 80s work, remembered by few, and even fewer are interested in creating works based on it. It's only at times like these that I can feel the spirit of the early internet when people around the world shared their achievements, exchanged experiences, and solved problems without reservation.
@settima_ai I just posted a piece of work comparing screenshots from animations using LoRa. AI create the image based on the storyboard. The effect is excellent. The resolution of the screenshots from the animation is very low, and there are no details. After the AI creates the image, it becomes very clear. It's a good creative approach.
@settima_ai I produced a new SDXL model yesterday. Unfortunately, SDXL has very high hardware requirements. I had to delete a lot of training materials to reduce the training time. If I didn't, I would have faced nearly 36 hours of training time. The final result wasn't actually good. Initially, I attempted training with the pony model, but the produced model was difficult to achieve the desired effect. Eventually, I switched models and used Animagine to train besides the SDXL model, which yielded slightly better results. However, compared to SD1.5, there doesn't seem to be a significant improvement, except for better drawing of the hands. The images trained with the new model look more like screenshots from an animation, whereas SD1.5's effect is more like a remastered version of the new technology. I'm quite disappointed and don't know how to solve the problem. It's very distressing.
@creamyws Yeah, i agree the SD Xl is ammature and quite disappointing. May be a question of CheckPoint still to train well, expecially for anime...
Anyway, based on my exeriments, is promisling... not only hands, but clothes details are much better, ad also it "knows" much more concepts...
@settima_ai The SDXL version released last night, based on the animaginV3.1 model, gave me hope for improvement. So, last night, I made the decision to train the computer with the complete training set. It has been running since last night and has not yet finished. It's expected to continue until around 1 a.m. tomorrow morning, making the entire training time approximately 26 hours. Therefore, I won't engage in any creative work tonight. Tomorrow night will be the moment of witnessing miracles. I hope to achieve a good result. Currently, I believe the advantages of SDXL lie in better human body and finger structures, higher completion rate of produced works, and lower error rate. Although there may still be some errors in human body construction and the number of fingers, whether too many or too few, compared to SD1.5, it has improved significantly. Tomorrow night, I should release some new works. If the results are good, I will swiftly release an improved version.