Happy Meek (umamusume)版本LoHa (ID: 75986)

Happy Meek (umamusume)版本LoHa (ID: 75986)

LoHa of Happy Meek (umamusume). Put LoHa file in your stable-diffusion-webui/models/lora folder and write LoRA notation in prompt to applying LoHa. <lora:your_loha_name:weight>

All of my works is free. If you'd like to support me, feel free to buy me a coffee. ?

NOTICE

  • After careful consideration, I decided not to produce LoRA of the characters below.

  • If anyone wants the characters above, I will consider making them.

  • Except for the above, I will make the rest of the characters in Umamusume: Pretty Derby, unless there is a special notice.

  • Something weird happened when training. If you interested in training, read comments on LoRA training info section.

Recommended options

  • LoRA weight 0.6 ~ 0.7

  • Trigger words

    • Base

      • happy meek \(umamusume\)

    • Original Art :

      • hair flower, hairclip, striped bowtie, blue bowtie, brooch, jewelry, collared shirt, dress shirt, white shirt, blue vest, white jacket, open clothes, open jacket, puffy short sleeves, frilled sleeves, white gloves, pleated skirt, blue skirt, boots, white footwear

    • Tracen summer uniform

      • tracen school uniform, summer uniform, serafuku, puffy short sleeves, purple bowtie, horseshoe ornament, sailor collar, sailor shirt, purple shirt, white skirt, pleated skirt, frilled skirt, miniskirt, zettai ryouiki, white thighhighs, brown footwear, loafers

    • Tracen winter uniform

      • tracen school uniform, ear ornament, hairclip, winter uniform, puffy long sleeves, sailor collar, purple shirt, white bowtie, horseshoe ornament, purple skirt, pleated skirt, purple thighhighs, brown footwear, loafers

Use a model derived from or mixed with animefull model.

  • Settings

    • Stable Diffusion WebUI by Automatic1111

    • DPM++ Series(SDE Karras, 2M Karras, etc.)

    • about 20 steps, CFG scale 3.5~6.5

    • I recommend kl-f8-anime2.vae

    • CLIP skip = 1

    • Use Hires.fix to get higher quality image

      • Upscaler

        • Latent, Latent (nearest-exact), Latent (bicubic antialiased) or other Latent series.

      • Denoising strength 0.50~0.65

LoRA training info

  • trained on sd-scripts by kohya_ss.

  • based on Animefull-pruned model

  • 129 images consist of

    • 4x low quality 47 images

    • 8x medium quality 46 images

    • 16x high quality 36 images

    • => 1 epoch = 1132 images

  • replaced character feature tags with happy meek \(umamusume\)

    • 1girl, red eyes, pink hair, simple background, solo, tail, braid, white hair, official alternate costume, long hair, alternate hairstyle, brown eyes, horse ears, purple eyes, horse girl, brown hair, blonde hair, alternate hair length, bangs, horse tail, blunt bangs, animal ears, pink eyes, yellow eyes, short hair, grey hair, light brown hair, ponytail, medium hair, seiun sky \(umamusume\), bob cut

    • I think these tags are the essentials of Happy Meek or

    • unnecessary for training because of duplication.

  • resolution : 768x768 with aspect ratio bucketing.

  • other training settings is included in training data or metadata of LoRA file.

  • Yukino Bijin, Neo Universe, Zenno Rob Roy, Light Hello, Tracen Staff, Ines Fujin, Inari One and Tanino Gimlet were trained with learning rate of text encoder(TE) which has five times higher than Unet learning rate.

    • This is very interesting. Most of people uses TE lr as smaller than Unet lr. I did mistakes but got a good results. I don't know why these LoRA works. It may a clue to help to set proper learning rate.

    • I found this phenomenon in training of happy meek. During training, loss goes to nan. Happy meek has two tokens, so I think that higher TE learning rate with small tokens of trigger word make model boom.

All uploaded images are generated by AOM3A1

And, special thanks to my best friend. DH.

描述:

LoHa of Happy Meek (umamusume). Please install extension on WebUI.

After installing extension, put LoHa file in your sd/models/lora folder and write LoRA notation in prompt to applying LoHa. <lora:your_loha_name:weight>

训练词语: happy meek \(umamusume\)

名称: happy_meek_loha-000010.safetensors

大小 (KB): 19257

类型: Model

Pickle 扫描结果: Success

Pickle 扫描信息: No Pickle imports

病毒扫描结果: Success

Happy Meek (umamusume)

Happy Meek (umamusume)

Happy Meek (umamusume)

Happy Meek (umamusume)

Happy Meek (umamusume)

Happy Meek (umamusume)

资源下载
下载价格VIP专享
仅限VIP下载升级VIP
犹豫不决让我们错失一次又一次机会!!!
原文链接:https://1111down.com/915812.html,转载请注明出处
由于网站升级,部分用户密码全部设置为111111,登入后自己修改, 并且VIP等级提升一级(包月提升至包季,包季提升到包年 包年提升至永久)
没有账号?注册  忘记密码?

社交账号快速登录