少年歌行-李寒衣
详情
下载文件
模型描述
版权说明:我的Lora训练均使用网络公开图片进行训练,训练后的Lora成品不推荐用于商业或者违法用途。如果你用于商业或者违法用途,一切风险由你自己承担,我不承担任何责任!
最近几个月比较忙,加上研究了一段时间的SDXL新模型,发现AI领域又流行起了倒卖、技术封闭的现象,国内经常出现这种问题,很难想象AI绘画刚开始的时候,你问tags或者使用模型推荐的时候,无话不说,现在问个底模,好像要命一样,有利益就有壁垒,但是AI绘画的使用有什么高深的技巧吗?有4090好显卡和没有好显卡的区别,仅此而已,目标是训练全的国漫角色模型,所有角色都会免费发布到C站或者对应网站。
QQ:3495154498 QQ群:744872631
Discord: https://discord.gg/MMcZec7D
C站: https://civitai.com/user/wzgrx/models
爱发电: https://afdian.net/a/wzgrx
吐司:wzgrx的个人资料 | 吐司 [Blocked Link] (tusiart.com)
lib:个人中心|LiblibAI
SD1.5主流模型Lora效果测试: 点击
模型训练流程,目前99%都差不多,没有一点点技术,都是显卡强度。
主流大模型,十个十一个都是这些大模型,其他私模都是互相融合,效果差不多,没有一点点区别。
chilloutmix-Ni.safetensors
perfectWorld_v5Baked.safetensors
3Guofeng3_v34.safetensors
素材质量大于一切,不要看什么优化器,训练参数,正则,花式玄学什么的,10%的作用都没有,大模型+素材质量99%的占比。
- 首先是素材准备,有8k分辨率优先,没有就4k,或者2k分辨率,发布的所有动画和微博剧照,超话等,逐帧截图,大概需要1-7天,根据动画集数来定,几百集时间翻倍。素材包括侧脸,正脸,全身,半身,各种角度,包括所有类型。
- 去除字幕水印和logo水印,还有抠图,1-3天
- 炼制Lora模型,分辨率1280,这个分辨率目前SD1.5和SDXL完全够用或者超出。
- 测试泛化性,保证可以换衣服、发色等
- 可以说质量保证,模型质量只敢说第一。
- 所有Lora模型支持2k放大,画质不清晰的自己放大,平时懒得跑2k图,很费时间,不代表不支持,包括C站作品,都是默认960分辨率跑的,糊是正常的,3090,24GB,1280分辨率已经是极限了,如果你有更好的显卡,可以尝试4k等分辨率。
- 跑图设置,可以搭配韩国娃娃模型,比重0.5-0.9,可以跑出皮肤很白的效果,如果你喜欢这种可以加大比例。
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3020062449, Size: 768x960, Model hash: 7234b76e42, Model: chilloutmix-Ni, Denoising strength: 0.55, Clip skip: 2, Hires upscale: 2, Hires steps: 10, Hires upscaler: SwinIR 4x, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: 完美世界-天狐仙子(41eaa017613e), AddNet Weight A 1: 0.75, AddNet Weight B 1: 0.75, AddNet Module 2: LoRA, AddNet Model 2: koreandolllikenessV20_v20(8f3d16e6eada),
AddNet Weight A 2: 0.75, AddNet Weight B 2: 0.75
模型是支持跑全身图的,自己搭配tags或者Control试,除了手和脚趾这种目前还没有什么好办法处理,其他都是最优解。
tags说明,参数
prompt
1girl,brown hair,long hair,looking at viewer,realistic,sleeveless,smile,solo,upper body,
个人tags只测试基础tags,人物,发色,镜头角度,服饰,
想加质量修饰词可以自己添加,这个很多,自己找就行。
(masterpiece), (best quality), (highly detailed), CG, unity, 8k wallpaper, (extremely detailed CG unity 8k wallpaper), HDR, (RAW photo), official wallpaper, official art, photo of the most beautiful artwork in the world, (realistic, photo realistic:1.3), an extremely delicate and beautiful,
negative prompt: fat, naked, nude, paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, ((monochrome)), ((grayscale)), bad anatomy, DeepNegative, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, bad feet, cropped, poorly drawn hands, poorly drawn face, mutation, deformed, extra fingers, extra limbs, extra arms, extra legs, malformed limbs, fused fingers, too many fingers, long neck, cross-eyed, mutated hands, polar lowres, bad body, bad proportions, gross proportions, text, error, missing fingers, missing arms, missing legs,
Steps: 15-35
Sampler: Restart,等
CFG scale: 7-10
分辨率:512-4k
Hires upscaler: 4x-UltraSharp,
附加Lora:韩国娃娃0.2-0.7,Pure Eros Face 0.2-0.7,其他Lora自行测试。
其他全身照角度tags可以自己加。
其他增强图片质量的方法可以看对应大模型的使用说明。跑图算法负面词什么的自己用自己常用的就行,没啥好说的。
- 模型质量不用怀疑,出图什么的在于你的显卡能跑多少分辨率和结合其他插件Control,或者细节Lora的搭配使用。
Copyright note: My Lora training is conducted using publicly available online images, and the trained Lora products are not recommended for commercial or illegal use. If you use it for commercial or illegal purposes, all risks are borne by you, and I do not assume any responsibility!
In recent months, I have been quite busy, and after studying the new SDXL model for a period of time, I have found that the AI field has become popular again, with reselling and technology being closed. This problem often occurs in China. It is difficult to imagine when you first asked tags or recommended models for AI painting, without saying anything. Now, when you ask the base model, it seems like it's deadly. There are benefits and barriers, but are there any advanced techniques in using AI painting? The difference between having a 4090 good graphics card and not having a good graphics card is just that. The goal is to train all character models in Chinese animation, and all characters will be published for free on the C website or corresponding website.




