NoobAI XL V-Pred 0.5
模型介绍
该图像生成模型基于 Laxhar/noobai-XL_v1.0,利用带有原生标签和自然语言字幕的完整 Danbooru 和 e621 数据集。
作为 v-prediction 模型(不同于 eps-prediction)实现,它需要特定的参数配置 – 详见以下章节。
特别感谢我的队友 euge 的编码工作,我们也感谢许多热心的社区成员提供的技术支持。
⚠️重要通知⚠️
该模型的工作原理与 EPS 模型不同!
请仔细阅读本指南!
模型详细信息
- 开发者:Laxhar Lab
- 模型类型:基于扩散的文本到图像生成模型
- 微调自:Laxhar/noobai-XL_v1.0
- 赞助方:Lanyun Cloud
如何使用该模型。
方法一:reForge
- (如果您尚未安装 reForge)按照存储库中的说明安装 reForge;
- 切换到
dev_upstream
分支:
git checkout dev_upstream
3.更新重铸:
git pull
4.启动WebUI并照常使用该模型!
方法二:ComfyUI
带节点的样本
方法三:WebUI
请注意,dev 分支不稳定并且可能包含错误。
- (如果您尚未安装 WebUI)按照存储库中的说明安装 WebUI。对于简单
- 切换到
dev
分支:
git switch dev
3.获取最新更新:
git pull
4.启动WebUI并照常使用该模型!
方法 IV:Diffusers
import torch
from diffusers import StableDiffusionXLPipeline
from diffusers import EulerDiscreteScheduler
ckpt_path = "/path/to/model.safetensors"
pipe = StableDiffusionXLPipeline.from_single_file(
ckpt_path,
use_safetensors=True,
torch_dtype=torch.float16,
)
scheduler_args = {"prediction_type": "v_prediction", "rescale_betas_zero_snr": True}
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, **scheduler_args)
pipe.enable_xformers_memory_efficient_attention()
pipe = pipe.to("cuda")
prompt = """masterpiece, best quality,artist:john_kafka,artist:nixeu,artist:quasarcake, chromatic aberration, film grain, horror \(theme\), limited palette, x-shaped pupils, high contrast, color contrast, cold colors, arlecchino \(genshin impact\), black theme, gritty, graphite \(medium\)"""
negative_prompt = "nsfw, worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, mammal, anthro, furry, ambiguous form, feral, semi-anthro"
image = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
width=832,
height=1216,
num_inference_steps=28,
guidance_scale=5,
generator=torch.Generator().manual_seed(42),
).images[0]
image.save("output.png")
注意:请确保您的机器上安装了 Git 并且环境配置正确。
推荐设置
参数
- CFG:4 ~ 5
- 步骤:28〜35
- 采样方法:欧拉(其他采样器将无法正常工作)
- 分辨率:总面积约为 1024×1024。最佳选择:768×1344、832×1216、896×1152、1024×1024、1152×896、1216×832、1344×768、1024×1536、1536×1024
提示
- 提示前缀:
masterpiece, best quality, newest, absurdres, highres, safe,
- 负面提示:
nsfw, worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, mammal, anthro, furry, ambiguous form, feral, semi-anthro
使用指南
Caption
<1girl/1boy/1other/...>, <character>, <series>, <artists>, <special tags>, <general tags>, <other tags>
品质标签
对于质量标签,我们通过以下流程评估图像流行度:
- 根据各种来源和评级的数据规范化。
- 根据日期新近度应用基于时间的衰减系数。
- 根据此处理对整个数据集内的图像进行排名。
我们的最终目标是确保优质标签能够有效跟踪近年来的用户偏好。
百分位范围质量标签> 95thmasterpiece> 85th, <= 95thbest quality> 60th, <= 85thgood quality> 30th, <= 60thnormal quality<= 30thworst quality
美学标签
标签描述
| very awa | Top 5% of images in terms of aesthetic score by waifu-scorer | | worst aesthetic | All the bottom 5% of images in terms of aesthetic score by waifu-scorer and aesthetic-shadow-v2 | | … | … |
日期标签
There are two types of date tags: year tags and period tags. For year tags, use year xxxx
format, i.e., year 2021
. For period tags, please refer to the following table:
年份范围时间段标签 2005-2010old2011-2014early2014-2017mid2018-2020recent2021-2024newest
模特执照
此模型的许可证继承自https://huggingface.co/OnomaAIResearch/Illustrious-xl-early-release-v0 fair-ai-public-license-1.0-sd 并添加了以下条款。此模型及其变体的任何使用均受此许可证约束。
一、使用限制
- 禁止用于有害、恶意或非法活动,包括但不限于骚扰、威胁和传播错误信息。
- 禁止生成不道德或冒犯性的内容。
- 禁止违反用户所在管辖区域内的法律法规的行为。
二、商业禁令
我们禁止任何形式的商业化,包括但不限于将模型、衍生模型或模型生成的产品货币化或进行商业使用。
三、开源社区
为了促进开源社区的繁荣,用户必须遵守以下要求:
- 开源衍生模型、合并模型、LoRA以及基于上述模型的产品。
- 分享合成公式、提示和工作流程等工作细节。
- 遵循 fair-ai-public-license 确保衍生作品保持开源。
四、免责声明
生成的模型可能会产生意外或有害的输出。用户必须承担使用过程中的所有风险和潜在后果。
作品参照
1girl, solo, floating hair, :o, fang, blush, red eyes, green eyes, black headband, hair intakes, white hair, absurdly long hair, heterochromia, fox ears, sundress, white dress, slit pupils, crossed bangs, <lora:hayateluc:0.8>, paper lantern, sky lantern, night, looking up, cloud, upper body, looking afar, dark
Negative prompt: worst quality,bad quality,logo,watermark,username,text,watermark,bad anatomy,bad proportions,extra limbs,extra digit,extra legs,extra legs and arms,disfigured,missing arms,too many fingers,fused fingers,missing fingers,unclear eyes,watermark,username,censored
Steps: 20, CFG scale: 3, Sampler: DPM++ SDE CFG++, Seed: 440996518, Size: 1344x1728, Model: noobaiXLNAIXL_vPred05Version, Version: f1.0.2-v1.10.1RC-1.10.1, Mask blur: 4, Model hash: 78748f163e, Inpaint area: Only masked, Schedule type: Align Your Steps, Denoising strength: 0.2, Masked area padding: 32
1girl, hatsune_miku, lips, thin lips, parted lips, solo, looking at viewer, camisole, upper body, dark, underlightling, masterpiece, best quality, very aesthetic
Negative prompt: worst quality, low quality, displeasing, text, watermark, bad anatomy,
Steps: 30, CFG scale: 5, Sampler: Euler a, Seed: 919909395, Size: 832x1216, Created Date: 2024-11-13T1705:30.8473792Z, Clip skip: 2
masterpiece, best quality, newest, absurdres,
Dutch Angle, a Hanfu Bride, her elegant silhouette outlined against the dark night sky., BREAK Her delicate red hanfu, embroidered with intricate patterns, looking_at_viewer, <lora:ponyv4_noob1_2_adamW-000017:0.7000000000000001> <lora:USNR_STYLE_XL_lokr:0.6>
Negative prompt: lowres, (worst quality, low quality:1.2),
bad anatomy, old, High contrast,
Steps: 20, CFG scale: 5, Sampler: Euler, Seed: 49716620009, Size: 1024x1536, Model: IL\noobaiXLNAIXL_vPred05Version.safetensors, Model hash: 78748f163e, Clip skip: 2
使用介绍
关于此版本
尊敬的各位AIGC爱好者,
很高兴向大家介绍Laxhar Dream Lab推出的:NoobAI-XL EPS 1.0
该模型使用了1270万张图像(最新的Danbooru和e621完整数据集),在32*H100 GPUs上进行了32个epoch的训练(共计38.4亿步),已支持D站solo count 80图的角色和风格。
本版本训练过程中,来自nieta的算法实习生@li_li对trainer发挥了重要作用,在此进行特别鸣谢,感谢li_li作为Laxhar Lab成员的辛勤付出。
Lanyun作为本项目的算力赞助商,其对于开源社区的巨大贡献我们无以言表,Liblib AI在训练过程中提供了测试设备,也一同在此致谢。
同时,解构原典社群的伙伴们也在训练过程中进行了详细的测试与辅助工作,限于人数众多,无法一一鸣谢,在此对各位一并致以最诚挚的感谢!
Laxhar Dream Lab目前正全力致力于进一步完善SDXL开源生态,后续我们的工作是开发v预测版与noob配套的专用controlNet,以及更多配套插件,提高模型的泛用度,这也是这个模型明明的初衷,即“菜鸟也能用的很好的模型。”
介绍
这是一个基于Illustrious-xl训练的图像生成模型,并由Laxhar Lab继续训练。
它利用最新的完整 Danbooru 和 e621 数据集进行训练,并带有原生标签标题。
此版本模型提升了Illustrious-XL 0.1ver中人物与风格的贴合度,人物的具体特征有了更好的体现。Laxhar lab目前在此beta版本基础上继续训练XL新版开源模型,希望可以最大程度减少lora的使用,发布更加Noob友好的一键SDXL动漫模型!
当前状态
这是一个 100% EPS 版本,可供使用。我们下一步的工作包括为 NoobAI-XL 开发 v 预测版本和专用的 controlNet,以及其他补充插件,以增强模型的通用性。这也是这个模型的初衷——成为“连菜鸟都能用好的模型”。
禁止盈利:
● 禁止将任何闭源的微调/合并模型货币化,这种模型不允许公众访问模型的源代码/权重及其用法。
标题
<1girl/1boy/1other/...>, <character>, <series>, <artists>, <special tags>, <general tags>
品质标签
对于质量标签,我们通过以下流程评估图像流行度:
- 根据各种来源和评级的数据规范化。
- 根据日期新近度应用基于时间的衰减系数。
- 根据此处理对整个数据集内的图像进行排名。
我们的最终目标是确保优质标签能够有效跟踪近年来的用户偏好。
百分位数范围质量标签
> 95th 部杰作
> 85th, <= 95th 最佳品质
> 60th, <= 85th 质量好
> 30th, <= 60th 正常质量
<= 30th 质量最差
在 CCIP 测试中,noobaiXL 相比基础模型提升了约 2%。基于超过 3,500 个角色的数据,89.2% 的角色获得了高于 0.9 的 CCIP 分数。鉴于目前的模型性能,有必要进一步扩展现有 CCIP 测试的数据集。
● 根据许可证,您必须公开发布任何衍生模型和变体。此模型旨在供开源使用,所有衍生产品都必须遵循相同的原则。
作品参照
artist:hxxg,{pigeon666},[ciloranko,wlop][artist:ask_(askzy)],{artist:redi_(rasec_asdjh)},artist:wanke,hikari\(blue archive\),nozomi\(blue archive\),masterpiece,best quality,8K,year 2024,2girls,dynamic_angle,white_pantyhose,happy halloween,green hair,Golden eyes,demon tail,
Negative prompt: 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,aidxlv05_neg,FastNegative,
Steps: 28, CFG scale: 6, Sampler: Euler a, Seed: 551104985, Size: 896x1344, Model: noobaiXLNAIXL_epsilonPred075, Version: v1.10.1, Model hash: 0efaa1e7e1, Schedule type: Automatic, Denoising strength: 0.4, Clip skip: 2
luo tianyi,sitting on boat,alternate_costume,(artist:atdan:0.9),(artist:torino_aqua:0.8),(artist:izumi_tsubasu:0.9),(artist:kedama_milk:1.05),(artist:nardack:0.75),(artist:ke-ta:0.95),(artist:toosaka_asagi:0.7),(artist:sho_(sho_lwlw):0.8),(artist:as109:0.8),{{{{{{{{ink_wash_painting,sketch,Chinese style}}}}}}},year 2024,
1girl,{{{portrait}}},close-up,{{{transparent clothes,see-through chinese hanfu}}},{{white trim}},{{{{naked blue chinese hanfu}}}},{hanfu neck strap},breasts strap,erotic lingerie,{{{side-tie}}},{{long rope}},strapless,ribbon,chinese knot,{{{see-through shawl,see-through capelet ,tassel,floating clothes,gold trim}}},lotus,cleavage,wide sleeves,sitting on boat,boat,lotus leaf,lotus flower,collarbone,long sleeves,temple,mountain,buddha,sun,clouds,{{fog}},rain,water,splashing,hydrokinesis,wading,huge leaf,holding huge leaf,rain,leaf umbrella,lotus leaf,
extremely detailed cg unity 8k wallpaper,masterpiece,original,highres,ray tracing,dynamic lighting,refined rendering,depth of field,ambience sense,detailed light,extremely light,beautiful detailed lighting,perfect lighting,best shadow,detailed shadow,
Negative prompt: (kedama:2),(luo_xiaohei :3),{{(glowing eyes:2)}},{{{{(mismatched pupils:2)}}}},pov,boy,nsfw,{{chibi doll}},{{{{{pubic hair}}}}},wolf,{animal},cake,elf ears,weapon,card,screen,bird,chibi,doll,milk,lowres,{bad},error,fewer,extra,missing,worst quality,jpeg artifacts,bad quality,watermark,unfinished,displeasing,chromatic aberration,signature,extra digits,artistic error,username,scan,[abstract],very displeasing,displeasing,lowres,lowres,bad,text,error,missing,extra,fewer,cropped,jpeg artifacts,worst quality,bad quality,watermark,displeasing,unfinished,chromatic aberration,scan,scan artifacts,photo,deformed,realism,disfigured,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,{{{adult}}},{{{{bad anatomy}}}},{bad feet},bad hands,{{{bad proportions}}},{blurry},cloned face,cropped,{{{deformed}}},{{{disfigured}}},error,{{{extra arms}}},{extra digit},{{{extra legs}}},extra limbs,{{extra limbs}},{fewer digits},{{{fused fingers}}},gross proportions,ink eyes,ink hair,jpeg artifacts,{{{{long neck}}}},low quality,{malformed limbs},{{missing arms}},{missing fingers},{{missing legs}},{{{more than 2 nipples}}},mutated hands,{{{mutation}}},normal quality,out of frame,owres,{{poorly drawn face}},{{poorly drawn hands}},reen eyes,signature,text,{{too many fingers}},{{{ugly}}},username,uta,watermark,worst quality,{{{more than 2 legs}}},{{{furry girl}}},
Steps: 28, CFG scale: 5.8, Sampler: DPM++ 2M, Seed: 2189985388, Size: 1024x1523, Model: NoobAI-XL (NAI-XL)_Eps-pred 77%, Version: v1.10.1, Model hash: a5e17da21b, Hires steps: 14, Hires upscale: 1.3, Schedule type: Karras, Hires upscaler: R-ESRGAN 4x+ Anime6B, Denoising strength: 0.3, Clip skip: 2
1girl, kiritani haruka, project sekai,
reoen, liduke, eboda-x, yuyumu, (ningen mame:0.8638), miyase mahiro, chen bin, (sho \(sho lwlw\):0.8638), (ogipote:0.8227), (ciloranko:0.8638), (cogecha:0.9524), (wlop:0.9524), dishwasher1910, ru zhai,
year 2023, cowboy shot, autumn leaves, black pantyhose, blue eyes, blue hair, breasts, from above, ginkgo leaf, grey shirt, holding, holding leaf, leaf, long sleeves, looking at viewer, miyamasuzaka girls' academy school uniform, neckerchief, pantyhose, parted lips, pleated skirt, red neckerchief, sailor collar, school uniform, serafuku, shirt, short hair, skirt, solo, standing, white sailor collar
very aesthetic, high definition, amazing quality product \(project sekai\), blue skirt, medium breast pocket, arm at side, miniskirt, hand up, looking away,
masterpiece, best quality, very aesthetic, highres, absurdres, sensitive
Negative prompt: text,watermark,bad anatomy, bad proportions, extra limbs, extra digit, extra legs, extra legs and arms, disfigured, missing arms, too many fingers, fused fingers, missing fingers, unclear eyes,watermark,username
Steps: 30, CFG scale: 6.5, Sampler: DPM++ 2M, Seed: 1438116811, VAE: vae_trainer_step_250000.safetensors, Size: 1024x1536, Model: OUE9checkpoint-e2_s82000, Version: v1.10.1, Model hash: 9c4a5dfd28, TIPO format: tag only (DTG mode), TIPO prompt: [object Object], Schedule type: Karras, TIPO nl prompt: , TIPO Parameters: [object Object], Wildcard prompt: [object Object], Clip skip: 2
使用介绍
关于此版本
- 传统模型训练的新版本支持围绕图像数量 150(样式和字符)的概念,迄今为止已对 1270 万张图像进行了 22 个时期的训练。
- 这仍然是一个中间版本,可以使用传统的 A1111 Webui 运行,同时 A1111 承诺目前在主分支中支持所有版本的 NoobAI XL(包括 v-pred 版本)
- V-pred 版本和 epsilon-pred 版本仍在 Laxhar 实验室的待办事项计划中
- 本次发布已经是支持实际计数数字超过 100 的概念,甚至支持一些计数只有 50 的字符的中后期版本!如果有你喜欢的角色和样式符合条件但不受支持,可能是因为实际数据尚未收集,Laxhar Lab 欢迎你的反馈,以使模型变得更好!ミ(・・)ミ
我们希望大家喜欢这个 75% 的版本,并且我们计划下个月发布 100% 的 EPS-Ver,同时开始 V-pred 版本完整版本的开发工作。
这是一个基于Illustrious-xl训练的图像生成模型,并由Laxhar Lab继续训练。
它利用最新的完整 Danbooru 和 e621 数据集进行训练,并带有原生标签标题。
此版本模型提升了Illustrious-XL 0.1ver中人物与风格的贴合度,人物的具体特征有了更好的体现。Laxhar lab目前在此beta版本基础上继续训练XL新版开源模型,希望可以最大程度减少lora的使用,发布更加Noob友好的一键SDXL动漫模型!
禁止盈利:
● 禁止将任何闭源的微调/合并模型货币化,这种模型不允许公众访问模型的源代码/权重及其用法。
标题
<1girl/1boy/1other/...>, , , , <special tags>, <general tags>
品质标签
对于质量标签,我们通过以下流程评估图像流行度:
- 根据各种来源和评级的数据规范化。
- 根据日期新近度应用基于时间的衰减系数。
- 根据此处理对整个数据集内的图像进行排名。
我们的最终目标是确保优质标签能够有效跟踪近年来的用户偏好。
百分位数范围质量标签
> 95th 部杰作
> 85th, <= 95th 最佳品质
> 60th, <= 85th 质量好
> 30th, <= 60th 正常质量
<= 30th 质量最差
在 CCIP 测试中,noobaiXL 相比基础模型提升了约 2%。基于超过 3,500 个角色的数据,89.2% 的角色获得了高于 0.9 的 CCIP 分数。鉴于目前的模型性能,有必要进一步扩展现有 CCIP 测试的数据集。
● 根据许可证,您必须公开发布任何衍生模型和变体。此模型旨在供开源使用,所有衍生产品都必须遵循相同的原则。
作品参照
1girl, jinhsi \(wuthering waves\), jue \(wuthering waves\), wuthering waves,
reoen, liduke, eboda-x, yuyumu, (rei \(sanbonzakura\):1.05), (ciloranko:1.05), (akizero1510:1.05), (as109:0.9524), lam \(ramdayo\), ningen mame, puuzaki puuna, (maccha \(mochancc\):0.9524), (wlop:0.9524), (henreader:0.9524),
cowboy shot, solo, perspective, black collar, black dress, bow, breasts, closed eyes, coat, collar, dragon, dress, earrings, eastern dragon, eastern dragon horns, hair bow, horns, jewelry, long hair, medium breasts, multiple rings, ring, two side up, very long hair, white bow, white coat, white hair, high definition, clothing cutout, cleavage cutout, cleavage, black horns, tacet mark \(wuthering waves\), thighhighs, amazing quality \(wuthering waves\),
very aesthetic, masterpiece, best quality, very aesthetic, highres, absurdres, sensitive
Negative prompt: text,watermark,bad anatomy, bad proportions, extra limbs, extra digit, extra legs, extra legs and arms, disfigured, missing arms, too many fingers, fused fingers, missing fingers, unclear eyes,watermark,username
Steps: 35, CFG scale: 6.5, Sampler: DPM++ 2M, Seed: 2799576287, VAE: vae_trainer_step_250000.safetensors, Size: 1024x1536, Model: checkpoint-e4_s126656, Version: v1.10.1, Model hash: 0efaa1e7e1, TIPO format: tag only (DTG mode), TIPO prompt: [object Object], Schedule type: Karras, TIPO nl prompt: , TIPO Parameters: [object Object], Clip skip: 2
plana \(blue archive\),arona \(blue archive\),
2girls,
(idol:1.2),(split theme:1.2),polar opposites,split screen,symmetry,(symmetrical pose:1.1),heart_hands,w,
[artist:alpha],(artist:ciloranko),solo,[artist:sho (sho lwlw)],[tianliang duohe fangdongye],[artist:rhasta],year 2023,(wlop:0.4),
best quality,amazing quality,very aesthetic,absurdres,
Negative prompt: holding whip,whip,lowres,worst quality,bad quality,bad hands,very displeasing,extra digit,fewer digits,jpeg,pregnant,artifacts,signature,username,reference,mutated,lineup,manga,comic,disembodied,turnaround,2koma,4koma,monster,cropped,amputee,text,bad foreshortening,what,guro,logo,bad anatomy,bad perspective,bad proportions,artistic error,anatomical nonsense,amateur,out of frame,multiple views,
Steps: 20, CFG scale: 7, Sampler: Euler a, Seed: 3805232715, Size: 1024x1536, Model: noobaiXLNAIXL_epsilonPred075, Version: v1.10.1, Model hash: 0efaa1e7e1, Schedule type: Automatic
artist:yoneyama mai, (artist:wanke:1.1), artist:mocha, (cotton:1.1), artist:date, (senpen:1.1), artist:ciloranko, artist:noyu, noyu23386566, (artist:sho:0.8264462809917354), (sho lwlw:0.9090909090909091), year 2023, (solo:1.3310000000000004), lineart, monochrome, (red sword:1.2100000000000002), (fighting stance:1.2100000000000002), (disheveled hair:1.1), (dark heme:1.2100000000000002), chiaroscuro, foreshortening, (colorful splashes:1.1), (shining:1.1), (focus on face:1.1), Glowing skin, Strong contrast, long sleeves, simple background, red eyes, weapon, hood, black background, sword, hooded jacket, meme, jacket, hood up, red long hair, glowing, half-closed eyes, looking at viewer
Negative prompt: worst quality,low quality,logo,text,watermark,signature,
Steps: 50, CFG scale: 6.5, Sampler: Restart, Seed: 2162009144, Size: 1024x1536, Model: checkpoint-e3_s178000, Version: f2.0.1v1.10.1-previous-506-gd1e40361, Model hash: 5d4fff5a03, Hires steps: 10, TIPO format: tag only (DTG mode), TIPO prompt: [object Object], Hires upscale: 1.5, Schedule type: Karras, Hires upscaler: 4x-UltraSharp, TIPO nl prompt: , ADetailer model: face_yolov8m.pt, TIPO Parameters: [object Object], ADetailer version: 24.4.2, Denoising strength: 0.44, ADetailer mask blur: 4, ADetailer model 2nd: Eyes.pt, ADetailer confidence: 0.3, Style Selector Style: base, ADetailer dilate erode: 4, Style Selector Enabled: True, ADetailer mask blur 2nd: 4, ADetailer confidence 2nd: 0.3, Style Selector Randomize: False, ADetailer inpaint padding: 32, ADetailer dilate erode 2nd: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding 2nd: 32, ADetailer denoising strength 2nd: 0.4, ADetailer inpaint only masked 2nd: True