本帖最后由 领带骑士 于 2023-3-31 17:00 编辑
好像你坛都没人提到过colab,素都在背地里自己爽吗
今天好奇用手机试了试,感觉很不错
可以快速切换各种模型,运算速度也比我的1060快多了
加上你坛又开新活动了,所以来浅浅分享一下
【colab】在线ai绘画的入门教程1.1 进入stable-diffusion-webui-colab的github主页 当然你也可以选择其他指定了模型的版本(比如ye18),但是推荐在后面webui上选择模型 1.3 点击运行 耐心等待,一般十分钟以内会完成下载安装运行 (繁忙时可能会失败,换个时间吧) 1.4 部署成功后会给你两个链接,哪个都行,点击进入stable diffusion的webUI界面
(pc安卓端都可访问) 1.5 进入c站,浏览并选择自己喜欢的模型
比如我想要这个模型 1.6 模型下载与切换 下载: 可以看到速度非常快,一两分钟就下好了 切换: 1.7 第一次使用模型,最好使用官方样图给的prompt
l 不要急着吟唱!每个模型适合的prompt都不一样,比如ye18必备的uploadon e621标签这里就没有。最好在这个模型的标准prompt上一步步修改 l cfg等3个参数同理,最好和模型给的一致,按照官方的来效果一般会更好 l 最重要的是要看模型的介绍描述部分 This is intended to be a semi specialized solo gay male pinup / character creator. If you are looking for a model strong at generating females, this is not it lol. It is able to create multiple character images, but that's not what it's ideal for. There is currently another project I'm working on which will be better suited for that. This models ideal range is femboy to twunk. See testing prompt for a good place to start. Early version, still lots of work to do, but does not re-use any previous furry or yiffy models, which always messed up the pps (uwu). Trained various combinations of around 120,000 male gay focused furry images, for use with stable diffusion web UI. Tagging uses comma separated e-621 options without any underscores. Exception to tags: YMMV, but underscores are still needed when specifying a specific character tags and MetaTags. Meta tags are preserved as well, and continue to use underscores. IE: Vincent_(FoxMcC) on e621 would be represented as Vincent_\(FoxMcC\) in this model. IE: hi_res and digital_media_(artwork) on e621 would be hi_res, digital_media_\(artwork\) Artist tags will usually work okay with just the name, but you'll have the best luck prefixing them with 'by' IE: by chunie, by richard foley, etc. Be VERY SPECIFIC with what you want in an image with this, down to specifying eye color. See sample prompt. It is intended to be more of a character or pinup generator, but it's able to do well with multiple characters as well with careful sculpting. Some kinks are not represented in the model, images which were tagged with the following are not included in the image training dataset -animated -humanonly -hyper -watersports -scat -cub -comic -diaper -pixel(artwork) -3d_(artwork) -vore -chibi This isn't because I have anything against those kinks, rather I'm trying to create a slight built-in bias or remove unexpected elements from the images it generates. The criteria to be included in the training was a favcount greater than 200 and a score over 100, which also contain the tags 'anthro' and 'male' Note: Specifying the number of fingers you want helps immensely, and put your most important criteria at the front of the prompt, and nice to haves near the back.
Testing configuration /prompts Euler A - 28 (Heun 30 usually is an interesting vatiation) CFG 8 (works well between 6 and 9 depending on the prompt) Seed - 855213688 (you can set it to whatever but this is what I test with) Optional Hires testing configuration Upscaler - Latent (nearest) Denoising strength 0.6 Upscale 2 Positive Prompt: wolf, wolfdog, striped hyena, anthro, male, detailed, (detailed fur), (athletic), athletic anthro, muscular, muscular anthro, muscular male, fluffy, fluffy fur, fluffy ears, fluffy hair, fluffy chest, mane, pinup, canine ears, canine penis, canine anatomy, chest tuft, (balls), big balls, nude, nude male, solo, (fully sheathed), (penis tip), thigh up, (yellow eyes), ((sheath)), (white sheath), (black sheath), (brown sheath), tan sheath, gray sheath, retracted sheath, smug, grin, horny, lustful gaze, [orange fur], [red fur], [white fur], [blue fur], [yellow fur], two tone fur, two tone body, two tone hair, multi tone hair, sitting, posing, reclining, detailed background, five fingers, rating:explicit, digital_media_\(artwork\)
Negative Prompt: deformed, blurry, multi limb, taur, multiple limbs, bad anatomy, disfigured, poorly drawn face, mutated, extra limb, ugly, woman, female, multiple characters, feral, deformed hands, face out of frame, multiple tails, over saturated, hdr, high dynamic range, sketch, comic, human penis, no pupils, simple background, multiple characters, cub, child, jock, underwear, clothing, chastity, cage, abstract, chastity cage
Possible future testing versions can be found in Furry Diffusion discord or at https://wolf.gay/No commercial for profit use of any kind, or selling images made by this model, or any derivations of this model / images created by them. |
机翻: 这是为了成为一个半专业化的男同性恋单口相声/角色创作者。如果你正在寻找一个擅长生成女性的模特,这不是它,哈哈。它可以创建多个角色图像,但这不是它的理想选择。目前我正在进行的另一个项目将更适合这一点。这款车型的理想续航里程为女性至青少年。请参阅测试提示以获得一个好的起点。 早期版本,还有很多工作要做,但没有重复使用任何以前的毛茸茸的或毛茸茸的模型,这些模型总是把pps(uwu)搞砸。 训练了大约120000张以男性同性恋为焦点的毛茸茸的图片的各种组合,用于稳定的扩散web UI。 标记使用逗号分隔的e-621选项,没有任何下划线。 标签例外:YMMV,但在指定特定字符标签和MetaTags时仍需要下划线。元标记也得到了保留,并继续使用下划线。 (括号需要加反斜杠\进行转义,下划线不需要) IE:在这个模型中,e621上的Vincent_(FoxMcC)将被表示为Vincent_\(FoxMcC\)。 IE:e621上的hi_res和digital_media_(艺术作品)将是hi_res,digital_mmedia_\(artwork\) 艺术家标签通常只适用于名字,但你最好在它们前面加上“by” IE:by chunie, by richard foley, etc. 非常具体地说明你想要的图像,直到指定眼睛的颜色。请参见示例提示。它更像是一个角色或拼音生成器,但它可以很好地处理多个角色,也可以进行仔细的雕刻。 一些扭结没有在模型中表示,标记有以下内容的图像不包括在图像训练数据集中-动画-仅限人类-圣水-粪便-幼崽-漫画-尿布-像素(艺术品)-3d_(艺术品)-vore-chibi 这并不是因为我反对这些扭结,而是我试图创建一个轻微的内置偏见,或者从它生成的图像中删除意想不到的元素。 培训的标准是支持度超过200,得分超过100,其中还包含“antro”和“male”标签 注意:指定你想要的手指数量非常有帮助,并将你最重要的标准放在提示的前面,并将其放在后面。
测试配置/提示 Euler A -28 (Heun 30 usually is an interesting vatiation) CFG 8(works well between 6 and 9 depending on the prompt) Seed -855213688 (you can set it to whatever but this is what I test with) OptionalHires testing configuration Upscaler- Latent (nearest) Denoisingstrength 0.6 Upscale 2 PositivePrompt: wolf,wolfdog, striped hyena, anthro, male, detailed, (detailed fur), (athletic),athletic anthro, muscular, muscular anthro, muscular male, fluffy, fluffy fur,fluffy ears, fluffy hair, fluffy chest, mane, pinup, canine ears, canine penis,canine anatomy, chest tuft, (balls), big balls, nude, nude male, solo, (fullysheathed), (penis tip), thigh up, (yellow eyes), ((sheath)), (white sheath),(black sheath), (brown sheath), tan sheath, gray sheath, retracted sheath,smug, grin, horny, lustful gaze, [orange fur], [red fur], [white fur], [bluefur], [yellow fur], two tone fur, two tone body, two tone hair, multi tone hair,sitting, posing, reclining, detailed background, five fingers, rating:explicit,digital_media_\(artwork\) NegativePrompt: deformed, blurry, multi limb, taur,multiple limbs, bad anatomy, disfigured, poorly drawn face, mutated, extralimb, ugly, woman, female, multiple characters, feral, deformed hands, face outof frame, multiple tails, over saturated, hdr, high dynamic range, sketch,comic, human penis, no pupils, simple background, multiple characters, cub,child, jock, underwear, clothing, chastity, cage, abstract, chastity cage
没有任何形式的商业营利用途,也没有销售该模型制作的图像,或该模型/他们创建的图像的任何衍生产品。 1.8 输入测试prompt生成样图
至此算成功入门了 1.9 保存生成的图片 可以直接在webUI下载保存
也可以在colab文件目录查看生成过的所有图片
参考来源:
都看到这了,顺便给个追随吧
|