I don't know what my friends are using.Stable DiffusionWhen painting, have you ever suddenly run out of ideas? You want to draw something, but you just don’t know where to start. You wrote the beginning of 1girl, but you don’t know what to write next. Anyway, Tingyu often encounters this situation.
Or I just don't know.Prompt wordHow to write to be rich.
Used the old sixPluginsThere are random inspirations, but it seems that there are only a few prompts, and if you use them too much, you will find that they are repeated over and over again.
Today we will introduce another plug-in DTG, which has a richer tag library and can bring us more inspiration.
Okay, without further ado, let’s get started.
The full name of the plug-in is: z-a1111-sd-webui-dtg, which mainly uses DanTagGen to enhance the prompt word generation function.
So what is DanTagGen? To put it simply, it is the Danbooru tag generator, a large language model designed to generate Danbooru tags.
The goal is to provide users with a more convenient way to make prompt words for text-to-image models trained on the Danbooru dataset.
This extension can generate detailed tags/core tags based on the prompt words you enter, and can add some additional elements to your prompt words.
Well, that’s about it. Friends who are interested in this model can go and take a look at the face.
Website:
https://huggingface.co/KBlueLeaf/DanTagGen-beta
Then let’s talk about how to install and use this model. A great god has made a plug-in that allows us to access this model.
This plugin can currently only be installed from a URL.
Website:
https://github.com/KohakuBlueleaf/z-a1111-sd-webui-dtg
After the installation is complete and SD is restarted, you can see the directory of this plug-in on the interface.
The usage is also very simple. Although there seem to be many parameters below, we can actually use most of them by default.
Remember to click Enable.
The option we usually modify most often is the "Total tag length".
It is obvious at a glance. The four options below are: very short, short, long, and very long.
This means how many prompt word labels the model randomly generates. This should be easy to understand.
Let’s try the shortest and the longest one here.
Let’s start with some basic prompts. Wow, the model is well chosen, and the output of the basic prompts is also very beautiful.
1girl,souryuu asuka langley,neon genesis evangelion,very aesthetic,masterpiece, Negative prompt: wings,nsfw,low quality,worst quality,normal quality,aidxlv05_neg, Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 4103425799 , Size: 896x1152, Model hash: e3c47aedb0, Model: animagine-xl-3.1, FP8 weight: Enable for SDXL, Cache FP16 weight for LoRA: True, VAE hash: c6a580b13a, VAE: sdxl_vae.safetensors, ENSD: 31337, TI hashes : "aidxlv05_neg: 84a6266ecaa6, aidxlv05_neg: 84a6266ecaa6", Version: v1.8.0
Next, we enable DanTagGen and select "very short" to see the effect. Compared with our original prompt words, many more adjectives and detailed description words have been added.
Different images of the same character immediately emerged.
1girl, souryuu asuka langley, neon genesis evangelion, small breasts, black gloves, turtleneck, looking to the side, blush, long hair, hand on own hip, nerv, :d, looking at viewer, glowing, pilot suit, smile, narrowed eyes, upper body, red gloves, breasts apart, hair between eyes, blue background, foreshortening, gloves, bodysuit, open mouth, simple background, solo, breasts, blue eyes, floating hair, two side up, orange hair, from side, impossible clothes, headgear, gradient background, glowing eye, plugsuit, very aesthetic, masterpiece Negative prompt: wings,nsfw,low quality,worst quality,normal quality,aidxlv05_neg, Steps: 28, Sampler: Euler a, CFG scale: 7, Seed : 4103425799, Size: 896x1152, Model hash: e3c47aedb0, Model: animagine-xl-3.1, FP8 weight: Enable for SDXL, Cache FP16 weight for LoRA: True, VAE hash: c6a580b13a, VAE: sdxl_vae.safetensors, ENSD: 31337, Version: v1.8.0
Here I need to remind you that when you enable generation for the first time, the DanTagGen model will be automatically downloaded. You can check the background. The model will also be placed in the network disk at the end of the article. Those who need it can pick it up by themselves.
Let’s try “very long” again. Tingyu is one of the more common tags, but I didn’t expect this picture... But isn’t this the effect we want, right!
1girl, souryuu asuka langley, neon genesis evangelion, small breasts, nervous, twintails, looking to the side, standing, blush, long hair, hair over shoulder, pilot suit, interface headset, cowboy shot, aged up, wide hips, thick mustache , covered navel, souryuu asuka langley \(cosplay\), green eyes, smile, pinky out, red gloves, skin tight, crossed arms, cosplay, pregnant, imagining, nervous smile, grin, tsurime, curvy, very long hair, hair between eyes, red bodysuit, gloves, hand on own stomach, bodysuit, full stomach, solo, sidelocks, blue bodysuit, breasts, simple background, glasses, sideways mouth, mustache, low twintails, beard, facial hair, lens flare, parted bangs , red-framed eyewear, thought bubble, multicolored bodysuit, alternate hair color, headgear, tape, plugsuit, science fiction, brown hair, multicolored clothes, very aesthetic, masterpiece Negative prompt: wings,nsfw,low quality,worst quality,normal quality,aidxlv05_neg, Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 4103425799, Size: 896x1152, Model hash: e3c47aedb0, Model: animagine-xl-3.1, FP8 weight: Enable for SDXL, Cache FP16 weight for LoRA: True, VAE hash: c6a580b13a, VAE: sdxl_vae.safetensors, ENSD: 31337, Version: v1.8.0
Okay, let's continue reading:
Ban tags: These are tags that need to be banned. They support regular expressions. This model still needs this function... Some sensitive tags...
Prompt Format: The format of the generated prompt words, which is also the default. It mainly refers to the prompt word labels classified into different categories in a prompt word, such as quality label, artist label, etc.
This is the default. If you choose randomization, a fixed seed will be generated like the random seed number. After the fixed seed number is generated, the prompt word will be fixed and no new prompt words will be added.
Upsampling timing: The time point when the prompt word is enhanced, before or after other prompt word plug-ins are processed. This function can be used with other dynamic prompt words or wildcard function plug-ins.
We can choose the right time.
This parameter represents temperature: higher temperatures represent more diverse results, lower temperatures represent greater consistency with our original prompt word.
Let's compare the effects by adjusting "Temperature" to the lowest and highest settings and try them respectively:
Adjust it to the lowest setting, and the labels are pretty standard. I tried it several times, and the pictures I got are basically the same as the one below, and the labels are also similar every time.
1girl, souryuu asuka langley, neon genesis evangelion, pilot suit, interface headset, red bodysuit, bodysuit, solo, medium breasts, breasts, plugsuit, blue eyes, long hair, orange hair, very aesthetic, masterpiece Negative prompt: wings,nsfw, low quality,worst quality,normal quality,aidxlv05_neg, Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 4103425799, Size: 896x1152, Model hash: e3c47aedb0, Model: animagine-xl-3.1, FP8 weight: Enable for SDXL, Cache FP16 weight for LoRA: True, VAE hash: c6a580b13a, VAE: sdxl_vae.safetensors, ENSD: 31337, Version: v1.8.0
When the setting is set to the highest level, the style that comes out is different every time. The styles are varied and there are all kinds of images. I have run a lot and basically haven't seen any duplicate ones.
1girl, souryuu asuka langley, neon genesis evangelion, twitter username, holding polearm, musblstone, lance renguahau \(craft viwony itee croata\) \(cast bank broad symbol\) art numberfayw movie interior, holding cyrgesight, blue buruma, hands on weapon , biceps, medium breasts, smile, two-tone border, muscular female, plugsuit, logo, lance, blue egun, interface back dx, very aesthetic, masterpiece Negative prompt: wings,nsfw,low quality,worst quality,normal quality,aidxlv05_neg , Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 4103425799, Size: 896x1152, Model hash: e3c47aedb0, Model: animagine-xl-3.1, FP8 weight: Enable for SDXL, Cache FP16 weight for LoRA: True, VAE hash: c6a580b13a, VAE: sdxl_vae.safetensors, ENSD: 31337, Version: v1.8.0
Here is another little trick I would like to share. When you enable the plugin and set multiple batches in the total batch, the styles of the pictures that come out at the same time are all in the same series! Isn’t this feature pretty nice?
Okay, that’s all the functions of the plug-in have been introduced. I have to say that the card drawing function of this plug-in is really powerful and full of creativity.
But this card drawing, well, everyone should be careful when using it, it is also easy to be eye-catching...
Okay, that’s all for today’s sharing. If you are interested, go and try it!