Ai picture insensitivity that comes when you draw a lot of ai pictures... 1 person has seen too many pictures 1 person's pictures look the same... 1 person's brain that has been pickled in ai pictures The state of dopamine inflation So now let's use multiple loras and make at least 2 people Regional After installing the Prompter extension, check the activation of the Regional Prompter window, and we will try three horizontal and vertical masks that we will try . Divide Ratio = Screen division ratio This feeling is for 2 character rollers , the top row is 1,1 horizontally = both sides of the picture are common . If you interact here, hand-to-hand (holding hands) walk Use common prompt, so the common prom at the top is automatically added to the proms following BREAK . It comes out well even if it is vertical Let's put an expression on the expressionless character A smile behind the red character prom, an open mouse is completed The prom used behind Laura is applied separately. When using Laura, sometimes the picture becomes strange Try using another blue character Laura? Looking at the screen created in this way, when two rollers meet, the screen is broken like this . How to avoid the extension LoRA Block Weight Open the LoRA Block Weight section Complicated options at a glance... Let's use it as simply as possible Here you can open one more webui tip tab and use it Open the LoRA Block Weight section... It looks complicated, but don't be intimidated After the LoRA Block Weight XYZ setting LoRa prompt: XYZ LoRA Block Weight XYZ creation completed The blue character LoRa applied values were created to be easy to compare NONE = LoRa not applied ALL = (Overfitting symptom came out) Choose a useful one from the following Write MIDD If you use this, you can avoid the overfitting symptom. Return to the window that appeared strangely in the figure, and LoRA Block Weight is turned on. Paste :MIDD behind the overfitting lora. It comes out clearly If you add a related prom behind the character Laura, it becomes a little more natural Now it's a bit nice to see . Let's add one more green character Add the features and make 3 girls from the common upper row = left lower row = right character position in the picture Use 3 Laura Figure Because of the green character Laura, the picture shows overfitting symptoms again. When using Joe :MIDD, it was confirmed that only characters appeared without any background effect. There are exceptions depending on the type of Laura, but in most cases, when using :MIDD, overfitting could be avoided . Let's insert MIDD and run it. Looking at the finished picture, anyway, 3 LoRas are used without overfitting. 50% probability of 3 people? Those of you who have been drawing AI pictures for a while will know that AI pictures have great abilities in each wild, but are difficult to control. 1 out of 4 success rate? If you look at the picture on the far right, when the character invades the area, Laura's characteristics are mixed . Let's train them with ControlNet Arrange 3 people using the open pose editor and send them as text images to enter the ControlNet Activation model Open pose walking Hand-to-hand Walking holding hands Putting in and pulling out A picture came out... but it's 3 people... .The proportions of the human body are strange. 3 people who worked hard on the upper body... The open pose was not modified and the body shape came out as it is. What should I do? Adjust the control weight If the control weight is raised , it follows the skeleton, and if it is lowered, it moves away from the skeleton . It is to adjust the well. Then 4 rollers are possible? I made an open pose for 4 Lauras, and 4 of them entered. Regional Prompter 4 areas were created, and 1 more character Laura was added. To avoid overfitting :MIDD Prompt, 4girls. Will 4 Lauras come out? Completed 3 sheets are normal output As expected, unbridled foal ai picture... Chapter 1 has a character looking back If this is enough, 4 loras are output without problems. It is easy up to here. ? What if the character is located in a non-rectangular area? How do you do this? Let's use the Regional Prompter mask function. A vertical white canvas comes out. Figure out the position of the character in the open pose and draw the first character position to fill it. Press the Draw region to color this side . I will draw more characters in the middle I will fill in the lower area as well Fill in the area at the bottom 3 areas complete Let's put 3girls in the common prompt at the top and imagine...what are these 3 people doing? Let's put a jump prom in the blue sky. 3 characters jump into the blue sky . You might think of it as that, but what if the camera was raised up and the three characters were lying down instead of looking down? Lie down in the common prom Try entering Picnic on viewpoint It is completed like this. If you check it, it will appear as a scene with the Picnic on viewpoint prompt applied when lying down Until recently, there was only one LoRa... Now there is no problem using several LoRas... What is the development speed of ai ... The video part added later... It is more convenient to draw the area by inserting a picture Then, this time, Regional Prompter expansion can be used for screen division. Attention = 1 character roller Latent = 2 or more character rollers The method of dividing the aspect ratio is used. Divided Copy English as it is and paste it into the prompt window Go down one row with enter and enter the evening sunset on the top line (enter the number 0) Riverside = Riverside (enter the number 1) 1girl, bicycle, outfit (enter the number 2) Enter) Let's make 4 sheets. Top row Enter evening sunset (enter number 0) Riverside = Riverside (enter number 1) 1girl, bicycle, outfit (enter number 2) Apply OK Delete riverside and enter baseball field if? Top row Enter evening sunset (enter digit 0) Baseball field (enter digit 1) 1girl, bike, outfit (enter digit 2) Apply OK Send baseball field down and bike girl up . What if you delete it, change it to fireworks, and create it? Fireworks (enter the number 0) 1girl, bicycle, outfit (enter the number 1) baseball field (enter the number 2) What if this method is used for a full body shot of one character? Create three vertical divisions and designate each character separately for the head, upper body, and lower body. Separate them with BREAK and enter the form. This will result in a vertical picture in a 1:1:1 ratio . Can I specify the color and put it in? Colors are specified and output at a ratio of 1:1:1. What if the current ratio is modified to 1:1:1? What if the vertical ratio is 1:1:2? At a ratio of 1:1:2, the bottom takes up a ratio of 2. The shooting time is getting longer... I ate tteokbokki in Rosé . If you've been dealing with AI drawings for more than a month, there must be a lot more people with more model types than me. If you just collect them, it'll be like a list of steam games you don't play... You need to delete the drawings except for the ones you want. Then how do you conveniently organize them? Let's use the XYZ Plot convenient function. XYZ Plot X Horizontal Y Vertical Z on the script side Simply put the prompts separately and select X Checkpoint (model) And press this here to display the model you have. Press X here to subtract the unwanted model. It can be done like this once, and the vertical resolution output is output like this. If you check it , it is output for each model with the same seed and the same prompt. For the Y axis, use the Prompt S/R function. It was printed automatically replaced with the content. I 'll try it. It will take some time. It was printed like this . Try it? If you are curious about the picture change from cfg scale 1 to 20? XYZ plot cfg scale selection 1-20 (+1) If you enter this, 20 sheets are output in increments of 1 to 20 from 1 to 20. You can see the change from cfg scale 1 to 20 at a glance (you can check other numerical options like this) Let's make sure to organize the model... Before we suffer from SSD capacity shortage, let's find out about prompt weights . If you give and draw a monochrome sketch like this, it will be output as a black and white sketch. But if you use the weight function here, what if you want to give a more sketchy feeling in the picture? Take a good look at the prompt window keyboard here. If you move with the arrows on the keyboard, move one space. Move the arrow words with the keyboard while holding down Control. Select a word when moving the arrow with the keyboard while holding down Control Shift. If you press the up and down arrow keys while pressing Control + Shift, the weight is applied. The state without weight applied = 1 If the scale weight is 1.5? Result of scale weight 1.5 Result of scale weight 2 This weights What if we use the weights with the XYZ plot function to see the change? Insert X Prompt SR Sketch 0.1, and as we learned earlier, 0.1 words in the prompt are replaced and output continuously . It can also be applied to Laura. If you look closely at the prompt where Laura is entered, the default weight is 1. If you pull it out? Laura-applied picture that changes to the old drawing style This is also using the XYZ plot function, so 0.1 words in the prompt are replaced and output continuously. Laura weights can also be checked at a glance. Let's briefly look at the basic button functions . These also have a function. If you right-click the Generate Forever button on the Create button, it will be created infinitely until you cancel the right-click . What if you found a picture from the Sivitai site that you liked? Press copy all from Prom and paste it into the prompt box. In this state, if you press the arrow, everything below will be automatically entered like this. You can also copy all copies from the Prompt Search site. Is there a picture you like on the archive site? You know the program that displays EXIF information? (Link in the comment) (The person who uploaded the picture needs to upload the EXIF information so that the information appears) Copy the metadata and paste it into the Prom window Press the arrow to automatically enter the trash button Empty only the prompt field Save the prompt input want to do it? Can be saved by pressing the diskette shape. It is saved in the style box and can be recalled. If there is another word in the prompt, it is entered after it. The style box is entered, but nothing is deleted? how do you erase It is saved in the styles.csv file in the Stable Diffusion folder. When you open this, delete unwanted styles here. You can edit styles.csv to delete it or add it. If you are new to the Stable Diffusion AI program, you need a period of adaptation. Like me, English For depression, English input is burdensome (I'm used to it now, so it's okay...) There is a way to make English input easier. Chrome Extension Check DeepL Tray Icon Click DeepL Check Prom Window, then DeepL Extension icon appears . If you press +Y, the selected area changes to the translated English. Let's print out 4 pages. Sitting at the library desk, I am dozing off while studying. Translation is applied and a picture is created . The sentence is correct. Stable Diffusion WEBUI is the Novellai site hacked . +c (Press twice) The DeepL program window opens and translates into Korean. Even complicated prompts can be easily understood with the Control+c (Press twice) DeepL translation function. Let's overcome English anxiety! (The best thing is, let's study English...) And the emoji keyboard is also recognized in the prompt column. If you press the Windows key + .( period ), the emoji window comes out. Enter a smiley face emoji The picture prompt box also recognizes emojis If you put in a V mark emoji? Add smiley mark V mark firefighter emoji input Input smiley mark V mark firefighter emoji All are recognized and output ai tag library site prompt is shown as a picture, so it is easy to know the effect in advance and insert it Once you click the picture, the contents of the prom are copied from All you have to do is paste it in the box. It's comfortable to see the effect before putting it in. The outfit is also in a picture, so it's good to preview and insert it from the prom. I'm all annoyed... Do you want to automatically change the prompt for each picture and keep creating pictures? If you install the extension program Dynamic Prompt for those who are so bothered by it, there is a wildcard function. Wild Manager What is this? When entering the stable-diffusion-webui\extensions\sd-dynamic-prompts\wildcards path, create a new text document. Enter any name. Input 4 point prompts for testing . It looks like this . It goes in. Copy the viewpoint characters as they are If you enter the text in the Prom column, the content in the text (4 viewpoints) is randomly entered. The 4 viewpoints entered randomly . .. So if you enter 100 background places for ai picture prompts on the ChatGPT site? (When using the Prompt Genie Chrome extension, if you enter Korean, it will be automatically translated) ChatGPT is creating 100 backgrounds... ChatGPT has informed you of 100 backgrounds . And if you go back to webui and refresh the wild card , you have two points of view and location . You can put it in by telling ChatGPT to get 100 outfits. In this state, if you press the right-click generator Forever? It becomes an infinite meme creation machine... What you learn when you draw a picture... Due to the nature of AI, it can't draw its hand well.. Looking at it now, the character is dozing on the desk . But do you want to save this painting somehow? There is a way you can try Variation Seed Function Put a picture of a fool in PNG Info and import it as a text image, and the seed is fixed . We have nothing to do. Bring some popcorn. Robin... Variation seed 0.1 applied 20 sheets printed out There is a normal hand somewhere among them! If not, let's draw again... Still, you got two hand-retouched paintings worth seeing? There is a possibility of solving it somehow on the "click" line . If it doesn't come out, let's keep pulling it... It's an ai picture... I saw a picture on some site and want to see the EXIF information? But no information? When you put a picture in the img2img tab and press the Interrogate DeepBooru button, ai roughly analyzes the picture itself and gives you a rough prompt . Although it was made by analyzing ai with no EXIF information, it is still 30% similar? Let's use it for pictures that are very curious about EXIF And introduce two amazing sites. First, upload the character you made to the GIF creation site Make a moving effect like this . Yo , and the second strange site is a site that automatically picks up the noose (lucky). After clicking Upload, you can upload a picture file and specify the area to be cut out with the mouse . When I press cutout in , the one I just selected was cut out, so I right-clicked it and saved it as a cutout done. The video of looking at useful information on ai painting is finished. There is a lot of information about ai painting . Bun...oh...) Ai drawings with a lot of talk and a lot of escapades... I hope it helps those who need information. Thank you for watching the long video.
