search

hako-mikan / sd-webui-regional-prompter

set prompt to divided region
GNU Affero General Public License v3.0
1.47k stars 118 forks source link

readme

Regional Prompter

top

for LoHa, LoCon users

About LoRA/LoCon/LoHa
There are certain constraints due to the specifications of the Web-UI regarding the following: These constraints arise because the Web-UI cannot perform specific optimizations when applying LoRA and does not support mid-strength changes to LoRA.

LoRA/LoCon/LoHaについて LoRAの種類別の使用条件です。

Updates

Overview

Latent couple extention performs U-Net calculations on a per-prompt basis, but this extension performs per-prompt calculations inside U-Net. See here(Japanese) for details. Thanks to furusu for initiating the idea. Additional, Latent mode also supported.

index

Usage

This section explains how to use the following image, explaining how to create the following image.
sample
Here is the prompt.

green hair twintail BREAK
red blouse BREAK
blue skirt

setting

Active : On
Use base prompt : Off
Divide mode : Vertical
Divide Ratio : 1,1,1
Base Ratio :

This setting divides the image vertically into three parts and applies the prompts "green hair twintail" ,"red blouse" ,"blue skirt", from top to bottom in order.

Active

This extention is enabled only if "Active" is toggled.

Prompt

Prompts for different regions are separated by BREAK keywords. Negative prompts can also be set for each area by separating them with BREAK, but if BREAK is not entered, the same negative prompt will be set for all areas.

Using ADDROW or ADDCOL anywhere in the prompt will automatically activate 2D region mode.

Use base prompt

Check this if you want to use the base prompt, which is the same prompt for all areas. Use this option if you want the prompt to be consistent across all areas. When using base prompt, the first prompt separated by BREAK is treated as the base prompt. Therefore, when this option is enabled, one extra BREAK-separated prompt is required compared to Divide ratios.

Automatically turned on when ADDBASE is entered.

Divide ratio

If you enter 1,1,1, the image will be divided into three equal regions (33,3%, 33,3%, 33,3%); if you enter 3,1,1, the image will be divided into 60%, 20%, and 20%. Fractions can also be entered: 0.1,0.1,0.1 is equivalent to 1,1,1. For greatest accuracy, enter pixel values corresponding to height / width (vertical / horizontal mode respectively), eg 300,100,112 -> 512.

Using a ; separator will automatically activate 2D region mode.

Base ratio

Sets the ratio of the base prompt; if base ratio is set to 0.2, then resulting images will consist of 20%*BASE_PROMPT + 80%*REGION_PROMPT. It can also be specified for each region, in the same way as "Divide ratio" - 0.2, 0.3, 0.5, etc. If a single value is entered, the same value will be applied to all areas.

split mode

Specifies the direction of division. Horizontal and vertical directions can be specified. In order to specify both horizontal and vertical regions, see 2D region mode.

calcutation mode

Internally, system use BREAK in Attention mode and AND in Latent mode. AND/BREAK is automatically converted depending on the mode being used, but there is no problem whether you input BREAK or AND in the prompt.

Attention

Normally, use this one.

Latent

Slower, but allows separating LoRAs to some extent. The generation time is the number of areas x the generation time of one pic. See known issues.

Example of Latent mode for nendoorid, figma LoRA separated into left and right sides to create.

Use common prompt

If this option enabled, first part of the prompt is added to all region parts.

Automatically turned on when ADDCOMM is entered.

best quality, 20yo lady in garden BREAK
green hair twintail BREAK
red blouse BREAK
blue skirt

If common is enabled, this prompt is converted to the following:

best quality, 20yo lady in garden, green hair twintail BREAK
best quality, 20yo lady in garden, red blouse BREAK
best quality, 20yo lady in garden, blue skirt

So you must set 4 prompts for 3 regions. If Use base prompt is also enabled 5 prompts are needed. The order is as follows: common, base, prompt1,prompt2,...

2D region assignment

You can specify a region in two dimensions. Using a special separator (ADDCOL/ADDROW), the area can be divided horizontally and vertically. Starting at the upper left corner, the area is splited by columns when separated by ADDCOL and rows when separated by ADDROW. The ratio of division is specified as a ratio separated by a semicolon. An example is shown below; although it is possible to use BREAK alone to describe only the ratio, it is easier to understand if COL/ROW is explicitly specified. Using ADDBASEas the first separator will result in the base prompt. If no ratio is specified or if the ratio does not match the number of separators, all regions are automatically treated as equal multiples. In this mode, the direction selected in Divide mode changes which separator is applied first:

In any case, the conversion of prompt clauses to rows and columns is from top to bottom, left to right.

(blue sky:1.2) ADDCOL
green hair twintail ADDCOL
(aquarium:1.3) ADDROW
(messy desk:1.2) ADDCOL
orange dress and sofa
Active : On
Use base prompt : Off
Main splitting : Columns
Divide Ratio : 1,2,1,1;2,4,6
Base Ratio :

2d

Visualise and make template

Areas can be visualized and templates for prompts can be created.

tutorial

Enter the area ratio and press the button to make the area appear. Next, copy and paste the prompt template into the prompt input field.

fantasy ADDCOMM
sky ADDROW
castle ADDROW
street stalls ADDCOL
2girls eating and walking on street ADDCOL
street stalls

Result is following, tutorial

This is an example of an area using 1,1;2,3,2;3,2,3. In Columns, it would look like this:
flip
In Rows, it would appear as follows:
flip
When the flip option is enabled in Rows, it would appear as follows:
flip

Mask regions aka inpaint+ (experimental function)

It is now possible to specify regions using either multiple hand drawn masks or an uploaded image containing said masks (more on that later).

RegionalMaskGuide2 RegionalMaskGuide2B

Here is sample and code
using mask and promptlandscape BREAK moon BREAK girl. Using XYZ plot prompt S/R, changed moon BREAK girl to others.
RegionalMaskSample

region specification by prompt (experimental)

The region is specified by the prompt. The picture below was created with the following prompt, but the prompt apple printed should only affect the shirt, but the actual apples are shown and so on. 

lady smiling and sitting, twintails green hair, white skirt, apple printed shirt

prompt If you enhance the effect of apple printed to :1.4, you get,

prompt The prompt region specification allows you to calculate the region for the "shirt" and adapt the "printed apples".

prompt

lady smiling and sitting, twintails green hair, white skirt, shirt BREAK
(apple printed:1.4),shirt

prompt

How to use

syntax

baseprompt target1 target2 BREAK
effect1, target1 BREAK
effect2 ,target2

First, write the base prompt. In the base prompt, write the words (target1, target2) for which you want to create a mask. Next, separate them with BREAK. Next, write the prompt corresponding to target1. Then enter a comma and write target1. The order of the targets in the base prompt and the order of the BREAK-separated targets can be back to back.

target2 baseprompt target1  BREAK
effect1, target1 BREAK
effect2 ,target2

is also effective.

threshold

The threshold used to determine the mask created by the prompt. This can be set as many times as there are masks, as the range varies widely depending on the target prompt. If multiple areas are used, enter them separated by commas. For example, hair tends to be ambiguous and requires a small value, while face tends to be large and requires a small value. These should be ordered by BREAK.

a lady ,hair, face  BREAK
red, hair BREAK
tanned ,face

threshold : 0.4,0.6 If only one input is given for multiple regions, they are all assumed to be the same value.

Prompt and Prompt-EX

The difference is that in Prompt, duplicate areas are added, whereas in Prompt-EX, duplicate areas are overwritten sequentially. Since they are processed in order, setting a TARGET with a large area first makes it easier for the effect of small areas to remain unmuffled.

Accuracy

In the case of a 512 x 512 image, Attention mode reduces the size of the region to about 8 x 8 pixels deep in the U-Net, so that small areas get mixed up; Latent mode calculates 64*64, so that the region is exact. 

girl hair twintail frills,ribbons, dress, face BREAK
girl, ,face

Prompt-EX/Attention prompt Prompt-EX/Latent prompt

Mask

When an image is generated, the generated mask is displayed. It is generated at the same size as the image, but is actually used at a much smaller size.

Difference between base and common

a girl ADDCOMM (or ADDBASE)
red hair BREAK
green dress

If there is a prompt that says a girl in the common clause, region 1 is generated with the prompt a girl , red hair. In the base clause, if the base ratio is 0.2, it is generated with the prompt a girl 0.2 + red hair 0.8. Basically, common clause combines prompts, and base clause combines weights (like img2img denoising strength). You may want to try the base if the common prompt is too strong, or fine tune the (emphasis). The immediate strength that corresponds to the target should be stronger than normal. Even 1.6 doesn't break anything.

Known issues

3) Wait until a fix is published.

Here are samples of a simple prompt, two loras with negative te/unet values per lora of: (0,0) default, (1,0), (0,1), (1,1). MeguminMigurdiaCmp

If you come across any useful insights on the phenomenon, do share.

How to Use via API

The following format is used when utilizing this extension via the API.

  "prompt": "green hair twintail BREAK red blouse BREAK blue skirt",
    "alwayson_scripts": {
        "Regional Prompter": {
            "args": [True,False,"Matrix","Vertical","Mask","Prompt","1,1,1","",False,False,False,"Attention",False,"0","0","0",""]
}}

Please refer to the table below for each setting in args. No. corresponds to the order. When the type is text, please enclose it with "". Modes 3-6 ignore submodes that do not correspond to the mode selected in mode 3. For the mask in 17., please specify the address of the image data. Absolute paths or relative paths from the web-ui root can be used. Please create the mask using the color specified in the mask item.

No. setting choice type default
1 Active True, False Bool False
2 debug True, False Bool False
3 Mode Matrix, Mask, Prompt Text Matrix
4 Mode (Matrix) Horizontal, Vertical, Colums, Rows Text Columns
5 Mode (Mask) Mask Text Mask
6 Mode (Prompt) Prompt, Prompt-Ex Text Prompt
7 Ratios Text 1,1,1
8 Base Ratios Text 0
9 Use Base True, False Bool False
10 Use Common True, False Bool False
11 Use Neg-Common True, False Bool False
12 Calcmode Attention, Latent Text Attention
13 Not Change AND True, False Bool False
14 LoRA Textencoder Text 0
15 LoRA U-Net Text 0
16 Threshold Text 0
17 Mask Text
18 LoRA stop step Text 0
19 LoRA Hires stop step Text 0
20 flip True, False Bool False

Example Settings

Matrix

  "prompt": "green hair twintail BREAK red blouse BREAK blue skirt",
    "alwayson_scripts": {
        "Regional Prompter": {
            "args": [True,False,"Matrix","Vertical","Mask","Prompt","1,1,1","",False,False,False,"Attention",False,"0","0","0",""]
}}

Result sample

Mask

   "prompt": "masterpiece,best quality 8k photo of BREAK (red:1.2) forest BREAK yellow chair BREAK blue dress girl",
    "alwayson_scripts": {
        "Regional Prompter": {
            "args": [True,False,"Mask","Vertical","Mask","Prompt","1,1,1","",False,True,False,"Attention",False,"0","0","0","mask.png"]

Mask used sample
Result sample

Prompt

 "prompt": "masterpiece,best quality 8k photo of BREAK a girl hair blouse skirt with bag BREAK (red:1.8) ,hair BREAK (green:1.5),blouse BREAK,(blue:1.7), skirt BREAK (yellow:1.7), bag",
    "alwayson_scripts": {
        "Regional Prompter": {
            "args": [True,False,"Prompt","Vertical","Mask","Prompt-EX","1,1,1","",False,True,False,"Attention",False,"0","0","0.5,0.6,0.5",""]
}}

sample

Acknowledgments

I thank furusu for suggesting the Attention couple, opparco for suggesting the Latent couple, and Symbiomatrix for helping to create the 2D generation code.

Updates