0. Preface

• This article is an additional information for this LoRA

• I would like to show how to use the LoRA with ControlNet Canny

• The LoRA tries to fix the problem that the neck isn't fully covered when generating characters wearing bodysuit

• If using the LoRA as usual, the problem is gone but it changes the output results

• So I use it with Canny to minimize the effect and keep the character and composition I want

• For the basics of ControlNet Canny, see another article written by me

• I won't explain too much here

• I'm not an expert of AI

• This method can't use in CivitAI on-site generator at the moment (it doesn't support ControlNet)

• It may not work in some cases

• THIS IS A TIME AND LABOR COMSUMING METHOD

• Using Platform: reForge (local)

• Graphic Card: GeForce RTX 3060

• RAM: 32GB

1. Example 1

• Here is a simple example

• First, Generate a image

• 

  (Used Checkpoint and Lora 1 2)

• The neck isn't covered

• So I will fix it using the LoRA and Canny

1. 

   Enable ControlNet Unit 0

2. Drop the generated image

3. Select "Canny"

4. Enable "Pixel Perfect"

5. Run the preprocessor

6. Click "Edit" to open Photopea

• 

  Make sure that the name of the layer imported is "unit-0" (the ControlNet Unit in use)

• if not, rename it, or you can't send the layer back to the ControlNet

• Right click on the layer and choose "Rasterize"

• 

  Using black Pencil Tool to remove lines of the neck carefully (A)

• Remove the redundant line on the image border (B)

• 

  After

• 

  Click "Send to ControlNet" on the top left hand corner

• In ControlNet, Keep the Control Weight to 1 (keep those lines unchanged)

• In prompt, add "covered neck", the Lora and set weight to 0.8 (<lora:coveredneck_v1-000006:0.8>), and generate

• 

  Compare two images and you can see the problem is fixed and some details are slightly changed

• If you happy with that, you can continue to generate through changing parameters and prompts

• For example, I replace "white background" with "outdoors, night, dramatic lighting", and generate

• 

  you also can change the suit color through replacing corresponding words

• See here for details

• 

  you can add "white neck" or/and change seed if necessary

• I add it the blue one without changing seed

2. Example 2

• Here is another example

• I will share some skills of generating this image

• First, I generate the image like this

• 

  (Used Checkpoint and Lora 1 2 3)

• the neck isn't fully covered. Someone may like it but I don't

• Also, there are some redundant lines on hair

• Put the same image to ControlNet and run the canny preprocessor

• 

  In Photopea, removed those lines carefully using black pencil

• 

  For me, I will keep those lines. The rest will be recreated

• 

  Remove those redundant hair lines carefully

• remove other unwanted lines and send it back to ControlNet

• Keep the Control Weight to 1

• In prompt, add "covered neck" and the LoRA with weight 0.8 (<lora:coveredneck_v1-000006:0.8>)

• Since the lines are fixed and some words like "aged down" will not longer affect the image greatly. You can remove it optionally

• Click "generate"

• 

  Looks great

• BONUS: if using the LoRA without ControlNet, it becomes

• 

  The result is changed and IT'S NOT GOOD

• That's why I'm using the LoRA with Canny

3. Conclusion

• Again, THE PROCESS IS TIME AND LABOR COMSUMING

• Some skills and sense are required

• It's not the best solution, but it's viable

Notes

• I usually start with white background

• It can be changed later and this makes it easier for me to edit the Canny images without annoying lines