InvokeAI/docs/features/INPAINTING.md

title
Inpainting

:octicons-paintbrush-16: Inpainting

Creating Transparent Regions for Inpainting

Inpainting is really cool. To do it, you start with an initial image and use a photoeditor to make one or more regions transparent (i.e. they have a "hole" in them). You then provide the path to this image at the dream> command line using the -I switch. Stable Diffusion will only paint within the transparent region.

There's a catch. In the current implementation, you have to prepare the initial image correctly so that the underlying colors are preserved under the transparent area. Many imaging editing applications will by default erase the color information under the transparent pixels and replace them with white or black, which will lead to suboptimal inpainting. It often helps to apply incomplete transparency, such as any value between 1 and 99%

You also must take care to export the PNG file in such a way that the color information is preserved. There is often an option in the export dialog that lets you specify this.

If your photoeditor is erasing the underlying color information, dream.py will give you a big fat warning. If you can't find a way to coax your photoeditor to retain color values under transparent areas, then you can combine the -I and -M switches to provide both the original unedited image and the masked (partially transparent) image:

invoke> "man with cat on shoulder" -I./images/man.png -M./images/man-transparent.png

If you are using Photoshop to make your transparent masks, here is a protocol contributed by III_Communication36 (Discord name):

Create your alpha channel for mask in photoshop, then run image/adjust/threshold on that channel. Export as Save a copy using superpng (3rd party free download plugin) making sure alpha channel is selected. Then masking works as it should for the img2img process 100%. Can feed just one image this way without needing to feed the -M mask behind it

Masking using Text

You can also create a mask using a text prompt to select the part of the image you want to alter, using the clipseg algorithm. This works on any image, not just ones generated by InvokeAI.

The --text_mask (short form -tm) option takes two arguments. The first argument is a text description of the part of the image you wish to mask (paint over). If the text description contains a space, you must surround it with quotation marks. The optional second argument is the minimum threshold for the mask classifier's confidence score, described in more detail below.

To see how this works in practice, here's an image of a still life painting that I got off the web.

You can selectively mask out the orange and replace it with a baseball in this way:

invoke> a baseball -I /path/to/still_life.png -tm orange

The clipseg classifier produces a confidence score for each region it identifies. Generally regions that score above 0.5 are reliable, but if you are getting too much or too little masking you can adjust the threshold down (to get more mask), or up (to get less). In this example, by passing -tm a higher value, we are insisting on a tigher mask. However, if you make it too high, the orange may not be picked up at all!

invoke> a baseball -I /path/to/breakfast.png -tm orange 0.6

The !mask command may be useful for debugging problems with the text2mask feature. The syntax is !mask /path/to/image.png -tm <text> <threshold>

It will generate three files:

• The image with the selected area highlighted.
  • it will be named XXXXX...selected.png
• The image with the un-selected area highlighted.
  • it will be named XXXXX...deselected.png
• The image with the selected area converted into a black and white image according to the threshold level
  • it will be named XXXXX...masked.png

The .masked.png file can then be directly passed to the invoke> prompt in the CLI via the -M argument. Do not attempt this with the selected.png or deselected.png files, as they contain some transparency throughout the image and will not produce the desired results.

Here is an example of how !mask works:

invoke> !mask ./test-pictures/curly.png -tm hair 0.5
>> generating masks from ./test-pictures/curly.png
>> Initializing clipseg model for text to mask inference
Outputs:
[941.1] outputs/img-samples/000019.curly.hair.deselected.png: !mask ./test-pictures/curly.png -tm hair 0.5
[941.2] outputs/img-samples/000019.curly.hair.selected.png: !mask ./test-pictures/curly.png -tm hair 0.5
[941.3] outputs/img-samples/000019.curly.hair.masked.png: !mask ./test-pictures/curly.png -tm hair 0.5

Original image "curly.png"

000019.curly.hair.selected.png

000019.curly.hair.deselected.png

000019.curly.hair.masked.png

It looks like we selected the hair pretty well at the 0.5 threshold (which is the default, so we didn't actually have to specify it), so let's have some fun:

invoke> medusa with cobras -I ./test-pictures/curly.png -M 000019.curly.hair.masked.png -C20
>> loaded input image of size 512x512 from ./test-pictures/curly.png
...
Outputs:
[946] outputs/img-samples/000024.801380492.png: "medusa with cobras" -s 50 -S 801380492 -W 512 -H 512 -C 20.0 -I ./test-pictures/curly.png -A k_lms -f 0.75

You can also skip the !mask creation step and just select the masked

region directly:

invoke> medusa with cobras -I ./test-pictures/curly.png -tm hair -C20

Inpainting is not changing the masked region enough!

One of the things to understand about how inpainting works is that it is equivalent to running img2img on just the masked (transparent) area. img2img builds on top of the existing image data, and therefore will attempt to preserve colors, shapes and textures to the best of its ability. Unfortunately this means that if you want to make a dramatic change in the inpainted region, for example replacing a red wall with a blue one, the algorithm will fight you.

You have a couple of options. The first is to increase the values of the requested steps (-sXXX), strength (-f0.XX), and/or condition-free guidance (-CXX.X). If this is not working for you, a more extreme step is to provide the --inpaint_replace 0.X (-r0.X) option. This value ranges from 0.0 to 1.0. The higher it is the less attention the algorithm will pay to the data underneath the masked region. At high values this will enable you to replace colored regions entirely, but beware that the masked region mayl not blend in with the surrounding unmasked regions as well.

---

Recipe for GIMP

GIMP is a popular Linux photoediting tool.

1. Open image in GIMP.
2. Layer->Transparency->Add Alpha Channel
3. Use lasso tool to select region to mask
4. Choose Select -> Float to create a floating selection
5. Open the Layers toolbar (^L) and select "Floating Selection"
6. Set opacity to a value between 0% and 99%
7. Export as PNG
8. In the export dialogue, Make sure the "Save colour values from transparent pixels" checkbox is selected.

---

Recipe for Adobe Photoshop

1. Open image in Photoshop

   

2. Use any of the selection tools (Marquee, Lasso, or Wand) to select the area you desire to inpaint.

   

3. Because we'll be applying a mask over the area we want to preserve, you should now select the inverse by using the ++shift+ctrl+i++ shortcut, or right clicking and using the "Select Inverse" option.

4. You'll now create a mask by selecting the image layer, and Masking the selection. Make sure that you don't delete any of the underlying image, or your inpainting results will be dramatically impacted.

   

5. Make sure to hide any background layers that are present. You should see the mask applied to your image layer, and the image on your canvas should display the checkered background.

   

6. Save the image as a transparent PNG by using File-->Save a Copy from the menu bar, or by using the keyboard shortcut ++alt+ctrl+s++

   

7. After following the inpainting instructions above (either through the CLI or the Web UI), marvel at your newfound ability to selectively invoke. Lookin' good!

   

8. In the export dialogue, Make sure the "Save colour values from transparent pixels" checkbox is selected.