SBS Guided Example

Step-by-Step Guided Example

Preface

Disclaimers

EXACTNESS OF IMAGE IS NOT GUARANTEED, PARTICULARLY IF USING XFORMERS and probably many other options but xformers is infamous for being somewhat nondeterministic

Assumptions

This example makes a few assumptions that may not be the same as how your local environment is configured. For example, any keyboard commands involving [CTRL] would be [CMD] for the Mac equivalent. The following assumptions are made:

• webUI will be hosted at http://127.0.0.1:7860
• openOutpaint will be hosted at http://127.0.0.1:3456

You can host openOutpaint and/or webUI on completely different machines, specify alternative ports, require HTTPS, etc - if you're doing so, simply use the value appropriate to your installation instead of the defaults listed above.

Prerequisites

This example also has the following prerequisites:

• webUI at commit 44c46f0 or more recent
• openOutpaint at v0.0.9/commit 0b5b018 or more recent
• model sd-v-1-5-inpainting.ckpt 3e16efc8 (and preferably the SD1.5 VAE (VAE installation instructions))
• SD webUI option Eta noise seed delta set to value 31337 - note this is not a prerequisite of openOutpaint, but merely to get as-close-to-identical results to this example guide's output

Intro

This example will be a step-by-step walkthrough with specific models, seeds, prompts, and other values used to ensure identical results are able to be achieved. Please don't deviate from the values supplied for anything not listed in the above assumptions section to ensure you receive the same output and essentially make sure all's working as expected.

The intent of this guide is by no means whatsoever to produce anything even remotely confusable with Art so don't expect a masterpiece, best quality, masterpiece, 4k, trending on artstation, greg rutkowski. This is merely a quick introduction to how openOutpaint's feature set works and is not a comprehensive usage manual.

Screenshots were taken using Firefox on Windows, so your browser may appear differently than some of the screenshots. This should not affect openOutpaint nor Stable Diffusion.

Step-by-Step

• In your webUI installation directory, open webui-user.sh/bat in a text editor, and ensure the line starting with set COMMANDLINE_ARGS= contains the following: --api --cors-allow-origins=http://127.0.0.1:3456, then save the file.

• Run the newly-saved webui-user script file and wait for it to report that webUI is ready.
  (please note: http://0.0.0.0:7860 is a result of the --listen COMMANDLINE_ARGS flag which is only necessary if you wish to access webUI from another computer on your LAN. Your instance will likely report http://127.0.0.1:7860 instead if you do not have that flag enabled.)

• Once webUI is successfully running, open the directory where you have cloned/unzipped openOutpaint into and run openOutpaint.bat/sh to launch the tiny Python local webserver. It will state that it is serving HTTP on port 3456 as specified in the batch or shell script file.

• Now that both applications are running, go to your web browser and navigate to http://127.0.0.1:3456. On the first launch, you will be prompted to confirm the webUI API address. This can of course be updated later if necessary.

• After confirming the host address, a few quick checks are made to ensure maximum outpainting compatibility, and openOutpaint offers to update a few webUI settings to the recommended values automatically if non-optimal values are detected. The changes are:

  • option Inpainting conditioning mask strength, sets value to 1.0
  • option Apply color correction to img2img results to match original colors., sets value to false/disabled

• Welcome to openOutpaint! We'll be leaving the default prompts alone, so let's start with the Stable Diffusion settings button. Click that to expand the available options, and set the following:

  • Model: SD v1.5 Inpainting (titled runwayML-sd-v1-5-inpainting in screenshots)
  • Sampler: DDIM
  • Seed: 2346
  • Auto txt2img HRfix: disabled
  • Resolution: 512
  • Steps: 30
  • CFG Scale: 7
  • Batch Size: (select a quantity appropriate for your available VRAM, max of 8)
  • Iterations: 1

• Once you've got those set, you can click the Stable Diffusion settings button again to minimize those options if desired. Move the mouse cursor somewhere over the checkerboard background and click to start the txt2img process which we will lovingly call "dreaming". You'll be presented with a progress estimations at the top of the dream region, and an Interrupt button to rudely halt the dreaming. When the final image appears, a new set of buttons will be presented beneath the dream - click "Y" or press your [↩] key to accept the first image for now.

• Time to outpaint! Leaving Snap to Grid enabled, click to the southwest of the original image, expanding the lower portion by 2 rows, and the left edge by 3 columns.

• Whoops, there's a pretty rough seam there, quite unpleasant. Let's click the "N" button or press your [ESC] key to cancel that batch.

• At the bottom of the main openOutpaint options panel, drag the Overmask px slider to 20 pixels, and outpaint over the same area of image once more. Much better. Click through the rendered images if your batch size is greater than 1 (you can also request an additional batch by clicking the "+" button or pressing the [+] key on your keyboard but it's useless with a fixed seed), however for the purposes of this example, select the first image to continue. You can quickly jump back to the first image in your current accept/reject batch by clicking the "X of Y" button between the < and > buttons.

• Let's give these fish more room to explore. Expand the left edge by 4 columns, keeping the bottom edge aligned with the new lower boundary. Unless otherwise noted, from here on out it'll be assumed that the first image in a batch will be used for the rest of this example guide.

• Now let's give them some vertical space too. Use your mouse scrollwheel or the Cursor Size slider to double the size of the reticle all the way up to 1024 px. Leave the existing dream collage at the southwest corner of the new boundary box.

• Hmm, kind of blurry isn't it? openOutpaint's cursor size isn't tied to the requested render size, meaning you can request a large dream in a small target area, or vice versa, or use it as "intended" and match them - all up to you. Click "N" to cancel the quarter-resolution dream. Click the Stable Diffusion settings panel button if you closed it previously, and ensure the following options are set as follows:

  • Auto txt2img HRfix: enabled
  • Resolution: 1024
  • Batch Size: (you may need to reduce this value in the event that your VRAM can't handle the quadrupled resolution demands, and you'll receive an error about it if that happens)

Now click in the same area as the previous step to give our fish some vertical room, but in the correct resolution this time.

• Certainly better coherence and visual clarity, but the sad guy in the middle is a little depressing. Let's fix him. Click the Mask Brush tool (keyboard shortcut: [M]) in the floating toolbar on the right-hand side of the screen, and let's paint out the sad like it never existed. You can click the "Preview" button in the Context Menu on the left-hand main options to see the mask without transparency if desired.

• Now that we've masked away the sad, select the dream tool (keyboard shortcut: [D]), scroll your mousewheel down to a cursor size of 512 and set the Stable Diffusion settings resolution to match, and make sure the txt2img HRfix option is disabled. Since the region is masked, you don't need to be precise about where you click on the canvas as long as the entire masked region is within the boundary reticle. NOTE: Resultant output images will certainly differ just due to human-drawn mask differences. Determinism is no longer applicable as human variables and behavior cannot be accurately quantified with current models.

• Hmm, that's really not a lot better. Cancel that by clicking "N", then choose the mask brush tool again and click the "Clear" button at the bottom of the Context Menu, or you can also "paint" with a right-click over the masked region to erase it. After you've erased the mask, select the img2img tool (keyboard shortcut: [I]), crank its Denoising Strength option to 0.95 and click over that sad fish walrus thing.

• Well alright, seems our sad creature has simply been absorbed into the coral, but that's probably for the best anyway. Let's do something with that other fish at the northwest corner. Click the Prompts button and change the primary prompt to an octopus, then click the img2img tool over the fish, aligning with the northwest corner of the dream collage.

• Now we have an octopus.

• Just out of curiosity, let's click the interrogate tool (keyboard shortcut: [N]) and click directly over the octopus

• Now I can't say I understand how deterministic CLIP is for given visual inputs and I'm pretty sure I've set some various options I've forgotten about so you may have different results, but I was informed that the octopus looks like the following:

• As long as you received some kind of result, simply click OK to replace your prompt just for fun. Select the img2img tool again and click the northeast corner of your dream collage. Click the Y button immediately without a second's thought when it completes.

• I, uh... I got this.

• Maybe I shouldn't have been so cavalier suggesting we click Y without a second's thought. Press [CTRL]+[Z] or click the "undo" button in the floating History palette to remove... whatever it is we just made. The history palette should reflect the changes we just made by highlighting the previous action. You can click around in here freely if you'd like, but return to the penultimate action once you're done.

• Let's try to forget the tragedy of the sad fish walrus thing. Click the "Add Layer" button in the floating Layers Palette, likely in the bottom-right of your window. Select the color brush tool (keyboard shortcut: [C]) and draw a barrel in the southeast corner of the dream collage. Let out your inner art major. Change paint colors with the large button on the bottom of the Context Menu, or use the eyedropper tool to select a color already on the canvas.

• Change your prompt to a barrel, select the img2img tool, disable the Invert Mask option, and click over the "barrel". NOTE: Resultant output images will guaranteed differ from this point forward unless you drew PRECISELY the same barrel as I did.

• Click the Select Image tool (keyboard shortcut: [S]) and draw a box around your shiny new barrel. Click the "Resource" button at the bottom of the Context Menu, and start randomly left-clicking to stamp your new resource. You can right-click to deactivate the stamp resource when you're done playing around, then [CTRL]+[Z] your way back out of that mess, or just click on the "Image Dream" history entry prior to the "Image Stamp" entries.

You can also scale stamped resources or selections using the handles arbitrarily.

At this point, you should have a decent understanding of the basic functionality of openOutpaint and how to use its main features - have fun and play around with all the options and tools! You can always reset back to factory state using the "Reset to defaults" button in the Debug Info menu if something goes particularly out of control.