描述Demonstration of inpainting and outpainting using Stable Diffusion (step 4 of 4).png
Demonstration of the usage of inpainting and outpainting techniques on algorithmically-generated artworks created using the Stable Diffusion V1-4 AI diffusion model. Not only is Stable Diffusion capable of generating new images from scratch via text prompt, it is also capable of providing guided image synthesis for enhancing existing images, through the use of the model's diffusion-denoising mechanism.
This image aims to illustrate the process in which Stable Diffusion can be used to perform both inpainting and outpainting, as one part out of four images showing each step of the procedure.
Procedure/Methodology
All artworks created using a single NVIDIA RTX 3090. Front-end used for the entire generation process is Stable Diffusion web UI created by AUTOMATIC1111.
First image: Generation via text prompt
An initial 512x768 image was algorithmically-generated with Stable Diffusion via txt2img using the following prompts:
Prompt: busty young girl, art style of artgerm and greg rutkowski
Negative prompt: (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing, two heads, four breasts
Then, two passes of the SD upscale script using "Real-ESRGAN 4x plus anime 6B" were run within img2img. The first pass used a tile overlap of 64, denoising strength of 0.3, 50 sampling steps with Euler a, and a CFG scale of 7. The second pass used a tile overlap of 128, denoising strength of 0.1, 10 sampling steps with Euler a, and a CFG scale of 7. This creates our initial 2048x3072 image to begin working with. Unfortunately for her (and fortunately for the purpose of this demonstration), it appears that the AI neglected to give this woman one of her arms.
Second image: Outpainting
Using the "Outpainting mk2" script within img2img, the bottom of the image was extended by 512 pixels (via two passes, each pass extending 256 pixels), using 100 sampling steps with Euler a, denoising strength of 0.8, CFG scale of 7.5, mask blur of 4, fall-off exponent value of 1.8, colour variation set to 0.03. The prompts used were identical to those utilised during the first step. This subsequently increases the image's dimensions to 2048x3584, while also revealing the woman's midriff, belly button and skirt, which were previously absent from the original AI-generated image.
Third image: Preparation for inpainting
In GIMP, I drew a very shoddy attempt at a human arm using the standard paintbrush. This will provide a guide for the AI model to generate a new arm.
Final image: Inpainting
Using the inpaint feature for img2img, I drew a mask over the arm drawn in the previous step, along with a portion of the shoulder. The following settings were used for all passes:
Inpaint masked
Masked content: original
Inpaint at full resolution, padding at 256 pixels
Steps: 80, Sampler: Euler a
An initial pass was run using the following prompts:
Prompt: perfect arm, young woman's arm, (((anterior elbow))), (((inside of elbow))), bent arm, slender arm, realistic arm, wrinkled short sleeve of white blouse, woman's shoulder, brown hair on top of sleeve, (((pale skin))), skin on arm, smooth skin, art style of artgerm and greg rutkowski
This created the arm; another subsequent pass was then done to fine-tune deformations and blemishes around the newly generated arm along the sleeve. Drawing a new mask over the shoulder, the following prompt was used:
Prompt: brown hair on top of sleeve and arm, wrinkled short sleeve of white blouse, young woman's upper arm beside her chest, woman's shoulder, skin under sleeve, art style of artgerm and greg rutkowski
As the creator of the output images, I release this image under the licence displayed within the template below.
Stable Diffusion AI model
The Stable Diffusion AI model is released under the CreativeML OpenRAIL-M License, which "does not impose any restrictions on reuse, distribution, commercialization, adaptation" as long as the model is not being intentionally used to cause harm to individuals, for instance, to deliberately mislead or deceive, and the authors of the AI models claim no rights over any image outputs generated, as stipulated by the license.
Addendum on datasets used to teach AI neural networks
Artworks generated by Stable Diffusion are algorithmically created based on the AI diffusion model's neural network as a result of learning from various datasets; the algorithm does not use preexisting images from the dataset to create the new image. Ergo, generated artworks cannot be considered derivative works of components from within the original dataset, nor can any coincidental resemblance to any particular artist's drawing style fall foul of de minimis. While an artist can claim copyright over individual works, they cannot claim copyright over mere resemblance over an artistic drawing or painting style. In simpler terms, Vincent van Gogh can claim copyright to The Starry Night, however he cannot claim copyright to a picture of a T-34 tank painted with similar brushstroke styles as Gogh's The Starry Night created by someone else.
I didn't like how dark the arm turned out, so I re-did it again. Redrew the makeshift arm in a lighter shade in GIMP, then ran two passes of inpainting: Steps: 80, Sampler: Euler a, CFG scale: 17, Seed: 525737653, Denoising strength: 0.6 to generate arm; Steps: 80, Sampler: Euler a, CFG scale: 7, Seed: 653575127, Denoising strength: 0.4 for cleanup.
{{Information |Description=Demonstration of the usage of inpainting and outpainting techniques on algorithmically-generated artworks created using the [https://github.com/CompVis/stable-diffusion Stable Diffusion V1-4] AI diffusion model. Not only is Stable Diffusion capable of generating new images from scratch via text prompt, it is also capable of providing guided image synthesis for enhancing existing images, through the use of the model's diffusion-denoising mechanism. This image aims t...