Description:
Thanks to [Mr. Bat](/profiles/Mr%2BBat) for the use of his wonderful Daybreaker Dreambooth.
This is something I've been thinking about doing for some time, combining the results of multiple specialized models into one coherent image. Naturally I'm being a little bit conservative with this first outing as the characters are not truly interacting and this more merely demonstrates the models ability to not immediately explode any lesser understood characters who happen to overlap with a primary subject. Woo, small victories. In the future I'd like to explore more dynamic interactions like hugging, etc but in the meantime, this will do.
I started by generating two 1:1 images of Daybreaker and NMM each, using Mr. Bat's Daybreaker model, my own NMM model and a few of my own embeddings. In photoshop I quite crudely stitched the images together and began img2img runs until they had a coherent shared background across the width of the image. Following that I began hotswapping between models to rebuild and redesign both characters before taking the image back together to paint in Daybreakers mane, the mist effect and both 'glowing horn' effects, all of which are largely original painting with the exception of daybreakers mane which was largely lost over further iterations and error corrections slowly eroded any original work.
once I had a near final 512x768 image, I used the lollypop upscale model to bring the image up to its current resolution, after that I split the image into separate 1280x1280 chunks and began error correcting some of the more egregious 'ai upscale' artifacts, either by hand or successive low denoise inpainting. This largely took the form of correcting blurred lines, shapes that dissolved into nothingless, coloring artifacts and one very troublesome feather.
After that, I reconstructed the image and did some simple denoising in photoshop plus /sharpen/minor color correction/etc.
This is something I've been thinking about doing for some time, combining the results of multiple specialized models into one coherent image. Naturally I'm being a little bit conservative with this first outing as the characters are not truly interacting and this more merely demonstrates the models ability to not immediately explode any lesser understood characters who happen to overlap with a primary subject. Woo, small victories. In the future I'd like to explore more dynamic interactions like hugging, etc but in the meantime, this will do.
I started by generating two 1:1 images of Daybreaker and NMM each, using Mr. Bat's Daybreaker model, my own NMM model and a few of my own embeddings. In photoshop I quite crudely stitched the images together and began img2img runs until they had a coherent shared background across the width of the image. Following that I began hotswapping between models to rebuild and redesign both characters before taking the image back together to paint in Daybreakers mane, the mist effect and both 'glowing horn' effects, all of which are largely original painting with the exception of daybreakers mane which was largely lost over further iterations and error corrections slowly eroded any original work.
once I had a near final 512x768 image, I used the lollypop upscale model to bring the image up to its current resolution, after that I split the image into separate 1280x1280 chunks and began error correcting some of the more egregious 'ai upscale' artifacts, either by hand or successive low denoise inpainting. This largely took the form of correcting blurred lines, shapes that dissolved into nothingless, coloring artifacts and one very troublesome feather.
After that, I reconstructed the image and did some simple denoising in photoshop plus /sharpen/minor color correction/etc.
not provided yet
Comments
0 comments posted