HOME | DD

Description
Yes, there's a bit of tree on the sky. That's how giant tree reproducers, almost like a dandelion.

8K image available as premium download (actual 8K image, not upscaled full HD image).

This image was part of my test:
Testing out AI upscaling methodsHello,I've spent last few days trying out different upscaling method, so I wanted to share my findings, as well as ask for opinions. Please excuse my language, English is not my native language, and I'm not a good writer either way. Short intro to methods: Previously (labeled as old upscaler), I have been using "classic" upscalers, specifically mix of SwinIR_4x and 4x-UltraSharp. Those methods will increase image resolution (in my case from Full HD to 4K) basically by making slight adjustments to stretched original image. They add no details, and do not look great when zoomed in - but improve image quality compared to same image without upscaler. New method I'll introduce to this post (labeled as new upscaler), is Ultimate Upscale: a plugin which (in conjunction with ControlNet tile model) splits image into smaller tiles, makes new image using old one as basis and adds detail, then places tiles together to get higher resolution image with more details. Basically every tile has same level of detail as if each was image in itself.* note: With current limitations, I can't use original model for upscaling (SDXL models are not yet supported by Ultimate Upscale in AUTO1111) so I used older 1.5 model for that. There are few implications of that, one being mismatched color tones, as well as potentially different understanding of prompt in some places. So lets dive into examples. I have 3 landscape images and 3 character images (starting at Full HD/1920x1080) and applied both upscaling methods on them (to get 4K/3840x2160 images). To compare, I'll use image for each method, and one which is split (roughly) in the middle for side by side comparison. Let's start with characters first:Character 1: Old Upscaler:, New Upscaler:, Comparison:, Comment: notice skull on the left side in 2nd image. This is example how my old 1.5 model didn't recognize a skull, so it tried to recreate a face. I'm often getting such "artifacts" with new method. Skin tone is also not consistent. On the other side, ship looks amazing. Character 2: Old Upscaler:, New Upscaler:, Comparison:, Comment: Her left hand, and again skulls got messed up with new method. Rock textures have however greatly improved.Character 3: Old Upscaler:, New Upscaler:, Comparison:, Comment: Again we have artifacts: her right hand and glass right next to it. Notice the leather on the roof above her right hand. On the other side, we see great improvements in quality of clothes, skin and hair.Character conclusion: New method definitely creates more details, but it also creates undesirable artifacts. Additionally, since image after upscaling is 4K, I have no easy way to inpaint those artifacts. For now, as Ultimate Upscale (new method) takes much longer (about 15 minutes compared to 1 minute, plus I have to alter between SDXL and SD1.5 models+vae) I think I'll stick to old upscaler for the most part. I might use it for images where character is very far from camera. Once I get ControlNet tile model working on SDXL, I'll revisit it.* note: Unlike characters, landscapes were actually upscaled to 8K, then downscaled to 4K. I tried same with characters, but got too many artifacts (it got exponentially worse).Landscape 1: Old Upscaler:, New Upscaler:, Comparison:, Comment: New method is simply better.Landscape 2: Old Upscaler:, New Upscaler:, Comparison:, Comment: Again, ultimate upscale improves quality.Landscape 3: Old Upscaler:, New Upscaler:, Comparison:, Comment: Better lightning, better colors.Landscape conclusion: Pain points I had with characters with Ultimate Upscaling (new method) are not present here, I simply got clearer images with more details. I can also get solid 8K images this way. And since I rarely inpaint landscape images, inability to do so is not an issue. I think I'll start using this method from now on on landscape shots. It does take quite longer though - about half hour for one image. I believe it's worth it, though. What do you think? Which method do you prefer? Do you have any tips? Questions? Thanks for reading this to the end!

Description
Yes, there's a bit of tree on the sky. That's how giant tree reproducers, almost like a dandelion.

8K image available as premium download (actual 8K image, not upscaled full HD image).

This image was part of my test:
Testing out AI upscaling methodsHello,I've spent last few days trying out different upscaling method, so I wanted to share my findings, as well as ask for opinions. Please excuse my language, English is not my native language, and I'm not a good writer either way. Short intro to methods: Previously (labeled as old upscaler), I have been using "classic" upscalers, specifically mix of SwinIR_4x and 4x-UltraSharp. Those methods will increase image resolution (in my case from Full HD to 4K) basically by making slight adjustments to stretched original image. They add no details, and do not look great when zoomed in - but improve image quality compared to same image without upscaler. New method I'll introduce to this post (labeled as new upscaler), is Ultimate Upscale: a plugin which (in conjunction with ControlNet tile model) splits image into smaller tiles, makes new image using old one as basis and adds detail, then places tiles together to get higher resolution image with more details. Basically every tile has same level of detail as if each was image in itself.* note: With current limitations, I can't use original model for upscaling (SDXL models are not yet supported by Ultimate Upscale in AUTO1111) so I used older 1.5 model for that. There are few implications of that, one being mismatched color tones, as well as potentially different understanding of prompt in some places. So lets dive into examples. I have 3 landscape images and 3 character images (starting at Full HD/1920x1080) and applied both upscaling methods on them (to get 4K/3840x2160 images). To compare, I'll use image for each method, and one which is split (roughly) in the middle for side by side comparison. Let's start with characters first:Character 1: Old Upscaler:, New Upscaler:, Comparison:, Comment: notice skull on the left side in 2nd image. This is example how my old 1.5 model didn't recognize a skull, so it tried to recreate a face. I'm often getting such "artifacts" with new method. Skin tone is also not consistent. On the other side, ship looks amazing. Character 2: Old Upscaler:, New Upscaler:, Comparison:, Comment: Her left hand, and again skulls got messed up with new method. Rock textures have however greatly improved.Character 3: Old Upscaler:, New Upscaler:, Comparison:, Comment: Again we have artifacts: her right hand and glass right next to it. Notice the leather on the roof above her right hand. On the other side, we see great improvements in quality of clothes, skin and hair.Character conclusion: New method definitely creates more details, but it also creates undesirable artifacts. Additionally, since image after upscaling is 4K, I have no easy way to inpaint those artifacts. For now, as Ultimate Upscale (new method) takes much longer (about 15 minutes compared to 1 minute, plus I have to alter between SDXL and SD1.5 models+vae) I think I'll stick to old upscaler for the most part. I might use it for images where character is very far from camera. Once I get ControlNet tile model working on SDXL, I'll revisit it.* note: Unlike characters, landscapes were actually upscaled to 8K, then downscaled to 4K. I tried same with characters, but got too many artifacts (it got exponentially worse).Landscape 1: Old Upscaler:, New Upscaler:, Comparison:, Comment: New method is simply better.Landscape 2: Old Upscaler:, New Upscaler:, Comparison:, Comment: Again, ultimate upscale improves quality.Landscape 3: Old Upscaler:, New Upscaler:, Comparison:, Comment: Better lightning, better colors.Landscape conclusion: Pain points I had with characters with Ultimate Upscaling (new method) are not present here, I simply got clearer images with more details. I can also get solid 8K images this way. And since I rarely inpaint landscape images, inability to do so is not an issue. I think I'll start using this method from now on on landscape shots. It does take quite longer though - about half hour for one image. I believe it's worth it, though. What do you think? Which method do you prefer? Do you have any tips? Questions? Thanks for reading this to the end!