Matthew Crook's profile

Upscale and Improve

I previously mentioned that to make an animation, the Deforum extension takes an image, applies any transformations (shifts, rotations, or zooms) that the user specifies, and then adds noise to it (see Deforum I). Then it has Stable Diffusion denoise the image according to the prompt to produce the next frame of the video.

Well, this can also be done to a single image. To show you what I mean, I took some of the images of dioramas from a recent post (see Dioramas III), upscaled them, added some noise back in, and had Stable Diffusion denoise them again. As you'll see, this results it a much more detailed image. Sometimes this means that details from the original are lost and replaced with something else. But, overall, the final outcome is superior to the original. (While I was at it, I also increased the contrast a little bit.)

Here are the results. On the left is the original image from Stable Diffusion 1.5 and on the left is the upscaled and improved image.

Alhambra

Barn

Baroque House

Bunker

Car Wash

Catacombs

Chinatown

Church

Crab Shack

Fallout Shelter

Ferris Wheel

Fountain (v1)

Fountain (v2)

Gas Refinery

Geodesic Dome

Graveyard

Haberdashery

Haunted House

Hospital

Jail Cell

Jumbo Jet

Landfill

Missile Silo

Oil Derrick

Opium Den

Pagoda

Pier

Prison

Roller Coaster

Skyscraper

Temple

Spray Fountain

Stadium

Tomb

Tudor House

Tunnel

Victorian House

Water Park

Water Tower

Zoo


These illustrations using Stable Diffusion 1.5 as implemented in SD.Next with the DreamShaper 7 model.
Upscale and Improve
Published: