This feature runs by scaling the latent sample of the image, and then running a second pass of img2img which results in insane quality boost!
- A tremendous quality boost
- Can sometimes make the output image more deformed
This guide was tested using BakaDiffusion
to exactly replice what i did. use the Fp16 Model from my HuggingFace Repo above!
Let's say you want to latent space upscale an image of Misato Katsuragi for example.
masterpiece, best quality, 1girl, eye focus, katsuragi misato, ribbed sweater, purple hair, brown eyes
Negative prompt: lowres, (bad anatomy), text, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, username, artist name, (out of frame), black and white, obese, censored, child, loli, long fingernails
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1506386276, Size: 512x512, Model hash: 63c46c36, ENSD: 31337
The rest of the settings will stay the same, We'll be changing out FirstPass and output resolution, We'll be doubling the original resolution(512x512) which will become(1024x1024) and we'll be pasting our original resolution(512x512) into our first pass! and our output into (1024x1024) if everything's done correct it should look like the image above! ⇈
Sweet spot for denoising setting would be around 0.5 to 0.7 for me, Anything lower than 0.5 may give you blurred image