VoltaMLHighres Fix
Method that allows you to generate high resolution images without duplicating subjects.
Currently available on the Text To Image page.
INFO
Generated using UniPCMultistep (diffusers)
Hires scale 2x in cases where hires is applied.
Image Upscaler RealESRGAN_x4plus_anime_6B
Latent Upscaler Bislerp (Tortured, fast)
Steps: 30 base, 40 hires Strength: 0.65
Usually, Latent mode tends to perform better than Image mode, but it will get blurry with strength under 0.6. It is usually further from the original image. See example below.
Base VS Latent Upscale
Base VS Image Upscale
Image Upscale VS Latent Upscale
Base VS High resolution base
Even though the model I use is trained on higher resolution images, it still duplicated the railings. This gets more noticeable with older models that were trained only on 512x512 images.
More recent models or SDXL might be totally fine with higher resolution images.
Workflow diagram
