Upscalar with Tiled Diffusion

About

This is the scalar presented by @philz1337x where he combines MultiDiffusion, ControlNet tiles and details LoRas to obtain an interesting upscalar similar to Magnific's AI. The purpose of this article is to simply illustrate his technique when using Automatic1111

Steps

1. Switch to img2img

   1. Enter positive and negative prompts

      1. masterpiece, best quality, highres, <lora:more_details:0.7> <lora:SDXLrender_v2.0:1>

      2. (worst quality, low quality, normal quality:2) JuggernautNegative-neg
         

2. Use the following settings
   Steps: 18
   Sampler: DPM++ 3M SDE Karras
   CFG Scale: 6
   Size: 1024x1024
   Model: juggernaut_reborn
   Denoising Strength: 0.35
   

3. Enable Tiled Diffussion, configure as follows
   Upscalar: 4x-UltraSharp
   Tiled Diffusion: MultiDiffusion
   Tile tile width: 112
   Tile tile height: 144
   Tile Overlap: 4
   Tile batch size: 8
   Upscale factor: 2
   Keep input size: true
   

   
   (https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111)
   

4. Enable ControlNet 0, configure as follows
   Module: tile_resample
   Model: control_v11f1e_sd15_tile
   Weight: 0.6
   Resize Mode: 1
   Low VRam: False
   Pixel Perfect: True

   

5. Drop an image on img2img

   1. Start with a 768x768 image, this process is memory hungry.
      

6. Generate