OpenSource Version of Magnific, 3-17-2024: https://twitter.com/philipp1337x/status/1768679154726359128

Replicate: https://replicate.com/philipp1337x/clarity-upscaler

Github: https://github.com/philz1337x/clarity-upscaler

Kijai CCSR node and tiled upscale workflow, 1-18-2024:

ccsr_tiled_upscale_kijai_01.json

ccsr_upscale_02.json

chrome_j4yQU752G8.gif

This workflow uses a wrapper node I made for CCSR (https://github.com/csslc/CCSR) :

https://github.com/kijai/ComfyUI-CCSR

As well as TiledDiffusion node from @mizu :https://github.com/shiimizu/ComfyUI-TiledDiffusion

CCSR is very good at accurate upscaling, but tends to introduce heavy seams when used in tiled mode, using another pass with TiledDiffusion clears them up nicely in most cases.

Depending on the input image you will probably have to adjust the start step until you see no seams.

Since I found that SDXL gave best results, and since SDXL lacks tile controlnet, I ended up using the recolor controlnet as it's the most similar: https://huggingface.co/stabilityai/control-lora/blob/main/control-LoRAs-rank256/control-lora-recolor-rank256.safetensors The downside is that it changes the colors a bit, which I'm compensating for with color-match.

chrome_zR7QzvtGXw.mp4

Draken Simple Highres Fix, 1-16-2024

simple_hiresfix.json

512.png

1024.png

Draken MultiDiffusion Tiled Upscale, 1-12-2024

Simple_MultiDiffusion_Upscale_Example.json

Untitled

We finally have a custom_node that does the full concept of Multidiffusion similar to how the multidiffusion upscaler for Auto1111 works !

This upscaling method is one of the best ive ever seen for trying to keep the upscale as close to the original SD output as possible. This is done by doing tiled diffusion(which has existed in comfy before), but it has controlnet support at the TILE LEVEL. This means each tile, knows exactly what is in its tile, thanks to using the Tile Controlnet(ya lots of tile lol).

As of now, this is very new, and there will most likely be a lot of bugs. I was struggling to do anything outside of 1:1 ratio, as an example. Will try keep this updated if the repo gets more changes, which i think it will. The dev added support for Kos Advanced ControlNets within hours of me asking, so that is really amazing of them! https://github.com/shiimizu/ComfyUI-TiledDiffusion

WORKFLOW: Pretty simple, will just generate an image with your prompts/seed/model and then upscale it right after. As with my other workflows posted, this is a simple workflow that lets you see what is going on a bit easier, which will make adding this to your own big workflows a lot easier.

RESOURCES: Upscale Model: https://openmodeldb.info/models/4x-Remacri

Perfume Commercial from SVD, Upres with AD