Image Upscalers and Detailers
This page is a work in progress
Simplest: Upscale Image By
Use stock ComfyUI node Upscale Image By with upscale_method=lanczoc
Detailer (SEGS)
There are two Github repositories to get nodes from
USDU
Ultimate SD Upascaler, a popular choice.
Interestingly includes a node Ultimate SD Upscale (No Upscale) which is used to fix artifacts.
GH:ssitu/ComfyUI_UltimateSDUpscale
SeedVR2
GH:numz/ComfyUI-SeedVR2_VideoUpscaler useful for both images and videos
ERSGAN
for upscaling, those models are like the ultimate adversarial models, they literally only have adversarial training
Chatter
In my tests, z image doesn’t work well for second pass. Instead, I’m using wan 2.2; you can also use flux
[general talk about training] you can do consistency model training for most of the steps then do adversarial at the end, IDK why that isn’t more popular its what the Flux devs recommended
For ZIT
- GH:wildminder/ComfyUI-DyPE “ok I give up on DyPE 4096, but going to keep using dype 2048, it’s just barely manageable” - overall negative sentiment like “I still can’t fix the stupid dype long body problem”
- GH:spacepxl/ComfyUI-VAE-Utils
- blend-masked-noise wf which allows exclusing skip from noise blend to prevent “zombie skin” while adding details
- combined ZIT upscaling “When the image is blurry, I use qwen edit to recover details: input > qwen edit > seedvr > zit”;
“Upscale the image to a higher resolution, reduce compression artifacts, noise, and blur while preserving original details and colors. Do not change content, pose, face, body, age, gender, clothing, background, or camera angle”;
“do you do any sort of tiling? zIT can’t really do more than 2.5MP without tiling? yes USDU”;
“seedvr2 first then usdu? I do the reverse lol, cause seedvr2 adds microdetails”;
“I do it this way because seedvr sometimes generate too much noise. and a second pass to reduce this noise”
- zimage-upscaleAny by VRGameDevGirl “it’s literally a 0.5 denoise of an upscaled image lol”
Hmm
What is Supir x3? TTP?
