Since many people ask me how I upscale my images, I'll write a brief explanation of my workflow.
Create Image
First, create an image with a normal prompt:
unsplash, hot summer, woman sun hat
Steps: 50, Sampler: DPM++ 2M SDE, Schedule type: Exponential, CFG scale: 5, Seed: 3134790085, Size: 864x1152, Model hash: 61fcd93735, Model: epiCRealismXL-LastFAME, Clip skip: 2, RNG: CPU, Version: f2.0.1v1.10.1-previous-662-g0ced1d0c
Upscale
Now we will make use of the Hires Fix with these Settings below (i'll explain more later)
The most important parts are:
using same Prompt and add the DMD Lora after it
use sampling method: LCM Exponential
or for more details but less restauration: Euler A Karras
Hires CFG: 1
Hires Steps: 4 (go lower on LCM if it gets to plastic or getting hallucinations)
Denoising: 0,3
If you dont see the additional Prompt Section, make sure you checked these under settings
Now, as we have all settings set. just use the Star-Button under the generated image to use the HiRes Settings without generating the image again.
The Result:
As we can see it is crystal clear with these settings. Probably to clear, so let me Explain in detail.
DMD2 Lora has strong Dataset which is great for "repairing" the image at upscaling but also adds the own "plastiness" if you use LCM Sampler. So if you want a bit less restauration or "plastiness" use lower Hires Steps like 2. Also if you see some "hallucinations" like double breasts or wrinkles and stuff.
You always can use Euler A Karras to keep the image as original as possible and just upscale, like below.
Side Notes
Another Example to see more clearly the restauration proccess and "plastiness" involvement of LCM compared to Euler A (DDIM also can work)
Initial Image:
full-body portrait of a young woman wearing a dress with floral patterns, subtle makeup
Steps: 30, Sampler: DPM++ 2M SDE, Schedule type: Exponential, CFG scale: 5, Seed: 1337, Size: 832x1216, Model hash: 61fcd93735, Model: epiCRealismXL-LastFAME, Clip skip: 2, RNG: CPU, Version: f2.0.1v1.10.1-previous-662-g0ced1d0c
LCM (4 Steps)
Conclusion: Good Eye and finger restauration, bad texture
Euler A:
Conclusion: Bad restaurations, better textures
What we can do to have something of both worlds is using upscale by 1.5x with Euler A Karras -> LCM Exponential or LCM Exponential -> Euler A Karras
Comparsion:
1.5x Euler A -> 1.5x LCM
1.5x LCM -> 1.5x Euler A
Funny Note
You also can play around with different prompt at upscaling and get funny results
unsplash, icecream, sunset
Steps: 30, Sampler: DPM++ 2M SDE, Schedule type: Exponential, CFG scale: 5, Seed: 2258927394, Size: 832x1216, Model hash: 61fcd93735, Model: epiCRealismXL-LastFAME, Clip skip: 2, RNG: CPU, Version: f2.0.1v1.10.1-previous-662-g0ced1d0c
Denoising strength: 0.4, Hires prompt: "unsplash, bird, sunset <lora:dmd2_sdxl_4step_lora_fp16:1>
Here the upscale process is looking for stuff that would match the prompt mostly and tries to implement it into the original image.
In most cases LCM Exponential with 4 Steps works great.
Have fun experiment around and see what it does for you.
Links
DMD2 Lora (i use dmd2_sdxl_4step_lora_fp16.safetensors)
Attachment
Added Simple ComfyUI Workflow as Attachment