Yo! A lot of people keep asking me how I generate my images so here's a quick guide.
It's very simple.
HQ LoRA 😎
Good model/checkpoint (I prefer Heavenorangemix or Meinamix)
VAE: ClearVAE or (default) vae-ft-mse-840000-ema-pruned
Step #1: (txt-to-img) Prompts
Positives: Be specific what you want to see. My order is usually something like this:
quality prompts -> lora character prompt -> face expression -> eyes, hair -> clothes -> color of the clothes -> position of chara -> background
(worst quality, low quality:1.4), monochrome, muscular, (multiple views:1.2), deleted, lowres, comic, bad anatomy, text, error, blurry, loli, cropped, jpeg artifacts, signature, watermark, username, render, censored, mosaic censoring, bar censor, multiple girls, 3d, (upside-down:1.2)
You can also try textual inversion as negatives, such as easynegative.
I've been using these negatives for so long that I never really touch them anymore because I'm happy with the results.
Step #2: (txt-to-img) Settings
Set the settings accordingly:
You can choose between two upscalers: 4x-AnimeSharp or R-ESRGAN 4x+ Anime6B
Denoising strength: 0.6-0.7 (you have to experiment a little, as every LoRA gives different results)
Step #3: (txt-to-img) Experimenting
Adjust the prompts if something feels off. Don't look at imperfections like hands, just character in general, such as position, background or face expression.
Once you are happy with the result, bump up batch count to 6, 9, 12, 16, 24, whatever you want and press generate. This will take a while...
AI is not perfect with drawing hands or clothes exactly, this is the only way to find the best looking image.
When you find a decent image it's time to move to the next step.
Send image to img-to-img.
Step #4: (img-to-img) Settings
Prompts: Don't touch them, has to be the same as in txt-to-img.
Sampling method: Same from step #2.
Sampling steps: Same from step #2.
Resize to: Maximum your PC can handle. (You'll have to experiment with this one)
For nVidia RTX 4070Ti it's around 1500x1500 max, so if previous image is generated as 512x768, you can set it to resize by 2 or enter values manually.
CFG Scale: Same from step #2.
!!! Denoising strength: 0.1 !!!
Step #5: (img-to-img)
Generate image and wait for the result.
In general this is rather fast (5-10sec) at least on my PC.
Once the image baking is done, it's time to finish it with a crust.
Send the image to extras.
Step #6: (extras)
As of now, the image should already be in a rather good quality
but with upscaling you can enhance it a little bit more.
In general I never go over ~3000x3000 resolution.
So as previous image is 1024x1536, resizing it by 2 is plenty enough.
Upscaler is again up to you but I prefer to use 4x-AnimeSharp.
Wait for it....
And TA-DAAAAAA. Enjoy HQ anime image!
I use this process only for anime/manhwa images.
I have no idea what results you'll get with semi-realistic or realistic images. (careful with upscaler)
100 people = 100 different methods to generate HQ images
This method might have downsides and upsides, I don't really care. I've been using it since I got into SD, and so far it never failed me.