Experiments With fantasticmixReal_v20 model for Stable Diffusion
You probably saw the image and went "what the f^$&*% is that.... and why did he post it on his shop...."
Well you see latent space, it piques my interest immensely....it's kinda like liminal spaces in the fact that normally it would merely be a vast wasteland connecting two spots in space. At least that's how I think of it, whether that is necessarily true, I'm still digging into that mountain of data science.
So what, exists there... Well in the above image all variables are identical with the exception of two.
Sampler - The algorithm that is used to generate the image.
CFG Scale - How strongly/heavily the system tries to stick to my prompts.
Here is the config I used to generate this massive example of how varying even the same words can be depending on only 2 of the settings being different
The Config!
Prompt: depth_of_field,deep_focus,sharp_focus,full_shot, style-empire, emma stone , (extremely_detailed_CG_unity_8k_wallpaper:1.2), , (lust_demon:1.3),spell-casting,sheer_blood_mage_robes,tattoos,(glowing_vivid_inked_tattoos:1.4), trending on ArtStation, trending on Cg-society, Intricate, High Detail, Sharp focus, dramatic, photo-realistic art , epic, heroic, grand, fantasy,blood_magic,radioactive_glowing_eyes
Negative prompt: ,simple_background,text,watermark,logo,signature, easynegative, watermark, logo, text,signature,bad_prompts:0.8
Steps: 20,
Sampler: Euler a,
CFG scale: 1.0,
Seed: 2421357785,
Size: 632x272,
Model hash: dc68ca14e7,
Model: fantasticmixReal_v20,
Variation seed: 3880289778,
Variation seed strength: 0.5,
Denoising strength: 0.63,
Hires upscale: 2,
Hires steps: 16,
Hires upscaler: R-ESRGAN General WDN 4xV3,
AddNet Enabled: True,
AddNet Module 1: LoRA,
AddNet Model 1: lightAndShadow_setOff(cd947236ac78),
AddNet Weight A 1: 0.3,
AddNet Weight B 1: 0.3,
AddNet Module 2: LoRA,
AddNet Model 2: darkAndLight_v10(a2e2d712da1e),
AddNet Weight A 2: 0.55,
AddNet Weight B 2: 0.55,
Script: X/Y/Z plot,
X Type: Sampler, X Values:
"Euler a, Euler, LMS, Heun, DPM2, DPM2 a, DPM++ 2S a, DPM++ 2M, DPM++ SDE, DPM fast, DPM adaptive, LMS Karras, DPM2 Karras, DPM2 a Karras, DPM++ 2S a Karras, DPM++ 2M Karras, DPM++ SDE Karras, DDIM, PLMS, UniPC",
Y Type: CFG Scale, Y Values:
"1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30"
How did this hit 30 CFG and not look TOASTED?!
Honestly, I'm not even sure yet, but if I had to guess, I'd say the person who mixed this model is either lucky as hell, or took great care in the concoction to produce this:
DPM2 Karras CFG 30.0
Is it the LoRA's balancing it out? Or perhaps the Upscaling Hi-Res fix, or the combination of all of it