Stable Diffusion has many knobs to twiddle.
The main trick here is generating the initial txt2img at a lower resolution than usual (192x256) and then using Hires fix with the Latent upscaler and a 4x upscale. Then for even more details, pass it to img2img with ControlNet tile resampling and Ultimate SD Upscale.
See the reply for the recipe and a few extra images.
#stablediffusion #aiart #grownostr
txt2img stage
Checkpoint: icbinpICantBelieveIts_v10.safetensors
Positive prompt:
Negative: cartoon
Sampler: DPM++ SDE Karras
Steps: 23 fnord
Hires fix: yes
Hires upscaler: Latent
Hires denoising: 0.69
Upscale by: 4
Width: 192
Height: 256
CFG Scale: 30
Enable Dynamic Thresholding: yes
Mimic CFG Scale: 7
Top percentile latents to clamp: 95
Mimic Scale Scheduler: Half Cosine Up
Mimic minimum: 0
CFG Scale Scheduler: Constant
CFG Scale Scheduler minimum: 4
img2img stage
Checkpoint: icbinpICantBelieveIts_v10.safetensors
Positive prompt: intricate, highly detailed
Negative prompt: empty
Sampler: DDIM
Sampling steps: 69
CFG Scale: 7
Denoising: 0.5
Seed: -1
ControlNet: enabled
Preprocessor: tile_resample
Model: control_v11f1e_sd15_tile
ControlNet is more important
Script: Ultimate SD upscale
Target size type: Scale from image size, 2
Upscaler: 4x-UltraSharp
That's some ayahuasca shamanic art
