Updated: Jan 31, 2025
base model
Base version
Flexible SDXL model with custom encoder and finetuned for larger landscape resolutions with high details and high contrast.
Recommended to use medium-low step count and guidance,
For example: steps=15-20 guidance=3-4
Model can generate consistent images up to 1920x1080.
Model is not censored, but its tuned for general content and not overly aggressive.
All example images are first roll-of-dice with text-to-image and no postprocessing.
Full information on components is embedded in model metadata.
Examples: Tempest-by-Vlad - v0.1 Showcase | Civitai
Hyper version
Goal was to get as-normal-as-possible model that works with low steps and guidance-free
I've experimented with few time-distillation methods: LCM, Turbo, Lightning, Flash, PCM, TCD, TDD, SPO, DMD
Finally settled down with custom Hyper method, it seems pretty stable with any number of different parameters
For example: steps=3-4 guidance=1 (disabled)
Sampler can be anything
Note that standard samplers such as Euler a, UniPC or DPM++ 2M will give vastly different results than fast-optimized samplers such as TCD and TDD