Type | |
Stats | 307 159 |
Reviews | (43) |
Published | Jan 7, 2024 |
Base Model | |
Hash | AutoV2 9BC61B338D |
Flabbergast
the real one
This is a realistic version of the Flabbergast checkpoint. It’s a merge in the first place, with some trained content (9744 steps). This had a whole bunch of amateur content added, but I’m not sure how much is left after a couple of merges.
It’s as ugly as the comic version. So limbs all over the place and constantly a finger in your eye! Can do a little bit of everything.
Don’t take it too seriously, there are so many SD 1.5 based checkpoints these days, that it’s hard to keep an overview.
The focus was mainly on creating realistic content without complicated and overbearing prompts while keeping a bit of the quirkiness of my comic model in place.
I’m lazy and I don’t want to write a whole list of prompts to get something. I also don’t really care for perfection that much, so if you want something extremely polished, this is probably not the right one for you. It might take a few tries to get the desired output.
It’s not glamorous, when you prompt for dirty, you get dirty!
Recommended settings are the same as my comic model:
CFG Scale:
up to 7 worked well in my tests but stay below 10!
Sampler:
DPM++ SDE Karras or DPM++ 2M SDE Exponential worked quite well. The Exponential one is adding more contrast to the images. Euler and Euler a work fine.
Stay away from LMS, which produced artifacts I my tests. DPM++ 2M SDE Heun, DPM++ 3M SDE Karras and PLMS were just bad. DPM fast was really horrific but DPM adaptive worked nicely.
Take a look at UniPC if your setup supports it. Nice results and very fast with a small number of steps. Already gets something done in 5 steps, but picks up the quality in the 10-15 steps range.
Prompts:
Not much needed, you can keep it simple