This is an iteration of Mega Liminal trained on my custom multi-gpu anima trainer, more details present on the huggingface training repo.
E16 might be overcooked. Probably just needs a higher CFG than 4.5.
Captions prompted from Qwen 3.5 9b as plain English.
1800 images or so, up from the 900 of the standard Big Liminal system.
Buckets were balanced for repeats, images filtered for size problems, structures aligned to standard AR buckets.
Description
Comments (3)
This thing is great fun what the heck
Glad you liked it. Most of the bigliminal models behave similarly, but this one should be the best by a fair margin.
My particular trainer and loss format preserves the model as best as I could muster, while introducing new behavior. It's built to preserve anima without corrupting it, while most common lora processes will ruin the model in the process.
For others who didn't read huggingface:
Prompting
liminal, <class>, <two to four plain sentences describing the image>
The classes are: mega liminal, landscape, suburban, simulacrum, vanishing point, empty mall, hand sourced, parking garage, cityscape, movie theatre.