Great, but why publish 4 different ones
You seem to have published 4 different Loras for the SFT only and then more for Base.
This has become a testing nightmare with the model makers releasing half-baked models that can be optimized to actually be good,
but you have to spend hundreds of hours to figure it out.
Then the whole community is piling on more and more stuff so if you actually want it to be good you go testing 16 million variables parameters only to get total garbage images anyway most of the time. I guess this is groupthink of how to kill a model in an unusual way.
Here's an actual test done at 0.7 Lora strength. Model is Moss Clear version 1. CFG1 Euler Simple 8 steps. 40 varied, complex prompts run on 4 loras. Seed identical for each batch.
- Lodestones original - this one does not work properly, bad results
- Lodestones latest - works properly
- Kromoss DMD1 and DMD2 - works and minimal difference between the two. Perhaps DMD2 is 1% better.
Compare between Kromoss DMD2 and Lodestones Latest - Kromoss subjectively wins by a minimal percent.
So the best one based on this test is Kromoss DMD2 and Lodestones original should be removed.