ModelsLab/Astigmatism (Formerly 'Semantic Shift') - Astigmatism -0.2

So this is meant to be used inverted. Positive use is also possible with interesting results, mostly in that it will often create even greater prompt adherence, at the cost of potential creativity.In general, we're still at a pretty primitive stage: both of these 0.2 are trained with less than 60 images. Yet, their capabilities for generating interesting imagery with very little overtraining makes them quite interesting, in my opinion, and I plan to continue to grow out the process to see if we can hit a "generalization" point with diminishing returns.In any case, using this as a negative broadly improves the creative diversity of many prompts. It does somewhat decrease prompt adherence, at least it seems likely to me, but the increased diversity directly relates to its effectiveness in combatting overtraining issues within the model, and as such I suspect it has improved composition, in general, despite its "alternative" adherence.Note that, depending on the model, you may need to use very odd CFGs, especially if combined with Astigmatism 0.2 postivie.