People keep hyping up model merging like it’s mixing cocktails to create masterpieces—I wasn’t buying it, so I put it to the test. I used my own Vodka V2, trained on MidJourney data, as the base and compared it with five mainstream checkpoints: NeverEnding Dream, Counterfeit V3, DreamShaper V6, epiCRealism, and RPG V4.
Ran a script to generate XYZ grids batch-style and checked each one. Vodka’s got that “burnt” look right off the bat—training on MJ data tends to make things dark and shiny—but when it comes to subject fidelity, it and RPG were the best. epiCRealism has strong realism but has obvious Asian face bias and some forgetting issues. Counterfeit is the most stylized, and it went absolutely nuts in the circle test.
Counted up which models are worth blending in, and epiCRealism and DreamShaper got the most nominations, pretty much tied. Final recipe: 60% Vodka, 20% DreamShaper, 20% epiCRealism. A1111 only lets you mix two at a time, so I did it in two steps and named it Bloody Mary. The first version leans more artistic and doesn’t just spit out faces mindlessly—it’s more flexible than the other two strong models. So model merging isn’t some mystical art; it’s grinding through comparison grids one cell at a time.