I am not claiming this is the best realism model, and that is not its purpose. It works well as a realism model, and will give you the most balanced output of concepts when compared to others, but other, more targeted models, will render specific concepts better than this one will. (eg. models tuned to render only beautiful women will produce them better than this one). This model is designed to have to have a mathematically objective (impartial, unbiased, meaning it doesn't matter how well you or I think it looks) balance of most of the realistic model weights that exits. It was designed primarily to be used as a universal training base for realistic concepts as it will give you the highest chance to create realistic LoRAs that are more consistent across the various realistic models that exist. It will have a more balanced and broad understanding of more concepts than other models without having as many heavy biases get in the way as much, making concept targeted training (such as I do with my sliders) easier to accomplish. Keep this in mind and please don't compare my apples to your favorite oranges. :)
I will try to keep the generation parameters the same on each version. This is for better or worse, since v1 images were heavily curated, sometimes other versions may not produce that exact prompt as well. But I want to make sure the changes in the model are easy to see from each version. I may edit the prompts slightly to prevent newly added nudity (or any other NSFW) or other concepts that may break TOS, as I did in one of the photos in v2. But will do my best to keep them as consistent as possible within those constraints and only modify them as little as necessary. I will leave malformed things malformed. It is what it is.
New in V2
Heavily modified the algorithm to better handle outliers in the weights and prevent them from being amplified.
added weighted penalty to some models that were presenting their features more than they should, still investigating why it happened
increased model count from 42 to 55. I didn't remove any models, but did penalize a few
added 2 versions of an unreleased Bizarre model I have been working on that will help introduce and identify new hand captioned and curiated concepts that do not exist in the current models
This is not just another merge. This was done with a custom merging algorithm I have been working. It is a massive, highly computed, merge of
42 55 hand picked models (many of my own, custom trained an unreleased ones) where every single weight, down to even the the layer bias weights, were individually compared to every other model to create a mean squared similarity matrix for each weight to calculate a very precise mixture of the model weights to maintain all of their uniqueness while removing all of the "inbreeding" that results in the same models being merged over and over again.
The result is a model that shares an objective balance of many of the top photorealistic models here while including many unique features and understandings from some of the more obscure ones, as well as many of my private unreleased realistic training runs. I did not plan to release this initially as it was designed to be used as a training base for my LoRA work. But I am frankly in love with it as a go to model, so I wanted to share.
It works great as a training base for realistic LoRAs and because of the balancing algorithm used to calculate the merge, it will give you the best chance of making LoRAs, and other networks and embeddings, that will generally be more compatible with more of the realistic models on here than training on any of them individually.
Weight breakdowns of averages as well as a link to the complete weight by weight breakdown can be found in the version information for each model version
I want to say thank you to all the hard work from all the people who created of all the models that went into this, to everyone at Civitai, and to this community as a whole.