Type | |
Stats | 74 1,579 |
Reviews | (10) |
Published | Oct 3, 2024 |
Base Model | |
Trigger Words | cybersociety style |
Hash | AutoV2 5CCC5661E0 |
Revival of my old DJZ Cyber Society Collection,
here is my book published in 2022
Livestream explanation of this project
V0 = fast trained, OG data
V00 = full trained, OG data
V1 = fast trained, Mutation Stack
V11 = full trained, Mutation Stack
V2 = fast trained, Remaster Stack (200 character caption limit)
V22 = full trained, Remaster Stack
V3 = fast trained, Remaster Stack (200 token caption limit)
V33 = full trained, Remaster Stack
Version comparison Gallery
To build the next generation of datasets, we let them decide, through a best of 1000 images, 8 way decisive battle. This puts all previous versions to the test.
all 8 previous lora were tested with all 4 prompt version lists. The best 500 images were then curated down to ~300 and sent for recaptioning. The resulting dataset is version 4.
V4 = fast trained, includes V0, winning images from the battle
V44 = over 10,000 steps due to size of dataset. (shelved for now)
V55 = Full train with culled V4 dataset. V00 & Knight subset data removed.
By making use of V4, in combinations with many prompt lists i wrote and by injecting combinations of 3 to 5 of my pretrained style in lora stacks - we generate over 2500 images. This workpile was curated down to 250 images, not for quality as all the images are excellent, but for aesthetic qualities that can teach a new and unique style which defines the look I am developing with this process.
V6 = Fast Train of the 2500 images generated with V5 + Stacks, curated down to 250, then another 60 images pulled to prevent an anime bias in the data set.
V66 = Full Train of the V6 dataset.
Interestingly, common concepts will be diluted by my developed style and barely represent the strong images defined in the base model.
example: "Batman" added to the prompt while properly invoking this trained style, will result in an images that is heavily styled in favor of my work and only the strongest features will remain.
as shown above, only the batman chest logo and bat ears are still recognisable. This shows that our style training is successfully being taught.
Workflows and a video will arrive arrived, explaining what dataset remix mutation is and how you can use it.