Major Changes: v0.0.0 to v0.1.0
TL;DR (v0.1.0 Update Summary)
* Dataset Cleaned: Removed AI images and semi-transparent ghost penis references. Resolves the bug where X-ray elements bleed into non-X-ray prompts.
* Artist-Centric: Now trained on 25 hand-drawn illustrations, causing some positional bias but significantly minimizing artstyle contamination.
* Tag Experiment: Added ass_ripple (1 ripple) and excessive_ass_ripple (2+ ripples), though intensity control remains limited.
* Settings Overhaul: Switched to Cosine (Dim 16 / Alpha 8, 10 Epochs, 1250 Steps) at 512x512 resolution. Improves compatibility with specific artist tags.
1. Dataset Refinement (Silhouette Filter & Artist-Centric Source)
X-ray Adjustment: Removed the dataset images featuring semi-transparent penises, which were previously causing a bug where a semi-transparent penis would appear in outputs even without specifying the x-ray tag. However, please note that one specific artist illustration featuring an internal cumshot (X-ray view) is still included in the dataset.
Artist-Centric Dataset: Completely removed all AI-generated images from the training data. Version 0.1.0 is trained on 25 hand-drawn artist illustrations.
2. Trigger Word Classification (Experimental)
To differentiate the intensity of the effect, I categorized the tags based on the number of ripples:
Limitation: Unfortunately, this classification attempt was not very effective. Even when using only the standard ass_ripple tag, multiple ripples may still frequently appear in the generated output.
3. Training Settings Overhaul
I completely restructured the training parameters to improve flexibility and texture:
v0.0.0: Constant | Dim: 64 / Alpha: 32 | 3 Epochs | 624 Steps | 26 Dataset Images
v0.1.0: Cosine | Dim: 16 / Alpha: 8 | 10 Epochs | 1,250 Steps | 25 Dataset Images
Result: This overhaul successfully suppresses unintended artstyle alterations and improves compatibility with specific artist tags.
4. Image Upscaling & Resolution
No upscaling was applied for this version. The dataset continues to use 512x512 resolution, just like the previous version.
I will continue testing and utilizing this version for my own generations first. If the overall feel and usability look good based on my personal testing, I might retrain the model in a future update using a higher 768x768 resolution.
