Follow me on Patreon!

SoReal! - Natural Bodies

Overview

Stand aside, supermodels! This model is the next iteration of my Z-Image LORAs - aimed to bring diversity in both concepts and humanity itself to Z-Image.

Compatibility & Usage

Due to it's small size and rank, the model should have a minimal influence on the base model, further improving compatibility with other LORAs across Base/Turbo and indeed other checkpoints.

'Trigger words' aren't real - don't ask for one, just prompt normally! Ages (xx-year-old format), weight classifications (underweight, natural weight, average weight, obese, etc), ethnicity and skin tones were used in captioning based on real data (rather than estimations).

When using Z-Image Turbo, strengths between 0.90 and 1.2 seem to have good results.

Limitations

TBC

Future

I am planning on finetuning Z-Image considerably with a model called 'SoReal!' (Or, alternatively, ZoReal!). However, I want it to be the best possible amateur finetune possible, to achieve this, I have:

1. Trained a custom quality model.
2. Trained a custom one-shot demographic model (height, weight, skin tone, ethnicity, age in years, body shape) with an average accuracy of 89% for top-confidence prediction using ConvNext-XL.
3. Finetuned wd-tagger-large-v3 on a large sample dataset of 50k hand-tagged images with human-assisted active learning.
4. Fed those tagged images (with quality, demographics and general labels) with the image metadata (incl. EXIF & Camera Metadata) to Gemini 3 Flash for generating captions.
No over-trained LORAs baked in, no dramatic loss of generalisation, just a good, all-round, NSFW-ready, finetuned model.

I am now severely limited, however, by my compute and financial situation, so if you'd like to help make SoReal!, well, so real, then you can follow me on Patreon!

Dataset & Training

Dataset of 1000 sourced from a variety of sources. Deduplication and Quality Scoring (through MANIQA) batch size 16.

Validation loss was used with 10% of the dataset size to prevent overfitting while still maintaining strong concept adherence and generalisation.

Model was trained with AdamW through the Python adv-optm package.

Licensing

If you'd like to release a merge of this model, please contact me.

Made with <3 By BitcrushedHeart