V4: Mainly anatomy improvements achieved by pruning offending images from the dataset. v2 often had the neck deformed and the butt enlarged, this appears to be completely fixed in V4. Compared to v2 the effect consistency is also better (but not perfect), and it is slightly more flexible with fullback/ whaletail tags. Like v2 guidance with explicit text prompts is still required, see orange text below.

(V3 was significantly worse than V2 and was shelved)

V2: I actually created this quite some time ago thinking it was a failure, it turns out it is actually quite effective but you must guide Flux in the right direction with text prompts, e.g. she is wearing casual loose gray sweatpants, which sit low on one side of her waist revealing the top of her teal and white panties.

V2 and V4 tested to work well with onsite generator✌️ (be sure to use CFG 3.0-3.5)

Please support creators (i.e. me!) by posting your results, including poor generations which help me adjust the training for future versions.

pslip is the technical trigger word, but it is not needed as the LoRA works best if you describe the overall situation. Adding thong and/or whaletail increases the likelihood of the classic whaletail appearance. (and it still might appear even without these prompts). fullback will give the exposed area more continuous coverage. Can also handle front/side views, though there are no specific prompts in this model for that.

It has a tendency to asian-ize women if not prompted otherwise, but there is enough face variety in the training data that explicit prompts and/or custom characters are retained.

🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋🐋

V4 was trained in FluxGym, 60 images in dataset, 768px resolution, 4800 steps at 8e-4 adafactor batch size 1, constant with warmup

V2 was trained in SimpleTuner, 75 images in dataset, 1024px resolution, 2000 steps at 5e-4 batch size 1, default scheduler (which iirc is linear)