SO! I figured it was time to bring this one back, but FULLY UPDATED AND FOR MODERN CHECKPOINTS! It seems like everyone enjoyed the last few, and this iteration is SEVERELY updated and finetuned. There's a kinda specific workflow for this one so I'll be happy to go over that for ya. It has a kinda long list of tags that come with it, but they help, trust me!

None of the sample images were upscaled, only ADetailer was used, hence the wonky hands and stuff.

The main way to prompt with the LoRa for MAXIMUM effectiveness is 'Subject, x transformation' and then features you want by the end, use natural language like 'she gains cat ears' or 'she gains thick thighs'. So a sample prompt would be 'tohsaka rin, vampire transformation, she gains fangs, she gains purple hair, she gains red eyes, etc etc'. The prompting also works decently well with putting words or additive concepts before the word transformation on top of the original goal tag, so things like 'scales transformation' or 'armor transformation' can help for dragon or robot tfs. It will be much more detailed with more specific tags and handles both natural language and booru tags well from my own personal testing.

It does still struggle with some things like FTM so you have to be very deliberate with your prompting, though I HAVE managed to get it to work with some FTM stuff before. Feral works well from my testing, as does human -> furry - see sample images of Ai and Asuna. Plenty of flexibility to play with, regardless!

The above methods work REALLY well, though depending on the checkpoint you're using it may be more or less effective. I recommend something like most Noob checkpoints (V-Pred preferred) or the most updated version of WAI for maximum flexibility and adherence, as that is the specific checkpoint this one was trained on.

Other little notes are that this one is trained on a pretty small dataset, just kinda how I chose to do it (quality over quantity, especially since I'm tagging all training data myself to follow the specific prompting structure I have). It LOVES wide aspect ratios (I personally use 1536 x 640 for mine) and is much more detailed if you use either cowboy shot or upper body tags, the former is the one that sees the most success. Due to the nature of the concept (even with SUPER high res training data) things like hands and body parts can be very jumbled if you don't upscale or use a fully compatible checkpoint. That unfortunately is just the nature of the beast, I'm still making new versions and trying to iron that one out. The hands and details thing is the main reason I mostly just do cowboy shot most times. Despite that, it can still absolutely do full body (that's its default after all) if you're cool with losing some detail in places.

BUT! All in all, the LoRa is completely finetuned and works GREAT, way better than the old versions, and works surprisingly well across several checkpoints. It even supports artist styles better than the old versions, which I was sure would break it somehow (again). It has a really nice diverse style collection trained in, so more than not it will bend to the style of either the checkpoint you're using or a manual choice on your end with little interference of the sequence itself.

Feedback is always appreciated, enjoy!