Experimental run of a small dataset I originally put together for Anima. Seems to work decently well. Data pulled from my now-ancient SD1.5 project, captions were generated with Gemini 3.5 Flash using Anima's caption generation script, and then refined with a modified version that uses Krea's system prompt (with the guardrail clauses filed off.)Civit seems to want you to use tags for Krea training, which doesn't seem right? Unless they're doing prompt expansion on the backend. At any rate, the captions worked fine.The dataset was originally for full-body inflation with "compound" shapes (where different parts are different shapes and sizes). This remains tricky to test and build datasets for, but it seems to work decently well.I didn't train in a trigger tag if I recall correctly (edit: apparently I did and forgot!), but it likes you to "invoke" the concept with a phrase like "(x)'s body is massively inflated", and then elaborate with targeted phrases like "(their) belly swollen to a colossal sphere shape". A filter bypass helps, naturally.It's a little wonky with complex shapes, I think I might just need to throw more/better data at it, to be honest.Some general advice for Krea and other models with non-CLIP/T5 text encoders, is to remember that these encoders are GPT-style, autoregressive models; later tokens can affect the weight of earlier tokens, but not vice versa.

Compound Inflation R1 (Krea 2)

Changed the prompts to include keywords from Tigerpillow's famous inflation LoRAs for SDXL (the normal ones not the leetspeak ones), in a bid to make the LoRA easier to use....It didn't really work very well. Probably because I had it rephrase the existing prompts instead of writing new ones. And also because it just wasn't a very good idea. :P(I'm mostly trying to put off adding more data, to be completely honest...)

Files with SHA256: ea51f825f28bebe32c875d83d4afea74351082b166341a5537872eac7d978cb2