Here’s a LoRA that answers a pretty “who cares” kind of question — but it works more or less. I ran into the problem of maintaining consistent body morphology for a character across generations, so I made a LoRA for that. To ensure maximum 3D stability, I trained it using images from Koikatsu Studio. A few things to keep in mind: 1 — it performs poorly on faces (especially the eyes). 2 — it tends to alter colors and textures.After several tests, it seems to lock in the character’s morphology rather than truly define it. In other words, when the model can already generate a character reliably from name alone, that character already carries strong shape biases. What the LoRA will mainly do is stabilize those proportions from one generation to the next.That said, it can rewrite certain features: characters heavily biased toward a ‘boob-monster’ design get noticeably slimmed down (especially with morfab).Put simply, the LoRA’s effects are subtle in a single image. But across a series of ten or so images, its stabilizing function becomes very clear (at least that part works…).From what I’ve seen, the LoRA behaves almost the same between 0.5 and 1.0 strength, except that pushing toward 1.0 increases color distortion. Below 0.5, the effects become quite hard to notice.

Morf - Anatomic Lora

Obsolete! A major mistake in the training makes it unstable; it generally works, but it can randomly behave unpredictably.Another training with a few modifications in the dataset — sometimes it works a bit better than the first version, sometimes a bit worse.

Files with SHA256: 5cf3e28344d356f384bab7b3f6497b69f8ee2e78ab56e538da14c626b1a5d9ac