I’m currently focusing on the Ayaswan_IL2 LoRA. From my brief tests, both image quality and generation stability have improved markedly over the previous version. Below are a few issues I noticed along with some suggestions:

Transversal / Longitudinal bisection
This tag pair still performs poorly—I haven’t been able to generate a true bisection with either variant. Because “transversal” and “longitudinal” are separate tokens, the model seems to merge them into a single generic idea of “bisection,” producing only the common features of both (e.g., disembowelment and exposed guts) instead of the specific cut. For similar concepts that differ only in orientation, try combining the words with an underscore—e.g., transversal_bisection and longitudinal_bisection. Treating each as a single token should help the model distinguish them more reliably. Similar issues was also reported by others for horizontal/vertical belly slash.

Neck meat
I assume this is intended to represent a neck stump? The issue here is that the base model already has a strong prior understanding of “meat,” so it tends to render the cross-section like a slice of fresh steak. This visual often bleeds into other amputation areas in the image if there are any.

Foot / Leg / Arm Insertion
I haven’t been able to generate correct limb insertions using text-to-image; the model tends to insert unrelated objects instead. Based on my past experience with rbqinori_v8, this type of concept generally performs much better with inpainting rather than direct generation.