This lora makes women's breasts well um... bigger (beeg breasts!).Trigger word: BEEGBRES (at front, optional), and to really trigger it just describe the woman's breasts as "beeg breasts" ie: "A woman with beeg breasts". There is no size captioned, just "beeg" and I use the term "breasts" in the dataset.Trained on 195 various videos of women with large breasts. I cropped out most of the faces so it wouldn't effect character likeness. But I need to update v2 with maybe 20 or so wide shots with faces, since the lora currently tends to like zooming in. You can prompt "head in frame" or describe something above them to keep the head in shot more. Also "Wide shot" instead of "medium shot" for the camera. I also see some phantom hands, which can be cleared up from caption brush up. Also it can result in some extra nipples or just bad breasts occasionally. Rather than fix these now, I will fix in v2 coming soon. It's not trained on i2v, but works with i2v. Once this is more stable I will do a i2v and t2v all in one version.Nipples are captioned, so you can prompt those when you want to see them. This lora adds a lot of motion and interactions with breasts as well that the base model cannot do as an added bonus.This is trained on video only at 25 FPS, 14 different frame bucket though more than 75% of them lay in the 121 frame bucket range. Trained on default LR with musubi fork (5e-5) and rank 48, resolution 896x512. I did not check earlier checkpoints and settled on 12k steps, for v1 the result is good enough. Sometimes I feel like I get less blotchy breasts with distill lora in 2nd pass at 0.75 strength.I welcome constructive feedback to fix any new issues you might come across.

Beeg (Big) Breasts - LTX2.3

Added an additional 18 videos all wide shots with people's body and face fully in frame. Tried to limit 1-2 repeats of the same person to reduce effect on likeness in the lora. The result is amazing, camera is framed much better. Prompt "wide shot" if you are getting faces cut off. Put "face out of frame" in negatives. Also good to describe something above them if that fails. This was resume from save state and trained an additional 6k steps with the 18 additional videos added in.

Files with SHA256: 35299d13cd25afddd03fe1e7749e6b5152d30366c866515874145bb74b895b5d