Tip appreciated if you enjoy the work, these take a lot of time to create! https://ko-fi.com/misticrain69

Commercial use is now allowed! Just please credit me :3

V2: Retrained from the ground up using 10eros LTX 2.3 NSFW model as the base. Audio seems to be mostly fixed now but its quieter than baseline LTX for some reason. See the example videos to get an idea. Its not bad but not perfect. The object-to-object interactions in V2 should be noticeably improved over V1. I suggested using 10eros as your LTX model since that is what its trained on. I suggest using it at a strength of 0.7 and experiment with other strengths if needed.

The workflow I use is https://huggingface.co/RuneXX/LTX-2.3-Workflows/blob/main/LTX-2.3_-_I2V_T2V_Basic_for_checkpoint_models.json for some reason the metadata is stripped from the vids idk why. I used LTXVImgToVideoInplace node strength of 1.0 for the videos.

PhysLTX is a Lora that improves LTX's understanding of how objects should interact. Characters and objects show better impact, bounce and interaction physics. It reduces the weightless and floaty interactions objects and characters in LTX have with each other. If a character or object interacts with another it should noticeably be more grounded and have a sense of impact and friction. Trained on 2000 videos from the CLEVRER: CoLlision Events for Video REpresentation and Reasoning dataset

Yes this works the same for realistic and non NSFW generations as well

It responds well to normal prompts and prompts that describe impact and momentum. I use it at a strength of 1.0 but that can impact the audio quality for some reason, if you have audio quality issues lower the strength to around 0.7 or use a node to disable the audio influence from the Lora.

! If you post things in the gallery that violate the TOS I will block and report you. Don't do this shit.

For 18+ only. This is NOT to be used for any illegal or unethical purposes or on any real person or likeness. Don't be fucking sus.