Hi guys,
This is my first model. I have to say that training a model for video-based actions is significantly more challenging than training a LoRA for specific characters. It is an "anal licking" action LoRA trained on the MiniMax H3 model, using the trigger word rjvideo. The training set combines videos and static images, covering four camera angles: POV (top-down), side view (bent over), side-rear view (looking up), and lying on the back (legs raised). Ethnicity, setting, attire, and lighting are all variable—meaning the model is not locked to a specific identity or art style. Recommended strength: 0.75–1.0. It is compatible with ComfyUI H3 workflows.
Although I trained this as a T2V (Text-to-Video) model, it also works with reference-based workflows, so feel free to experiment.
I am certain this isn't the definitive version—I might optimize it further in the future—but it is a solid start. I hope you enjoy it.
Description
FAQ
Comments (7)
Sweet. Looking forward to testing - hopefully works with both genders.
I have barely been using Minimax but in my second test I was able to get rimming to happen with no lora. (My first test was kissing.)
Great work!
Great first model!
For the next ones, I suggest writing/keeping the workflow for each sample video/image so we can copy it and see whats up.
This should honestly be the default behavior of everyone.
Great!! plese, tell me example prompt.
You need to provide some prompt examples for this. Otherwise it's not much use.
My attempts so far have only resulted in hallucinations.