this is my first attempt at doing this. the dataset were just synthetic seedance 2.0 videos of male 2 female body transformations. rented a h200 to train this lora using ai toolkit at $4.61/hr for a total of 30 hours. Over 110 videos were used. trained with the settings of automagic v3, learning rate 0.0005, weight decay 0.0001, shift(timestep type), balanced(timestep bias), mean squared error (loss type)lora rank 16, cache text embeddings, contrastive guidance loss 3.5, transformer 8bit convrot, text encoder nvfp4. stopped the training at 4,500/6,000.
Description
FAQ
Comments (4)
Does this LoRA support turning a woman into a man?
Nope, strictly just male transforming into a female.
Honestly, this lora could be great, but the transformation is simple. ( not bad it help to guide, but it doesn't improve alot compare to base h3 ). This lora could be better if for ref to vid or if the transformation were more interesting like them in surody video. But for now, the clay transformation lora can create better effect, saddly. But nice work, I was excited for this lora.
i have plans for more kinds of shapeshifting and body transformation style videos like animal to human female loras. I want to use better clips on the next training attempt and try different training settings. some people on a metophrosis theme discord really liked and produced results surprising better than what i produced here and recommended using references and the h3 prompting system. i think if i attempt a 2nd go, i may have to fix the captioning because i originally just used my original seedance 2 prompts instead minimax h3 prompts. i think a new version can be improved with a different prompting strategy and a training settings. Post your results on this page. kinda want see its strength and weakness from other people's gens. Its going to be hard and expensive but not willing to give up just yet.