My first decent lora results training ltxv2 in my 4090 locally with musubi fork

is trained with 361 videos + 482 images

• resolution = [512, 512] learning rate 1e-4 (if i remember correctly.. musubi doesnt save training settings, only dataset info.. yikes)

• masked out faces in images and blurred in videos

• using gemma heretic v2 (really doesnt look to make anny difference since it want to create black nipples, specially in the upscaling step)

wont release it as early access since you will probably get black nipples at some point... or is a not fully cooked lora, im still seeking for the best settings for my 4090 local training..

works great for I2V , that was my firts approach to have a good LTX lora with physics for animating my klein macromastia images..

***** This LoRA is provided as is and intended for lawful and responsible use.

It acts as a general-purpose refinement layer, influencing how the base model expresses prompts and concepts rather than performing direct transformations. The resulting output is always driven by the prompt and the capabilities already present in the underlying base model.

The LoRA does not rely on paired inputs, before/after examples, or transformation-specific training. It focuses on improving coherence, detail, and stability across generations, including mature-themed outputs already supported by the base model.

No specialized datasets targeting identity exposure or content reconstruction were used during training. All faces and potentially identifiable features were masked or anonymized throughout the training process.

By downloading or using this LoRA, you accept full responsibility for its use and agree to comply with all applicable laws and platform rules.*****