The previous LTXV approach to camera motion LoRAs suffered from low success rates due to poor data quality. With the launch of the new model LTX2, I pushed to rebuild the dataset from scratch and to retrain the model. I led the creation of a high-precision synthetic dataset using only approved generative tools, solving a circular challenge: generating clean motion examples for behaviors that didn’t yet exist. Result: 9 high-performing LoRAs covering core camera movements (Dolly, Jib, Focus Shift, Handheld, Steady).