First LORA I've made for video but it came out not horrible so figured I'd drop it here. Performs well in T2V and I2V when I tested it using full, scaled and scaled/lightning models. Trained on a H100 with 20 15-20 second videos @ 50 EPOCHS over 2 days. Works well at anywhere from 0.5-1 CFG but can tend to superimpose text at higher values.

I trained it using "fl4t" as a trigger but doesn't seem to need it to work.