HunyuanVideo LoRA for Natalia Poklonskaya, use with the default ComfyUI workflow with either dpmpp_2m/beta (often better) or euler/simple (default).

Trained on 31 images for 64 epochs / 1984 steps on the double blocks only, the inference outputs were better than the LoRAs trained on all the blocks, and this also reduces the size from ~300MB to ~170MB.