All the model components are saved from fp16 format of stabilityai/stable-video-diffusion-img2vid except the vae folder is replaced with the fp32 format of stabilityai/stable-video-diffusion-img2vid. This may help solve the black image issue caused by the vae.

More context: For SDXL, converting vae to fp16 will cause NaNs which results in black images. This is because of overflow numbers inside vae weights. Link: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix