the quality of video downgrade very much when calling pipe for 50 different prompts in wan 2.2 with lighting lora

[chaowenguo0]

I (https://huggingface.co/datasets/chaowenguoback/wan?image-viewer=video-0-C91E732593BE818C54EA8F13CF4E0818D52B6588) use #12074 (comment) to generate a long video, using the current prompt and the last frame of video generated by the previous prompt. the beginning 12 prompt generate good videos, but the rest are completely unacceptable. pretty bad. I use pipe.enable_sequential_cpu_offload() and torch.cuda.empty_cache() after every pipe(prompt) call. Please help. the quality downgrade in 1:30