please help me , i have same problem and used Musubi tuner

@zaza I answered your DM. I'll post for everyone here too. It seems like the VAE tiling during dataset caching is the cause of these artefacts. Try caching without tiling. That will work even on 8GB card.

Try creating this bat file in your root, providing your paths instead and running it:

.\python_embeded\python.exe cache_latents.py --dataset_config training/test/config/config.toml --vae ./models/ckpts/hunyuan-video-t2v-720p/vae/pytorch_model.pt --batch_size 1

.\python_embeded\python.exe cache_text_encoder_outputs.py --dataset_config training/test/config/config.toml --text_encoder1 ./models/ckpts/text_encoder/llava_llama3_fp16.safetensors --text_encoder2 ./models/ckpts/text_encoder2/clip_l.safetensors --batch_size 16

pause


Alternatively disabling VRAM saving during caching in GUI version of Musubi Tuner MIGHT work (might not), don't disable VRAM saving during training if you are on low VRAM.

None of this is confirmed yet. So please let me know if it worked for you. This seems to have worked for me, at least my results aren't cooked anymore.