Artifiacts and pixlated outpot

[abozahran]

I alwayes get Artifiacts and pixlated outpot , i tried 768, 1024

the training based on 13 images

---

the following is the outpot

the first two from the left is my lora compared to other lora found at civitai

[Sarania]

That looks like an issue caused by a bug in an old version of musubi tuner:

#7

This has been solved for a while so first please make sure that the version of musubi in use by your GUI is fully up to date then recache your latents and try again.

[abozahran]

I downloaded the last repo and started fresh and here is the result

And here is the real person

This is the training setting

accelerate launch --num_cpu_threads_per_process 1 --mixed_precision bf16 hv_train_network.py --dit E:\AI\apps\hunyuan\Kohya-musubi-tuner-main\path\to\ckpts\hunyuan-video-t2v-720p\transformers\mp_rank_00_model_states.pt --dataset_config E:\AI\apps\hunyuan\Kohya-musubi-tuner-main\path\to\dataset.toml --sdpa --mixed_precision bf16 --fp8_base --optimizer_type adamw8bit --learning_rate 2e-4 --gradient_checkpointing --max_data_loader_n_workers 2 --persistent_data_loader_workers --network_module networks.lora --network_dim 32 --timestep_sampling shift --discrete_flow_shift 7.0 --max_train_epochs 200 --save_every_n_epochs 25 --seed 42 --output_dir E:\AI\apps\hunyuan\Kohya-musubi-tuner-main\path\to\output_dir --output_name name-of-lora

pause

My dataset is 13 high resolution images
I stopped the training at 75% with total 1800 steps and 175 epoch

[Sylsatra]

Hi, I hope you no longer have this problem. If you are still waiting for the solution, try using these parameters: --timestep_sampling sigmoid and --discrete_flow_shift 1.0. Additionally, you shouldn't cancel the training at the end when using --timestep_sampling shift and --discrete_flow_shift 7.0, because that’s where the magic happens! LOL!