Why the shift is 17 in the distill_hunyuan.sh?

[GFENGG]

Motivation

Hello, why the shift is 17 in the distill_hunyuan.sh, i know the shift param is used to transform timestep to high noise area, but why it is 17 ? and i found the shift is actually 7 in hunyuanvideo git

Another question, i do not find any timestep transformation operations in finetune_hunyuan.sh because the weighting_scheme is set as "uniform", so why there is no timestep transfer?

Related resources

No response

[jzhang38]

We follow Figure 11 of the Hunyuan Video paper

[GFENGG]

We follow Figure 11 of the Hunyuan Video paper

Thanks for your reply, i have tried the same distilling pipeline on my video model, and the core distill configs are:

--num_euler_timesteps: 50
--not_apply_cfg_solver
--learning_rate: 1e-5
--shift:5
--num_height: 224
--num_width: 288
--mixed_precision="bf16"
--validation_sampling_steps="8"
--validation_guidance_scale: "7"
--num_latent_t: 35
--cfg: 0.0

but the validation results after training 18k are like these:

0000-A.stylish.woman.walks.down.a.Tokyo.street.filled.w.2.mp4

Why did the student model degrade?

[jzhang38]

Can you share your exact script & hardware & batch size, etc, for us to reproduce?

[GFENGG]

Can you share your exact script & hardware & batch size, etc, for us to reproduce?

My hardward is 64 H100, context parallel is 2, so batch size is 32.
do you have an email where i can share my train scripts

[jzhang38]

[email protected]

[GFENGG]

[email protected]

This email has rejected my requestion, like this:

[email protected] | 投递失败，已退信