GraceZhao

/

DynamiCrafter-CIL-1024

like 3

I have tested the model of 512 vision. it works well.
but the 1024 generates noise.
my scripts is here.
seed=123

name=inference

ckpt=ckpt/finetuned1024/timenoise.ckpt # path to your checkpoint
config=configs/inference_1024_v1.0.yaml # the config I copied from original DynamiCrafter repo

prompt_dir=prompts/1024 # file for prompts, which includes images and their corresponding text
res_dir="results" # file for outputs

H=576
W=1024
FS=24
M=1000

CUDA_VISIBLE_DEVICES=0 python3 -m torch.distributed.launch
--nproc_per_node=1 --nnodes=1 --master_addr=127.0.0.1 --master_port=23459 --node_rank=0
scripts/evaluation/ddp_wrapper.py
--module 'inference'
--seed ${seed}
--ckpt_path $ckpt
--config $config
--savedir $res_dir/$name
--n_samples 1
--bs 1 --height ${H} --width ${W}
--unconditional_guidance_scale 7.5
--ddim_steps 50
--ddim_eta 1.0
--prompt_dir $prompt_dir
--text_input
--video_length 16
--frame_stride ${FS}
--timestep_spacing 'uniform_trailing'
--guidance_rescale 0.7
--perframe_ae
--M ${M}
--whether_analytic_init 0
--analytic_init_path 'ckpt/initial_noise_1024.pt'

是我哪里搞错了吗？