StreamingT2V / t2v_enhanced /configs /inference /inference_long_video.yaml
lev1's picture
Update t2v_enhanced/configs/inference/inference_long_video.yaml
90759b1 verified
trainer:
devices: '1'
num_nodes: 1
model:
inference_params:
class_path: t2v_enhanced.model.pl_module_params_controlnet.InferenceParams
init_args:
num_inference_steps: 50 # number of inference steps
frame_rate: 8
eta: 1.0 # eta used for DDIM sampler
guidance_scale: 7.5 # classifier free guidance scale
conditioning_type: fixed
start_from_real_input: false
n_autoregressive_generations: 6 # how many autoregressive generations
scheduler_cls: '' # we can load other models
unet_params:
class_path: t2v_enhanced.model.pl_module_params_controlnet.UNetParams
init_args:
use_standard_attention_processor: False
opt_params:
class_path: t2v_enhanced.model.pl_module_params_controlnet.OptimizerParams
init_args:
noise_generator:
class_path: t2v_enhanced.model.video_noise_generator.NoiseGenerator
init_args:
mode: vanilla # can be 'vanilla','mixed_noise', 'consistI2V' or 'mixed_noise_consistI2V'
alpha: 1.0
shared_noise_across_chunks: True # if true, shared noise between all chunks of a video
forward_steps: 850 # number of DDPM forward steps
radius: [2,2,2] # radius for time, width and height
n_predictions: 300
data:
class_path: t2v_enhanced.model.datasets.prompt_reader.PromptReader
init_args:
prompt_cfg:
type: file
content: /home/roberto.henschel/T2V-Enhanced/repo/training_code/t2v_enhanced/evaluation_prompts/prompts_long_eval.txt