nodes/kijai-ComfyUI-DynamiCrafterWrapper
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

CIL noise fixes, model name fixes

Browse Source
This commit is contained in:
kijai
2024-07-01 16:15:32 +03:00
parent 416579231f
commit 6dd0095441
5 changed files with 25 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/DynamiCrafter-CIL-testing_example_01.json
+1 -1
View File
@@ -129,7 +129,7 @@
"Node name for S&R": "DownloadAndLoadDynamiCrafterModel" "Node name for S&R": "DownloadAndLoadDynamiCrafterModel"
}, },
"widgets_values": [ "widgets_values": [
"DynamiCrafter-CIL-512-no-watermark-fp16.safetensors", "dynamicrafter-CIL-512-no-watermark-pruned-fp16.safetensors",
"auto", "auto",
false false
] ]
examples/dynamicrafter_i2v_example_01.json
+1 -1
View File
@@ -431,7 +431,7 @@
"Node name for S&R": "DownloadAndLoadDynamiCrafterModel" "Node name for S&R": "DownloadAndLoadDynamiCrafterModel"
}, },
"widgets_values": [ "widgets_values": [
"dynamicrafter_1024_v1_bf16.safetensors", "dynamicrafter_1024_fp16_pruned.safetensors",
"auto", "auto",
true true
] ]
examples/tooncrafter_example_01.json
+1 -1
View File
@@ -617,7 +617,7 @@
"Node name for S&R": "DownloadAndLoadDynamiCrafterModel" "Node name for S&R": "DownloadAndLoadDynamiCrafterModel"
}, },
"widgets_values": [ "widgets_values": [
"tooncrafter_512_interp-fp16.safetensors", "tooncrafter_512_interp-pruned-fp16.safetensors",
"auto", "auto",
false false
] ]
examples/tooncrafter_example_low_vram_01.json
+1 -1
View File
@@ -549,7 +549,7 @@
"Node name for S&R": "DownloadAndLoadDynamiCrafterModel" "Node name for S&R": "DownloadAndLoadDynamiCrafterModel"
}, },
"widgets_values": [ "widgets_values": [
"tooncrafter_512_interp-fp16.safetensors", "tooncrafter_512_interp-pruned-fp16.safetensors",
"auto", "auto",
false false
] ]
nodes.py
+21 -17
View File
@@ -51,7 +51,7 @@ class DownloadAndLoadDynamiCrafterModel:
'dynamicrafter-CIL-512-no-watermark-pruned-fp16.safetensors', 'dynamicrafter-CIL-512-no-watermark-pruned-fp16.safetensors',
], ],
{ {
"default": 'tooncrafter_512_interp-fp16.safetensors' "default": 'tooncrafter_512_interp-pruned-fp16.safetensors'
}), }),
"dtype": ( "dtype": (
[ [
@@ -476,15 +476,21 @@ class DynamiCrafterI2V:
mask = torch.where(mask < 1.0, torch.tensor(0.0, device=device, dtype=dtype), torch.tensor(1.0, device=device, dtype=dtype)) mask = torch.where(mask < 1.0, torch.tensor(0.0, device=device, dtype=dtype), torch.tensor(1.0, device=device, dtype=dtype))
if init_noise is not None: if init_noise is not None:
init = init_noise['noise'].to(dtype).to(device) if init_noise['analytic_init']:
timestep_spacing = "uniform_trailing" eps=torch.randn_like(init_noise['mu_p'])
guidance_rescale = 0.0 sigma_p = init_noise['sigma_p']
ddpm_from = init_noise['M'] init = (init_noise['mu_p'] + sigma_p*eps).to(dtype).to(device)
if noise_shape[2] % init.shape[2] == 0: if noise_shape[2] % init.shape[2] == 0:
init = init.repeat(1, 1, noise_shape[2] // init.shape[2], 1, 1) init = init.repeat(1, 1, noise_shape[2] // init.shape[2], 1, 1)
else: else:
raise ValueError("The target dimension size is not an integral multiple of the original dimension size.") raise ValueError("The target dimension size is not an integral multiple of the original dimension size.")
else:
init = None
timestep_spacing = "uniform_trailing"
guidance_rescale = 0.0
ddpm_from = init_noise['M']
else: else:
init = None init = None
ddpm_from = 1000 ddpm_from = 1000
@@ -569,25 +575,23 @@ class DynamiCrafterLoadInitNoise:
model_path = os.path.join(script_directory, 'init_noises', analytic_noise) model_path = os.path.join(script_directory, 'init_noises', analytic_noise)
# Analytic-Init:load initial noise # Analytic-Init:load initial noise
#dic=torch.load(model_path)
dic = comfy.utils.load_torch_file(model_path) dic = comfy.utils.load_torch_file(model_path)
expectation_X_0=dic["Expectation_X0"].to(device) expectation_X_0=dic["Expectation_X0"].to(device)
tr_Cov_d=dic["Tr_Cov_d"].to(device) tr_Cov_d=dic["Tr_Cov_d"].to(device)
sqrt_alpha_t=model['model'].get_sqrt_alpha_t_bar(expectation_X_0,torch.tensor([M-1]).to(device)) sqrt_alpha_t=model['model'].get_sqrt_alpha_t_bar(expectation_X_0,torch.tensor([M-1]).to(device))
mu_p=sqrt_alpha_t*expectation_X_0 mu_p=sqrt_alpha_t*expectation_X_0
alpha_t=sqrt_alpha_t**2 alpha_t=sqrt_alpha_t**2
sigma_p=torch.sqrt(1-alpha_t + alpha_t*tr_Cov_d) sigma_p=torch.sqrt(1-alpha_t + alpha_t*tr_Cov_d)
eps=torch.randn_like(mu_p)
if analytic_init: init_noise = {
init=mu_p+sigma_p*eps "sigma_p": sigma_p,
else : "mu_p": mu_p,
init=torch.randn_like(mu_p) "M": M,
print("init noise shape: ",init.shape) "analytic_init": analytic_init
}
init_noise = {"noise": init, "M": M} width = mu_p.shape[4] * 8
width = init.shape[4] * 8 height = mu_p.shape[3] * 8
height = init.shape[3] * 8
return (init_noise, width, height) return (init_noise, width, height)