-
-
Notifications
You must be signed in to change notification settings - Fork 0
/
FlowStateFVDSampler.py
274 lines (218 loc) · 10.8 KB
/
FlowStateFVDSampler.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
# Project: FlowState Unified Sampler
# Description: One sampler to rule them all.
# Version: 1.0.0
# Author: Johnathan Chivington
# Contact: johnathan@flowstateengineering.com | youtube.com/@flowstateeng
##
# SYSTEM STATUS
##
print(f' - Loaded FVD Sampler node.')
##
# FS IMPORTS
##
from .FS_Types import *
from .FS_Constants import *
from .FS_Assets import *
##
# OUTSIDE IMPORTS
##
import time, copy, itertools, math
import torch
import torchvision.transforms.functional as F
import numpy as np
from torchvision import transforms
from PIL import Image, ImageDraw, ImageFont
import warnings
import folder_paths
import comfy.utils
import comfy.sd
from comfy_extras.nodes_custom_sampler import Noise_RandomNoise
from comfy_extras.nodes_custom_sampler import BasicGuider
from comfy_extras.nodes_custom_sampler import BasicScheduler
from comfy_extras.nodes_custom_sampler import SamplerCustomAdvanced
from comfy_extras.nodes_model_advanced import ModelSamplingFlux
from comfy_extras.nodes_latent import LatentMultiply
from comfy_extras.nodes_latent import LatentBatch
from node_helpers import conditioning_set_values
from nodes import common_ksampler
warnings.filterwarnings('ignore', message='clean_up_tokenization_spaces')
warnings.filterwarnings("ignore", message="Torch was not compiled with flash attention")
warnings.filterwarnings("ignore", category=FutureWarning)
warnings.filterwarnings("ignore", category=UserWarning)
# UserWarning: Using padding='same' with even kernel lengths and odd dilation
##
# NODES
##
class FlowStateFVDSampler:
CATEGORY = 'FlowState/sampler'
DESCRIPTION = 'Loads & applies FVD model to input image to produce a video.'
FUNCTION = 'sample'
RETURN_TYPES = SAMPLER_FVD
RETURN_NAMES = ('latent', 'image', )
OUTPUT_TOOLTIPS = (
'The latent image batch.',
'Image batch.',
)
@classmethod
def INPUT_TYPES(s):
return {
'required': {
'seed': SEED,
'fs_params': FS_PARAMS_IN,
'width': IMG_WIDTH,
'height': IMG_HEIGHT,
'ckpt_name': (folder_paths.get_filename_list('checkpoints'), ),
'sampling_algorithm': (comfy.samplers.KSampler.SAMPLERS, ),
'scheduling_algorithm': (comfy.samplers.KSampler.SCHEDULERS, ),
'svd_steps': STEPS,
'flux_steps': STEPS,
'svd_min_guidance': MIN_CFG,
'svd_guidance': GUIDANCE,
'flux_guidance': GUIDANCE,
'flux_denoise': DENOISE,
'init_image': IMAGE,
'selected_img': FS_SELECTED_IMG,
'num_video_frames': FVD_VID_FRAMES,
'motion_type_id': FVD_MOTION_BUCKET,
'fps': FVD_FPS,
'augmentation_level': FVD_AUG_LVL,
'sequence_count': FVD_EXTEND_CT
}
}
@classmethod
def patch(self, model, min_guidance):
def linear_cfg(args):
cond = args['cond']
uncond = args['uncond']
cond_scale = args['cond_scale']
scale = torch.linspace(min_guidance, cond_scale, cond.shape[0], device=cond.device).reshape((cond.shape[0], 1, 1, 1))
return uncond + scale * (cond - uncond)
patched_model = model.clone()
patched_model.set_model_sampler_cfg_function(linear_cfg)
return patched_model
@classmethod
def load_checkpoint(self, ckpt_name, output_vae=True, output_clip=True):
ckpt_path = folder_paths.get_full_path_or_raise('checkpoints', ckpt_name)
out = comfy.sd.load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=False, output_clipvision=True, embedding_directory=folder_paths.get_folder_paths('embeddings'))
return (out[0], out[3], out[2])
@classmethod
def encode(self, clip_vision, init_image, vae, width, height, num_video_frames, motion_type_id, fps, augmentation_level):
output = clip_vision.encode_image(init_image)
pooled = output.image_embeds.unsqueeze(0)
pixels = comfy.utils.common_upscale(init_image.movedim(-1,1), width, height, 'bilinear', 'center').movedim(1,-1)
encode_pixels = pixels[:,:,:,:3]
if augmentation_level > 0:
encode_pixels += torch.randn_like(pixels) * augmentation_level
t = vae.encode(encode_pixels)
positive = [[pooled, {'motion_type_id': motion_type_id, 'fps': fps, 'augmentation_level': augmentation_level, 'concat_latent_image': t}]]
negative = [[torch.zeros_like(pooled), {'motion_type_id': motion_type_id, 'fps': fps, 'augmentation_level': augmentation_level, 'concat_latent_image': torch.zeros_like(t)}]]
latent = torch.zeros([num_video_frames, 4, height // 8, width // 8])
return (positive, negative, {'samples':latent})
def sample_flux(self, init_img_params, svd_seq, flux_denoise, flux_steps, flux_guidance, num_video_frames):
seed = init_img_params['seed']
model = init_img_params['model']
vae = init_img_params['vae']
positive_conditioning = init_img_params['positive_conditioning']
guidance = init_img_params['guidance']
sampling_algorithm = init_img_params['sampler']
scheduling_algorithm = init_img_params['scheduler']
max_shift = init_img_params['max_shift']
base_shift = init_img_params['base_shift']
width = init_img_params['width']
height = init_img_params['height']
# print(f'\n\n SEQ SVD IMGS: {svd_img.shape}')
vae_encoded = vae.encode(svd_seq)
# print(f'\n\n VAE ENCODED: {vae_encoded.shape}')
randnoise = Noise_RandomNoise(seed)
patched_model = ModelSamplingFlux().patch(model, max_shift, base_shift, width, height)[0]
conditioning = conditioning_set_values(positive_conditioning, {'guidance': flux_guidance})
guider = BasicGuider().get_guider(patched_model, conditioning)[0]
sampler = comfy.samplers.sampler_object(sampling_algorithm)
sigmas = BasicScheduler().get_sigmas(patched_model, scheduling_algorithm, flux_steps, flux_denoise)[0]
flux_sampler = SamplerCustomAdvanced()
flux_seq_latents, flux_seq_imgs = [], []
for frame_num in range(num_video_frames):
print(f'\nFrame: {frame_num + 1} of {num_video_frames}.')
frame = vae_encoded[frame_num, :, :, :].unsqueeze(0)
flux_latent = flux_sampler.sample(randnoise, guider, sampler, sigmas, {'samples': frame})[1]['samples']
flux_img = vae.decode(flux_latent)
flux_seq_latents.append(flux_latent)
flux_seq_imgs.append(flux_img)
flux_seq_latents = torch.cat(flux_seq_latents, dim=0)
flux_seq_imgs = torch.cat(flux_seq_imgs, dim=0)
return flux_seq_latents, flux_seq_imgs
def sample(self, seed, fs_params, width, height, ckpt_name, sampling_algorithm, scheduling_algorithm, svd_steps,
flux_steps, svd_min_guidance, svd_guidance, flux_guidance, flux_denoise, init_image, selected_img, num_video_frames, motion_type_id,
fps, augmentation_level, sequence_count):
print_fs_params = copy.deepcopy(fs_params)[selected_img - 1]
del print_fs_params['model']
del print_fs_params['vae']
del print_fs_params['positive_conditioning']
print(
f'\n FlowState FVD Sampler - Loading models.\n'
f'\n - seed: {seed}'
f'\n - fs_params: {print_fs_params}'
f'\n - width: {width}'
f'\n - height: {height}'
f'\n - sampling_algorithm: {sampling_algorithm}'
f'\n - scheduling_algorithm: {scheduling_algorithm}'
f'\n - svd_steps: {svd_steps}'
f'\n - flux_steps: {flux_steps}'
f'\n - min_guidance: {svd_min_guidance}'
f'\n - svd_guidance: {svd_guidance}'
f'\n - flux_guidance: {flux_guidance}'
f'\n - flux_denoise: {flux_denoise}'
f'\n - num_video_frames: {num_video_frames}'
f'\n - motion_type_id: {motion_type_id}'
f'\n - fps: {fps}'
f'\n - augmentation_level: {augmentation_level}'
f'\n - sequence_count: {sequence_count}\n\n'
)
start_time = time.time()
model_components = self.load_checkpoint(ckpt_name)
model_patcher = model_components[0]
clip_vision = model_components[1]
svd_vae = model_components[2]
patched_model = self.patch(model_patcher, svd_min_guidance)
init_img_params = fs_params[selected_img - 1]
print(f'\n FlowState FVD Sampler - Sampling.')
sequence_latents = []
sequence_imgs = []
for seq_num in range(sequence_count):
print(f' - Generating SVD Video Sequence: {seq_num + 1} of {sequence_count}.\n')
init_frame = init_image if seq_num == 0 else sequence_imgs[-1][num_video_frames - 1, :, :, :].unsqueeze(0)
encoded_components = self.encode(clip_vision, init_frame, svd_vae, width, height, num_video_frames, motion_type_id, fps, augmentation_level)
positive_conditioning = encoded_components[0]
negative_conditioning = encoded_components[1]
latent_frames = encoded_components[2]
latent_batch = common_ksampler(
patched_model, seed, svd_steps, svd_guidance, sampling_algorithm, scheduling_algorithm, positive_conditioning,
negative_conditioning, latent_frames, denoise=1
)
svd_latents = latent_batch[0]['samples']
svd_imgs = svd_vae.decode(svd_latents)
print(f'\n FlowState FVD Sampler - Processing frames with Flux.')
flux_seq_latents, flux_seq_imgs = self.sample_flux(init_img_params, svd_imgs, flux_denoise, flux_steps, flux_guidance, num_video_frames)
sequence_latents.append(flux_seq_latents)
sequence_imgs.append(flux_seq_imgs)
latent_out = {'samples': torch.cat(sequence_latents, dim=0)}
print(
f'\n FlowState FVD Sampler - Sampling Complete.'
f'\n - Decoding Batch: {latent_out["samples"].shape}\n'
)
img_out = torch.cat(sequence_imgs, dim=0)
sampling_duration = time.time() - start_time
sampling_mins = int(sampling_duration // 60)
sampling_secs = int(sampling_duration - sampling_mins * 60)
print(
f'\n FlowState FVD Sampler - Generation complete.'
f'\n - Total Generated Frames: {latent_out["samples"].shape[0]}'
f'\n - Output Resolution: {img_out.shape[1]} x {img_out.shape[2]}\n'
f'\n - Generation Time: {sampling_mins}m {sampling_secs}s\n'
)
return (latent_out, img_out, )
# q, k, v, out - clip attn presets
# update node definitions - use anything multi - is_changed - selected_img
# no multiple llm gens for same prompt / different settings (qkv)
# x/y output grid
# new stream schedule announcement