Spaces:
Runtime error
Runtime error
foz
commited on
Commit
•
aada7c5
1
Parent(s):
046b08b
Fix requirements
Browse files- app.py +7 -14
- app_pose.py +0 -2
- model.py +68 -96
- requirements.txt +0 -1
- utils.py +4 -6
app.py
CHANGED
@@ -1,17 +1,14 @@
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
|
4 |
-
from model import Model
|
5 |
from app_pose import create_demo as create_demo_pose
|
6 |
import argparse
|
7 |
import os
|
8 |
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
parser.add_argument('--public_access', action='store_true',
|
13 |
-
help="if enabled, the app can be access from a public url", default=False)
|
14 |
-
args = parser.parse_args()
|
15 |
|
16 |
|
17 |
with gr.Blocks(css='style.css') as demo:
|
@@ -22,10 +19,6 @@ with gr.Blocks(css='style.css') as demo:
|
|
22 |
'''
|
23 |
|
24 |
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
else:
|
29 |
-
_, _, link = demo.queue(api_open=False).launch(
|
30 |
-
file_directories=['temporal'], share=args.public_access)
|
31 |
-
print(link)
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
|
4 |
+
from model import Model
|
5 |
from app_pose import create_demo as create_demo_pose
|
6 |
import argparse
|
7 |
import os
|
8 |
|
9 |
+
model = Model()
|
10 |
+
|
11 |
+
|
|
|
|
|
|
|
12 |
|
13 |
|
14 |
with gr.Blocks(css='style.css') as demo:
|
|
|
19 |
'''
|
20 |
|
21 |
|
22 |
+
|
23 |
+
demo.launch(debug=True)
|
24 |
+
|
|
|
|
|
|
|
|
app_pose.py
CHANGED
@@ -1,7 +1,5 @@
|
|
1 |
from model import Model
|
2 |
import gradio as gr
|
3 |
-
import os
|
4 |
-
on_huggingspace = os.environ.get("SPACE_AUTHOR_NAME") == "PAIR"
|
5 |
|
6 |
examples = [
|
7 |
['Motion 1', "An astronaut dancing in the outer space"],
|
|
|
1 |
from model import Model
|
2 |
import gradio as gr
|
|
|
|
|
3 |
|
4 |
examples = [
|
5 |
['Motion 1', "An astronaut dancing in the outer space"],
|
model.py
CHANGED
@@ -4,111 +4,95 @@ import numpy as np
|
|
4 |
import torch
|
5 |
|
6 |
|
7 |
-
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
|
8 |
-
from diffusers import StableDiffusionInstructPix2PixPipeline, StableDiffusionControlNetPipeline, ControlNetModel, UNet2DConditionModel
|
9 |
-
from diffusers.schedulers import EulerAncestralDiscreteScheduler, DDIMScheduler
|
10 |
|
11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
|
14 |
import utils
|
15 |
import gradio_utils
|
16 |
import os
|
17 |
-
on_huggingspace = os.environ.get("SPACE_AUTHOR_NAME") == "PAIR"
|
18 |
|
19 |
from einops import rearrange
|
20 |
|
|
|
21 |
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
|
26 |
class Model:
|
27 |
-
def __init__(self,
|
28 |
-
self.
|
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 |
-
self.pipe
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
latents = None
|
66 |
-
if 'latents' in kwargs:
|
67 |
-
latents = kwargs.pop('latents')[frame_ids]
|
68 |
-
if 'image' in kwargs:
|
69 |
-
kwargs['image'] = kwargs['image'][frame_ids]
|
70 |
-
if 'video_length' in kwargs:
|
71 |
-
kwargs['video_length'] = len(frame_ids)
|
72 |
-
return self.pipe(prompt=prompt[frame_ids].tolist(),
|
73 |
-
negative_prompt=negative_prompt[frame_ids].tolist(),
|
74 |
-
latents=latents,
|
75 |
-
generator=self.generator,
|
76 |
-
**kwargs)
|
77 |
|
78 |
def inference(self, **kwargs):
|
79 |
-
|
80 |
-
return
|
81 |
-
|
82 |
seed = kwargs.pop('seed', 0)
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
86 |
|
87 |
-
if 'image' in kwargs:
|
88 |
-
f = kwargs['image'].shape[0]
|
89 |
-
else:
|
90 |
-
f = kwargs['video_length']
|
91 |
|
92 |
assert 'prompt' in kwargs
|
93 |
prompt = [kwargs.pop('prompt')] * f
|
94 |
negative_prompt = [kwargs.pop('negative_prompt', '')] * f
|
95 |
|
96 |
frames_counter = 0
|
97 |
-
|
98 |
-
# Processing frame_by_frame
|
99 |
result = []
|
100 |
-
for i in range(f):
|
101 |
-
frame_ids = [0] + [i]
|
102 |
-
self.generator.manual_seed(seed)
|
103 |
print(f'Processing frame {i + 1} / {f}')
|
104 |
-
result.append(self.
|
105 |
prompt=prompt,
|
106 |
negative_prompt=negative_prompt,
|
107 |
-
|
|
|
108 |
frames_counter += 1
|
109 |
-
|
110 |
-
break
|
111 |
-
result = np.concatenate(result)
|
112 |
return result
|
113 |
|
114 |
def process_controlnet_pose(self,
|
@@ -120,33 +104,22 @@ class Model:
|
|
120 |
seed=42,
|
121 |
eta=0.0,
|
122 |
resolution=512,
|
123 |
-
use_cf_attn=True,
|
124 |
save_path=None):
|
125 |
print("Module Pose")
|
126 |
video_path = gradio_utils.motion_to_video_path(video_path)
|
127 |
-
if self.model_type != ModelType.ControlNetPose:
|
128 |
-
controlnet = ControlNetModel.from_pretrained(
|
129 |
-
"fusing/stable-diffusion-v1-5-controlnet-openpose", torch_dtype=torch.float16)
|
130 |
-
self.set_model(ModelType.ControlNetPose,
|
131 |
-
model_id="runwayml/stable-diffusion-v1-5", controlnet=controlnet)
|
132 |
-
self.pipe.scheduler = DDIMScheduler.from_config(
|
133 |
-
self.pipe.scheduler.config)
|
134 |
|
135 |
-
video_path = gradio_utils.motion_to_video_path(
|
136 |
-
video_path) if 'Motion' in video_path else video_path
|
137 |
|
138 |
added_prompt = 'best quality, extremely detailed, HD, ultra-realistic, 8K, HQ, masterpiece, trending on artstation, art, smooth'
|
139 |
negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic'
|
140 |
|
141 |
video, fps = utils.prepare_video(
|
142 |
-
video_path, resolution,
|
143 |
control = utils.pre_process_pose(
|
144 |
-
video, apply_pose_detect=False)
|
|
|
|
|
145 |
f, _, h, w = video.shape
|
146 |
-
|
147 |
-
latents = torch.randn((1, 4, h//8, w//8), dtype=self.dtype,
|
148 |
-
device=self.device, generator=self.generator)
|
149 |
-
latents = latents.repeat(f, 1, 1, 1)
|
150 |
result = self.inference(image=control,
|
151 |
prompt=prompt + ', ' + added_prompt,
|
152 |
height=h,
|
@@ -156,9 +129,8 @@ class Model:
|
|
156 |
guidance_scale=guidance_scale,
|
157 |
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
158 |
eta=eta,
|
159 |
-
latents=latents,
|
160 |
seed=seed,
|
161 |
output_type='numpy',
|
162 |
)
|
163 |
-
return utils.create_gif(result, fps, path=save_path)
|
164 |
|
|
|
4 |
import torch
|
5 |
|
6 |
|
|
|
|
|
|
|
7 |
|
8 |
+
|
9 |
+
import jax
|
10 |
+
import jax.numpy as jnp
|
11 |
+
import numpy as np
|
12 |
+
from flax.jax_utils import replicate
|
13 |
+
from flax.training.common_utils import shard
|
14 |
+
from PIL import Image
|
15 |
+
from diffusers import FlaxStableDiffusionControlNetPipeline, FlaxControlNetModel
|
16 |
|
17 |
|
18 |
import utils
|
19 |
import gradio_utils
|
20 |
import os
|
|
|
21 |
|
22 |
from einops import rearrange
|
23 |
|
24 |
+
import matplotlib.pyplot as plt
|
25 |
|
26 |
+
def create_key(seed=0):
|
27 |
+
return jax.random.PRNGKey(seed)
|
|
|
28 |
|
29 |
class Model:
|
30 |
+
def __init__(self, **kwargs):
|
31 |
+
self.base_controlnet, self.base_controlnet_params = FlaxControlNetModel.from_pretrained(
|
32 |
+
#"JFoz/dog-cat-pose", dtype=jnp.bfloat16
|
33 |
+
"lllyasviel/control_v11p_sd15_openpose", dtype=jnp.bfloat16, from_pt=True
|
34 |
+
)
|
35 |
+
self.pipe, self.params = FlaxStableDiffusionControlNetPipeline.from_pretrained(
|
36 |
+
"runwayml/stable-diffusion-v1-5", controlnet=self.base_controlnet, revision="flax", dtype=jnp.bfloat16,# from_pt=True,
|
37 |
+
)
|
38 |
+
|
39 |
+
def infer_frame(self, frame_id, prompt, negative_prompt, rng, **kwargs):
|
40 |
+
|
41 |
+
print(prompt, frame_id)
|
42 |
+
|
43 |
+
num_samples = 1
|
44 |
+
prompt_ids = self.pipe.prepare_text_inputs([prompt[frame_id]]*num_samples)
|
45 |
+
negative_prompt_ids = self.pipe.prepare_text_inputs([negative_prompt[frame_id]] * num_samples)
|
46 |
+
processed_image = self.pipe.prepare_image_inputs([kwargs['image'][frame_id]]*num_samples)
|
47 |
+
|
48 |
+
self.params["controlnet"] = self.base_controlnet_params
|
49 |
+
|
50 |
+
|
51 |
+
p_params = replicate(self.params)
|
52 |
+
prompt_ids = shard(prompt_ids)
|
53 |
+
negative_prompt_ids = shard(negative_prompt_ids)
|
54 |
+
processed_image = shard(processed_image)
|
55 |
+
|
56 |
+
output = self.pipe(
|
57 |
+
prompt_ids=prompt_ids,
|
58 |
+
image=processed_image,
|
59 |
+
params=p_params,
|
60 |
+
prng_seed=rng,
|
61 |
+
num_inference_steps=50,
|
62 |
+
neg_prompt_ids=negative_prompt_ids,
|
63 |
+
jit=True,
|
64 |
+
).images
|
65 |
+
|
66 |
+
output_images = np.asarray(output.reshape((num_samples,) + output.shape[-3:]))
|
67 |
+
return output_images
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
|
69 |
def inference(self, **kwargs):
|
70 |
+
|
|
|
|
|
71 |
seed = kwargs.pop('seed', 0)
|
72 |
+
|
73 |
+
rng = create_key(0)
|
74 |
+
rng = jax.random.split(rng, jax.device_count())
|
75 |
+
|
76 |
+
f = len(kwargs['image'])
|
77 |
+
print('frames', f)
|
78 |
|
|
|
|
|
|
|
|
|
79 |
|
80 |
assert 'prompt' in kwargs
|
81 |
prompt = [kwargs.pop('prompt')] * f
|
82 |
negative_prompt = [kwargs.pop('negative_prompt', '')] * f
|
83 |
|
84 |
frames_counter = 0
|
85 |
+
|
|
|
86 |
result = []
|
87 |
+
for i in range(0, f):
|
|
|
|
|
88 |
print(f'Processing frame {i + 1} / {f}')
|
89 |
+
result.append(self.infer_frame(frame_id=i,
|
90 |
prompt=prompt,
|
91 |
negative_prompt=negative_prompt,
|
92 |
+
rng = rng,
|
93 |
+
**kwargs))
|
94 |
frames_counter += 1
|
95 |
+
result = np.stack(result, axis=0)
|
|
|
|
|
96 |
return result
|
97 |
|
98 |
def process_controlnet_pose(self,
|
|
|
104 |
seed=42,
|
105 |
eta=0.0,
|
106 |
resolution=512,
|
|
|
107 |
save_path=None):
|
108 |
print("Module Pose")
|
109 |
video_path = gradio_utils.motion_to_video_path(video_path)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
110 |
|
|
|
|
|
111 |
|
112 |
added_prompt = 'best quality, extremely detailed, HD, ultra-realistic, 8K, HQ, masterpiece, trending on artstation, art, smooth'
|
113 |
negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic'
|
114 |
|
115 |
video, fps = utils.prepare_video(
|
116 |
+
video_path, resolution, False, output_fps=4)
|
117 |
control = utils.pre_process_pose(
|
118 |
+
video, apply_pose_detect=False)
|
119 |
+
|
120 |
+
print('N frames', len(control))
|
121 |
f, _, h, w = video.shape
|
122 |
+
|
|
|
|
|
|
|
123 |
result = self.inference(image=control,
|
124 |
prompt=prompt + ', ' + added_prompt,
|
125 |
height=h,
|
|
|
129 |
guidance_scale=guidance_scale,
|
130 |
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
131 |
eta=eta,
|
|
|
132 |
seed=seed,
|
133 |
output_type='numpy',
|
134 |
)
|
135 |
+
return utils.create_gif(result.astype(jnp.float16), fps, path=save_path)
|
136 |
|
requirements.txt
CHANGED
@@ -7,7 +7,6 @@ git+https://github.com/huggingface/diffusers@main
|
|
7 |
torch
|
8 |
accelerate
|
9 |
decord==0.6.0
|
10 |
-
diffusers==0.16.1
|
11 |
einops
|
12 |
gradio
|
13 |
imageio
|
|
|
7 |
torch
|
8 |
accelerate
|
9 |
decord==0.6.0
|
|
|
10 |
einops
|
11 |
gradio
|
12 |
imageio
|
utils.py
CHANGED
@@ -15,7 +15,7 @@ from controlnet_aux import OpenposeDetector
|
|
15 |
|
16 |
apply_openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
|
17 |
|
18 |
-
def prepare_video(video_path:str, resolution:int,
|
19 |
vr = decord.VideoReader(video_path)
|
20 |
initial_fps = vr.get_avg_fps()
|
21 |
if output_fps == -1:
|
@@ -37,7 +37,7 @@ def prepare_video(video_path:str, resolution:int, device, dtype, normalize=True,
|
|
37 |
video = video.asnumpy()
|
38 |
_, h, w, _ = video.shape
|
39 |
video = rearrange(video, "f h w c -> f c h w")
|
40 |
-
video = torch.Tensor(video)
|
41 |
|
42 |
# Use max if you want the larger side to be equal to resolution (e.g. 512)
|
43 |
# k = float(resolution) / min(h, w)
|
@@ -63,10 +63,8 @@ def pre_process_pose(input_video, apply_pose_detect: bool = True):
|
|
63 |
detected_map = img
|
64 |
H, W, C = img.shape
|
65 |
detected_map = cv2.resize(detected_map, (W, H), interpolation=cv2.INTER_NEAREST)
|
66 |
-
detected_maps.append(detected_map
|
67 |
-
|
68 |
-
control = torch.from_numpy(detected_maps.copy()).float() / 255.0
|
69 |
-
return rearrange(control, 'f h w c -> f c h w')
|
70 |
|
71 |
|
72 |
|
|
|
15 |
|
16 |
apply_openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
|
17 |
|
18 |
+
def prepare_video(video_path:str, resolution:int, normalize=True, start_t:float=0, end_t:float=-1, output_fps:int=-1):
|
19 |
vr = decord.VideoReader(video_path)
|
20 |
initial_fps = vr.get_avg_fps()
|
21 |
if output_fps == -1:
|
|
|
37 |
video = video.asnumpy()
|
38 |
_, h, w, _ = video.shape
|
39 |
video = rearrange(video, "f h w c -> f c h w")
|
40 |
+
video = torch.Tensor(video)
|
41 |
|
42 |
# Use max if you want the larger side to be equal to resolution (e.g. 512)
|
43 |
# k = float(resolution) / min(h, w)
|
|
|
63 |
detected_map = img
|
64 |
H, W, C = img.shape
|
65 |
detected_map = cv2.resize(detected_map, (W, H), interpolation=cv2.INTER_NEAREST)
|
66 |
+
detected_maps.append(Image.fromarray(detected_map))
|
67 |
+
return detected_maps
|
|
|
|
|
68 |
|
69 |
|
70 |
|