SteerX/eval_steerx.py at master · byeongjun-park/SteerX · GitHub

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
import os

os.environ["OPENBLAS_NUM_THREADS"] = "4"
os.environ["MKL_NUM_THREADS"] = "4"
os.environ["RAYON_NUM_THREADS"] = "4"


from argparse import ArgumentParser
from glob import glob
from tqdm import tqdm
import pandas as pd
import torch
from diffusers.utils import load_image, export_to_video
from steerx_diffusers import load_steerx
import cv2
from pathlib import Path

os.environ['PYTHONPATH'] = os.getcwd()


torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True

parser = ArgumentParser()
parser.add_argument('--model', type=str, default='splatflow', choices=['splatflow', 'dimensionx', 'cog', 'mochi', 'hunyuan'])
parser.add_argument('--num_particles', type=int, default=4)  # batch size
args = parser.parse_args()

pipe, reward_fn = load_steerx(args.model, device='cuda')

save_path = os.path.join('samples', args.model, str(args.num_particles))

# In eval mode, we disable tqdm
pipe.set_progress_bar_config(disable=True)
torch.set_num_threads(4)
cv2.setNumThreads(0)

if args.model == 'splatflow':  # test t3bench
    from gs_refine.refine import load_refiner, refine_3DGS
    refiner = load_refiner('cuda')

    with open('benchmarks/t3bench.txt', 'r') as f:
        text = f.readlines()

    for idx, row in tqdm(enumerate(text)):
        PROMPT = row.strip()

        path = os.path.join(save_path, str(idx))

        if os.path.exists(path):
            continue

        os.makedirs(path, exist_ok=True)

        pipeout = pipe(
            [PROMPT] * args.num_particles,
            reward_fn=reward_fn,
            num_steering=4,
            height=256,
            width=256,
            num_inference_steps=200,
            num_frames=8,
            generator=torch.Generator(device="cuda").manual_seed(42),
        )

        reward_fn.export_scene(pipeout, path)
        refine_3DGS(refiner, Path(path), PROMPT)

elif args.model == 'dimensionx':
    from gs_refine.refine import load_refiner, refine_3DGS
    refiner = load_refiner('cuda')

    fnames = sorted(glob('benchmarks/VBench/*.jpg'))

    for idx, fname in tqdm(enumerate(fnames)):
        PROMPT = fname.split('/')[-1].split('.')[0]

        image = load_image(fname)
        path = os.path.join(save_path, PROMPT)

        if os.path.exists(path):
            continue

        os.makedirs(path, exist_ok=True)

        pipeout = pipe(
            [image] * args.num_particles,
            [PROMPT] * args.num_particles,
            reward_fn=reward_fn,
            num_steering=4,
            num_inference_steps=50,
            generator=torch.Generator(device="cuda").manual_seed(42),
            use_dynamic_cfg=True
        )

        export_to_video(pipeout.video[0], os.path.join(path, '_video.mp4'))
        reward_fn.export_scene(pipeout, path)
        refine_3DGS(refiner, Path(path), PROMPT)

elif args.model == 'cog':  # test VBench
    fnames = sorted(glob('benchmarks/VBench/*.jpg'))
    for fname in tqdm(fnames):
        PROMPT = fname.split('/')[-1].split('.')[0]
        negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"

        image = load_image(fname)
        path = os.path.join(save_path, PROMPT)

        if os.path.exists(path):
            continue

        os.makedirs(path, exist_ok=True)

        pipeout = pipe(
            [image] * args.num_particles,
            [PROMPT] * args.num_particles,
            negative_prompt=negative_prompt,
            reward_fn=reward_fn,
            num_steering=2,
            height=768,
            width=768,
            num_inference_steps=50,
            num_frames=49,
            generator=torch.Generator(device="cuda").manual_seed(42),
            use_dynamic_cfg=True
        )

        export_to_video(pipeout.video[0], os.path.join(path, '_video.mp4'))
        reward_fn.export_scene(pipeout, path)

else:  # test penguin
    df = pd.read_csv('benchmarks/penguin_camera.csv', index_col=0)

    for idx, row in tqdm(df.iterrows()):
        PROMPT = row.to_list()[0]

        path = os.path.join(save_path, str(idx))
        if os.path.exists(path):
            continue

        os.makedirs(path, exist_ok=True)

        pipeout = pipe(
            [PROMPT] * args.num_particles,
            reward_fn=reward_fn,
            num_steering=2,
            height=480,
            width=480,
            num_frames=49,
            num_inference_steps=50,
            generator=torch.Generator(device="cuda").manual_seed(42)
        )

        export_to_video(pipeout.video[0], os.path.join(path, '_video.mp4'))
        reward_fn.export_scene(pipeout, path)