Unison/inference.py at main · ali-vilab/Unison · 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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
import os
import re
import argparse
import ffmpeg
import warnings
warnings.filterwarnings('ignore')

import torch
import torch.nn as nn
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

import wan
from wan.utils.utils import cache_video, cache_image
from vace.models.wan.configs import WAN_CONFIGS

from module import ProjMLP, WanVaceProj, process_vace_data


def get_parser():
    parser = argparse.ArgumentParser()
    parser.add_argument("--qwenvl_path", type=str, required=True)
    parser.add_argument("--vace_path", type=str, required=True)
    parser.add_argument("--proj_path", type=str, required=True)
    parser.add_argument("--prompt", type=str, required=True)
    parser.add_argument("--save_path", type=str, required=True)
    return parser


def main():
    args = get_parser().parse_args()
    cfg = WAN_CONFIGS["vace-1.3B"]

    prompt = args.prompt
    visual_content_list = re.findall(r'###(.*?)###', prompt)
    for visual_content_path in visual_content_list:
        if visual_content_path.endswith((".png", ".jpg", ".jpeg")):
            PAD_TOKEN = "<IMGPAD>"
        elif visual_content_path.endswith((".mp4", ".avi", ".mov")):
            PAD_TOKEN = "<VIDPAD>"
        else:
            assert False, "Unsupported file type"
        prompt = prompt.replace(f"###{visual_content_path}###", PAD_TOKEN)

    qwenvl_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(args.qwenvl_path, device_map="auto").to("cuda")
    qwenvl_processor = AutoProcessor.from_pretrained(args.qwenvl_path)

    wan_vace = WanVaceProj(config=cfg, checkpoint_dir=args.vace_path, device_id=0)
    wan_vace.model.text_len = 2048
    projector = ProjMLP(input_dim=qwenvl_model.config.text_config.hidden_size, t5_dim=4096).to("cuda")
    state_dict = torch.load(args.proj_path)
    projector.load_state_dict({k.replace('module.', ''): v for k, v in state_dict.items()})

    messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
    with torch.no_grad():
        inputs = qwenvl_processor.apply_chat_template(
            messages,
            tokenize=True,
            add_generation_prompt=True,
            return_tensors="pt",
            return_dict=True,
        )
        inputs = inputs.to("cuda")
        generated_ids = qwenvl_model.generate(**inputs, max_new_tokens=256)
        generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
        output_text = qwenvl_processor.batch_decode(generated_ids_trimmed, skip_special_tokens=False, clean_up_tokenization_spaces=False)
        output_text = output_text[0].replace('<|im_end|>', '').strip()

    size = (832, 480)
    src_mask = src_video = src_ref = None
    if '<CFI>' in output_text:
        print(output_text)
        num_frames = 1
        if '<BORES>' in output_text:
            match = re.search(r'<BORES>([^<]+)<EORES>', output_text)
            match_str = match.group(1)
            width, height = map(int, match_str.split(','))
            size = (width, height)
        if '<BOEDIT>' in output_text:
            match = re.search(r'<BOEDIT>([^<]+)<EOEDIT>', output_text)
            match_str = match.group(1)
            mask_id, source_id = map(int, match_str.split(','))
            mask_path = visual_content_list[mask_id]
            src_path = visual_content_list[source_id]
            result = process_vace_data(task="inpainting", mode="mask", video=src_path, mask=mask_path, save_fps=24)
            src_mask = result['src_mask']
            src_video = result['src_video']
        if '<REF>' in output_text:
            ref_path = visual_content_list[0]
            result = process_vace_data(task="image_reference", mode="plain", image=ref_path)
            src_ref = result['src_ref_images']
    elif '<CFV>' in output_text:
        print(output_text)
        num_frames = 81
        if '<BORES>' in output_text:
            match = re.search(r'<BORES>([^<]+)<EORES>', output_text)
            match_str = match.group(1)
            width, height = map(int, match_str.split(','))
            size = (width, height)
        if '<BONF>' in output_text:
            match = re.search(r'<BONF>([^<]+)<EONF>', output_text)
            match_str = match.group(1)
            num_frames = int(match_str) + 1
        if '<BOFIDX>' in output_text:
            match = re.search(r'<BOFIDX>([^<]+)<EOFIDX>', output_text)
            match_str = match.group(1)
            mode = "firstframe" if match_str == "match_str" else "lastframe"
            result = process_vace_data(task="frameref", mode=mode, image=visual_content_list[0])
            src_mask = result['src_mask']
            src_video = result['src_video']
        if '<BOEDIT>' in output_text:
            src_path = visual_content_list[0]
            result = process_vace_data(task="inpainting", mode="salient", video=src_path, save_fps=24)
            src_mask = result['src_mask']
            src_video = result['src_video']
        if '<REF>' in output_text:
            ref_path = visual_content_list[0]
            result = process_vace_data(task="image_reference", mode="plain", image=ref_path)
            src_ref = result['src_ref_images']
        if '<CTRL>' in output_text:
            src_video = visual_content_list[0]
    else:
        if len(visual_content_list) == 0:
            messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
        else:
            messages = [{"role": "user", "content": [{"type": "text", "text": prompt}, {"type": "image", "image": visual_content_list[0]}]}]
        with torch.no_grad():
            inputs = qwenvl_processor.apply_chat_template(
                messages,
                tokenize=True,
                add_generation_prompt=True,
                return_tensors="pt",
                return_dict=True,
            )
            inputs = inputs.to("cuda")
            generated_ids = qwenvl_model.generate(**inputs, max_new_tokens=256)
            generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
            output_text = qwenvl_processor.batch_decode(generated_ids_trimmed, skip_special_tokens=False, clean_up_tokenization_spaces=False)
            output_text = output_text[0].replace('<|im_end|>', '').strip()
            print(output_text)
        return

    with torch.no_grad():
        inputs_states = qwenvl_processor.apply_chat_template(
            messages,
            tokenize=True,
            add_generation_prompt=False,
            return_tensors="pt",
            padding=True
        ).to("cuda")
        qwenvl_outputs = qwenvl_model.model(inputs_states).last_hidden_state.to("cuda")
        qwenvl_proj_feat = projector(qwenvl_outputs[:, 1:, :].float())

    src_video, src_mask, src_ref_images = wan_vace.prepare_source([src_video], [src_mask], [None if src_ref is None else src_ref.split(',')], num_frames, size, "cuda")
    video = wan_vace.generate(
        prompt,
        qwenvl_proj_feat,
        src_video,
        src_mask,
        src_ref_images,
        size=size,
        frame_num=num_frames,
        seed=0
    )

    os.makedirs(os.path.split(args.save_path)[0], exist_ok=True)
    if num_frames == 1:
        cache_image(tensor=video[:, 0, ...], save_file=args.save_path, nrow=1, normalize=True, value_range=(-1, 1))
    else:
        cache_video(tensor=video[None], save_file=args.save_path, fps=24, nrow=1, normalize=True, value_range=(-1, 1))

if __name__ == "__main__":
    main()