English
medical
video
generation
File size: 6,599 Bytes
aacd658
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
MediGen-14B - Medical Video Generation

Fine-tuned Wan2.1-T2V-14B on MedVideoCap-55K dataset for generating
medical-domain videos from text descriptions.

Usage:
    # Single video generation
    python inference.py --prompt "A doctor examining a patient" --output exam.mp4

    # Batch generation from JSON file (list of strings or objects with "prompt" key)
    python inference.py --batch prompts.json --output_dir results/

    # Use a specific GPU
    python inference.py --prompt "..." --gpu 1 --output result.mp4

    # Custom resolution and seed
    python inference.py --prompt "..." --height 480 --width 832 --seed 123
"""

import torch
import os
import json
import argparse
import gc
from pathlib import Path
from safetensors.torch import load_file

# ---------------------------------------------------------------------------
# Paths - all model weights are stored under models/ relative to this script
# ---------------------------------------------------------------------------
SCRIPT_DIR = Path(__file__).resolve().parent
MODELS_DIR = SCRIPT_DIR / "models"

# Default negative prompt to suppress common artifacts in generated videos
NEGATIVE_PROMPT = (
    "Distorted, blurry, low quality, watermark, text overlay, "
    "static image, worst quality, JPEG artifacts, deformed, "
    "extra limbs, bad anatomy"
)


def load_pipeline(device="cuda"):
    """Load the full MediGen-14B pipeline.

    This involves three steps:
    1. Load the base Wan2.1-T2V-14B model (DIT + T5 text encoder + VAE)
    2. Load the UMT5-XXL tokenizer for text encoding
    3. Apply the fine-tuned DIT weights on top of the base model

    Args:
        device: Target device, default "cuda".

    Returns:
        WanVideoPipeline ready for inference.
    """
    from diffsynth.pipelines.wan_video import WanVideoPipeline, ModelConfig

    print("Loading MediGen-14B...")

    # Step 1: Load base pipeline components
    # - DIT (Diffusion Transformer): 6 shards, ~27GB total
    # - T5 text encoder: converts text prompts to embeddings
    # - VAE: decodes latent representations into video frames
    pipe = WanVideoPipeline.from_pretrained(
        torch_dtype=torch.bfloat16,
        device=device,
        model_configs=[
            ModelConfig(
                model_id="Wan-AI/Wan2.1-T2V-14B",
                origin_file_pattern="diffusion_pytorch_model*.safetensors",
            ),
            ModelConfig(
                model_id="Wan-AI/Wan2.1-T2V-14B",
                origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth",
            ),
            ModelConfig(
                model_id="Wan-AI/Wan2.1-T2V-14B",
                origin_file_pattern="Wan2.1_VAE.pth",
            ),
        ],
        tokenizer_config=ModelConfig(
            model_id="Wan-AI/Wan2.1-T2V-1.3B",
            origin_file_pattern="google/umt5-xxl/",
        ),
    )

    # Step 2: Apply fine-tuned DIT weights
    # Only the DIT component was fine-tuned; T5 and VAE remain unchanged
    ckpt_path = MODELS_DIR / "medigen-14b.safetensors"
    state_dict = load_file(str(ckpt_path))
    pipe.dit.load_state_dict(state_dict, strict=False)

    # Free checkpoint memory after loading
    del state_dict
    gc.collect()
    torch.cuda.empty_cache()

    print("MediGen-14B ready.")
    return pipe


def generate_video(pipe, prompt, output_path, seed=42, height=480, width=832):
    """Generate a single video from a text prompt.

    Args:
        pipe: Loaded WanVideoPipeline instance.
        prompt: Text description of the desired medical video.
        output_path: Path to save the output .mp4 file.
        seed: Random seed for reproducibility.
        height: Video height in pixels (default 480).
        width: Video width in pixels (default 832).
    """
    from diffsynth.utils.data import save_video

    print(f"Generating: {prompt[:80]}...")

    # Run diffusion inference with 50 denoising steps (default)
    # tiled=True enables tiled VAE decoding to reduce VRAM usage
    video = pipe(
        prompt=prompt,
        negative_prompt=NEGATIVE_PROMPT,
        seed=seed,
        height=height,
        width=width,
        tiled=True,
    )

    # Save as MP4 at 15fps with quality level 5
    save_video(video, output_path, fps=15, quality=5)
    print(f"Saved: {output_path}")


def main():
    parser = argparse.ArgumentParser(
        description="MediGen-14B Medical Video Generation"
    )
    parser.add_argument("--prompt", type=str, help="Text prompt for generation")
    parser.add_argument("--batch", type=str, help="JSON file with prompts")
    parser.add_argument("--output", type=str, default="output.mp4",
                        help="Output path for single video (default: output.mp4)")
    parser.add_argument("--output_dir", type=str, default="outputs",
                        help="Output directory for batch mode (default: outputs/)")
    parser.add_argument("--seed", type=int, default=42,
                        help="Random seed for reproducibility (default: 42)")
    parser.add_argument("--height", type=int, default=480,
                        help="Video height in pixels (default: 480)")
    parser.add_argument("--width", type=int, default=832,
                        help="Video width in pixels (default: 832)")
    parser.add_argument("--gpu", type=int, default=0,
                        help="GPU device ID (default: 0)")
    args = parser.parse_args()

    # Set visible GPU before any CUDA operations
    os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)

    # Load model (takes ~3-5 minutes for 14B model)
    pipe = load_pipeline()

    if args.batch:
        # Batch mode: read JSON file containing a list of prompts
        # Accepts either ["prompt1", "prompt2", ...] or [{"prompt": "..."}, ...]
        with open(args.batch) as f:
            prompts = json.load(f)
        os.makedirs(args.output_dir, exist_ok=True)

        for i, item in enumerate(prompts):
            prompt = item if isinstance(item, str) else item.get("prompt", "")
            out_path = os.path.join(args.output_dir, f"{i:03d}.mp4")
            generate_video(
                pipe, prompt, out_path,
                seed=args.seed + i,  # Different seed per video
                height=args.height, width=args.width,
            )

    elif args.prompt:
        # Single video mode
        generate_video(
            pipe, args.prompt, args.output,
            seed=args.seed, height=args.height, width=args.width,
        )

    else:
        print("Error: provide --prompt or --batch")
        parser.print_help()


if __name__ == "__main__":
    main()