test_stablehairv2.py

#!/usr/bin/env python3
import argparse
import logging
import sys
import os
import random
import numpy as np
import cv2
import torch
from PIL import Image
from transformers import AutoTokenizer, CLIPVisionModelWithProjection
from diffusers import AutoencoderKL, UniPCMultistepScheduler, UNet2DConditionModel
from src.models.unet_3d import UNet3DConditionModel
from ref_encoder.reference_unet import CCProjection
from ref_encoder.latent_controlnet import ControlNetModel
from src.pipelines.pipeline_pose2vid_long import Pose2VideoPipeline as Hair3dPipeline
from src.utils.util import save_videos_grid
from omegaconf import OmegaConf
from HairMapper.hair_mapper_run import bald_head


# face align
def _maybe_align_image(image_path: str, output_size: int, prefer_cuda: bool = True):
    """Align and crop a face image to FFHQ-style using FFHQFaceAlignment if available.
    Falls back to simple resize if alignment fails.
    Returns an RGB uint8 numpy array of shape (H, W, 3).
    """
    try:
        ffhq_dir = os.path.join(os.path.dirname(__file__), 'FFHQFaceAlignment')
        if ffhq_dir not in sys.path:
            sys.path.insert(0, ffhq_dir)
        # Lazy imports to avoid hard dependency if user doesn't enable alignment
        from lib.landmarks_pytorch import LandmarksEstimation
        from align import align_crop_image

        # Read image as RGB uint8
        img_bgr = cv2.imread(image_path, cv2.IMREAD_COLOR)
        if img_bgr is None:
            raise RuntimeError(f"Failed to read image: {image_path}")
        img = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB).astype('uint8')

        device = torch.device('cuda' if prefer_cuda and torch.cuda.is_available() else 'cpu')
        le = LandmarksEstimation(type='2D')

        img_tensor = torch.tensor(np.transpose(img, (2, 0, 1))).float().to(device)
        with torch.no_grad():
            landmarks, _ = le.detect_landmarks(img_tensor.unsqueeze(0), detected_faces=None)
        if len(landmarks) > 0:
            lm = np.asarray(landmarks[0].detach().cpu().numpy())
            aligned = align_crop_image(image=img, landmarks=lm, transform_size=output_size)
            if aligned is None or aligned.size == 0:
                return cv2.resize(img, (output_size, output_size))
            return aligned
        else:
            return cv2.resize(img, (output_size, output_size))
    except Exception:
        # Silent fallback to simple resize on any failure
        img_bgr = cv2.imread(image_path, cv2.IMREAD_COLOR)
        img = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB).astype('uint8') if img_bgr is not None else None
        if img is None:
            raise
        return cv2.resize(img, (output_size, output_size))


def log_validation(
        vae, tokenizer, image_encoder, denoising_unet,
        args, device, logger, cc_projection,
        controlnet, hair_encoder, feature_extractor=None
):
    """
    Run inference on validation pairs and save generated videos.
    """
    logger.info("Starting validation inference...")

    # Initialize inference pipeline
    pipeline = Hair3dPipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        image_encoder=image_encoder,
        feature_extractor=feature_extractor,
        controlnet=controlnet,
        vae=vae,
        tokenizer=tokenizer,
        denoising_unet=denoising_unet,
        safety_checker=None,
        revision=args.revision,
        torch_dtype=torch.float16 if args.use_fp16 else torch.float32,
    ).to(device)
    pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config)
    pipeline.set_progress_bar_config(disable=True)

    # Create output directory
    output_dir = os.path.join(args.output_dir, "validation")
    os.makedirs(output_dir, exist_ok=True)

    print(output_dir)

    # Speed/length overrides via env/args
    import os as _os
    steps = int(_os.getenv('SH_STEPS', getattr(args, 'num_inference_steps', 30)))
    gscale = float(_os.getenv('SH_GUIDANCE', getattr(args, 'guidance_scale', 1.5)))
    vlen = int(_os.getenv('SH_VIDEO_LENGTH', getattr(args, 'video_length', 21)))
    # 统一时序长度：上下文帧数始终等于视频帧数（不再读取 SH_CONTEXT_FRAMES）
    cframes = vlen

    # Generate camera trajectory with exactly vlen frames
    angles = np.linspace(0, 2 * np.pi, vlen, endpoint=False)
    X = 0.4 * np.sin(angles)
    Y = -0.05 + 0.3 * np.cos(angles)
    x_tensor = torch.tensor(X, dtype=torch.float32).unsqueeze(1).to(device)
    y_tensor = torch.tensor(Y, dtype=torch.float32).unsqueeze(1).to(device)

    # # Load reference images
    # id_image = cv2.cvtColor(cv2.imread(args.validation_ids[0]), cv2.COLOR_BGR2RGB)
    # id_image = cv2.resize(id_image, (512, 512))
    # Load reference images (optionally align)
    align_enabled = getattr(args, 'align_before_infer', True)
    align_size = getattr(args, 'align_size', 1024)
    prefer_cuda = True if device.type == 'cuda' else False
    if align_enabled:
        id_image = _maybe_align_image(args.validation_ids[0], output_size=align_size, prefer_cuda=prefer_cuda)
    else:
        id_image = cv2.cvtColor(cv2.imread(args.validation_ids[0]), cv2.COLOR_BGR2RGB)
        id_image = cv2.resize(id_image, (512, 512))

    # ===== ���� HairMapper ͺͷ�� =====
    temp_bald_path = os.path.join(args.output_dir, "bald_id.png")
    cv2.imwrite(temp_bald_path, cv2.cvtColor(id_image, cv2.COLOR_RGB2BGR))  # �������ͼ
    bald_head(temp_bald_path, temp_bald_path)  # ͺͷ�������DZ���
    # ���¼���ͺͷͼ�� (RGB)
    id_image = cv2.cvtColor(cv2.imread(temp_bald_path), cv2.COLOR_BGR2RGB)
    id_image = cv2.resize(id_image, (512, 512))

    id_list = [id_image for _ in range(vlen)]
    if align_enabled:
        hair_image = _maybe_align_image(args.validation_hairs[0], output_size=align_size, prefer_cuda=prefer_cuda)
        prompt_img = _maybe_align_image(args.validation_ids[0], output_size=align_size, prefer_cuda=prefer_cuda)
    else:
        hair_image = cv2.cvtColor(cv2.imread(args.validation_hairs[0]), cv2.COLOR_BGR2RGB)
        hair_image = cv2.resize(hair_image, (512, 512))
        prompt_img = cv2.cvtColor(cv2.imread(args.validation_ids[0]), cv2.COLOR_BGR2RGB)
        prompt_img = cv2.resize(prompt_img, (512, 512))
    hair_image = cv2.resize(hair_image, (512, 512))
    prompt_img = cv2.resize(prompt_img, (512, 512))

    prompt_img = [prompt_img]

    # Perform inference and save videos

    for idx in range(args.num_validation_images):
        result = pipeline(
            prompt="",
            negative_prompt="",
            num_inference_steps=steps,
            guidance_scale=gscale,
            width=512,
            height=512,
            controlnet_condition=id_list,
            controlnet_conditioning_scale=1.0,
            generator=torch.Generator(device).manual_seed(args.seed),
            ref_image=hair_image,
            prompt_img=prompt_img,
            reference_encoder=hair_encoder,
            poses=None,
            x=x_tensor,
            y=y_tensor,
            video_length=vlen,
            context_frames=cframes,
        )
        video = torch.cat([result.videos, result.videos], dim=0)
        video_path = os.path.join(output_dir, f"generated_video_{idx}.mp4")
        save_videos_grid(video, video_path, n_rows=5, fps=24)
        logger.info(f"Saved generated video: {video_path}")


def parse_args():
    parser = argparse.ArgumentParser(
        description="Inference script for 3D hairstyle generation"
    )
    parser.add_argument(
        "--pretrained_model_name_or_path", type=str, required=True,
        help="Path or ID of the pretrained pipeline"
    )
    parser.add_argument(
        "--model_path", type=str, required=True,
        help="Path or ID of the pretrained pipeline"
    )
    parser.add_argument(
        "--image_encoder", type=str, required=True,
        help="Path or ID of the CLIP vision encoder"
    )
    parser.add_argument(
        "--controlnet_model_name_or_path", type=str, default=None,
        help="Path or ID of the ControlNet model"
    )
    parser.add_argument(
        "--revision", type=str, default=None,
        help="Model revision or Git reference"
    )
    parser.add_argument(
        "--output_dir", type=str, default="inference_output",
        help="Directory to save inference results"
    )
    parser.add_argument(
        "--seed", type=int, default=42,
        help="Random seed for reproducibility"
    )
    parser.add_argument(
        "--num_validation_images", type=int, default=3,
        help="Number of videos to generate per input pair"
    )
    parser.add_argument(
        "--validation_ids", type=str, nargs='+', required=True,
        help="Path(s) to identity conditioning images"
    )
    parser.add_argument(
        "--validation_hairs", type=str, nargs='+', required=True,
        help="Path(s) to hairstyle reference images"
    )
    parser.add_argument(
        "--use_fp16", action="store_true",
        help="Enable fp16 inference"
    )
    parser.add_argument(
        "--align_before_infer", action="store_true", default=True,
        help="Align and crop input images to FFHQ style before inference"
    )
    parser.add_argument(
        "--align_size", type=int, default=1024,
        help="Output size for aligned images when alignment is enabled"
    )
    return parser.parse_args()


def main():
    args = parse_args()
    # Setup device and logger
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger = logging.getLogger(__name__)

    # Set random seed
    torch.manual_seed(args.seed)
    if device.type == "cuda":
        torch.cuda.manual_seed_all(args.seed)

    # Load models
    tokenizer = AutoTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer",
        revision=args.revision
    )
    image_encoder = CLIPVisionModelWithProjection.from_pretrained(
        args.image_encoder,
        revision=args.revision
    ).to(device)
    vae = AutoencoderKL.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="vae",
        revision=args.revision
    ).to(device)

    infer_config = OmegaConf.load('./configs/inference/inference_v2.yaml')

    unet2 = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", use_safetensors=True, revision=args.revision,
        torch_dtype=torch.float16
    ).to(device)
    conv_in_8 = torch.nn.Conv2d(8, unet2.conv_in.out_channels, kernel_size=unet2.conv_in.kernel_size,
                                padding=unet2.conv_in.padding)
    conv_in_8.requires_grad_(False)
    unet2.conv_in.requires_grad_(False)
    torch.nn.init.zeros_(conv_in_8.weight)
    conv_in_8.weight[:, :4, :, :].copy_(unet2.conv_in.weight)
    conv_in_8.bias.copy_(unet2.conv_in.bias)
    unet2.conv_in = conv_in_8

    # Load or initialize ControlNet
    controlnet = ControlNetModel.from_unet(unet2).to(device)
    # state_dict2 = torch.load(os.path.join(args.model_path, "pytorch_model.bin"), map_location=torch.device('cpu'))
    # state_dict2 = torch.load(args.model_path, map_location=torch.device('cpu'))
    state_dict2 = torch.load(os.path.join(args.model_path, "pytorch_model.bin"), map_location=torch.device('cpu'))
    controlnet.load_state_dict(state_dict2, strict=False)

    # Load 3D UNet motion module
    prefix = "motion_module"
    ckpt_num = "4140000"
    save_path = os.path.join(args.model_path, f"{prefix}-{ckpt_num}.pth")

    denoising_unet = UNet3DConditionModel.from_pretrained_2d(
        args.pretrained_model_name_or_path,
        save_path,
        subfolder="unet",

        unet_additional_kwargs=infer_config.unet_additional_kwargs,
    ).to(device)

    # Load projection and hair encoder
    cc_projection = CCProjection().to(device)
    state_dict3 = torch.load(os.path.join(args.model_path, "pytorch_model_1.bin"), map_location=torch.device('cpu'))
    cc_projection.load_state_dict(state_dict3, strict=False)

    from ref_encoder.reference_unet import ref_unet
    Hair_Encoder = ref_unet.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, low_cpu_mem_usage=False,
        device_map=None, ignore_mismatched_sizes=True
    ).to(device)

    state_dict2 = torch.load(os.path.join(args.model_path, "pytorch_model_2.bin"), map_location=torch.device('cpu'))
    # state_dict2 = torch.load(os.path.join('/home/jichao.zhang/code/3dhair/train_sv3d/checkpoint-30000/', "pytorch_model.bin"))
    Hair_Encoder.load_state_dict(state_dict2, strict=False)

    # Run validation inference
    log_validation(
        vae, tokenizer, image_encoder, denoising_unet,
        args, device, logger,
        cc_projection, controlnet, Hair_Encoder
    )


if __name__ == "__main__":
    main()