Stable3DGen/hi3dgen/pipelines/hi3dgen.py

# MIT License

# Copyright (c) Microsoft

# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:

# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.

# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

# Copyright (c) [2025] [Microsoft]
# Copyright (c) [2025] [Chongjie Ye] 
# SPDX-License-Identifier: MIT
# This file has been modified by Chongjie Ye on 2025/04/10
#
# Original file was released under MIT, with the full license text
# available at https://github.com/atong01/conditional-flow-matching/blob/1.0.7/LICENSE.
#
# This modified file is released under the same license.
from typing import *
from contextlib import contextmanager
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from tqdm import tqdm
from torchvision import transforms
from PIL import Image
from .base import Pipeline
from . import samplers
from ..modules import sparse as sp
import os

class Hi3DGenPipeline(Pipeline):

    def __init__(
        self,
        models: dict[str, nn.Module] = None,
        sparse_structure_sampler: samplers.Sampler = None,
        slat_sampler: samplers.Sampler = None,
        slat_normalization: dict = None,
        image_cond_model: str = None,
    ):
        if models is None:
            return
        super().__init__(models)
        self.sparse_structure_sampler = sparse_structure_sampler
        self.slat_sampler = slat_sampler
        self.sparse_structure_sampler_params = {}
        self.slat_sampler_params = {}
        self.slat_normalization = slat_normalization
        self._init_image_cond_model(image_cond_model)

    @staticmethod
    def from_pretrained(path: str) -> "Hi3DGenPipeline":
        """
        Load a pretrained model.

        Args:
            path (str): The path to the model. Can be either local path or a Hugging Face repository.
        """
        pipeline = super(Hi3DGenPipeline, Hi3DGenPipeline).from_pretrained(path)
        new_pipeline = Hi3DGenPipeline()
        new_pipeline.__dict__ = pipeline.__dict__
        args = pipeline._pretrained_args

        new_pipeline.sparse_structure_sampler = getattr(samplers, args['sparse_structure_sampler']['name'])(**args['sparse_structure_sampler']['args'])
        new_pipeline.sparse_structure_sampler_params = args['sparse_structure_sampler']['params']

        new_pipeline.slat_sampler = getattr(samplers, args['slat_sampler']['name'])(**args['slat_sampler']['args'])
        new_pipeline.slat_sampler_params = args['slat_sampler']['params']

        new_pipeline.slat_normalization = args['slat_normalization']

        new_pipeline._init_image_cond_model(args['image_cond_model'])

        return new_pipeline
    
    def _init_image_cond_model(self, name: str):
        """
        Initialize the image conditioning model.
        """

        try:
            dinov2_model = torch.hub.load(os.path.join(torch.hub.get_dir(), 'facebookresearch_dinov2_main'), name, source='local',pretrained=True)
        except:
            dinov2_model = torch.hub.load('facebookresearch/dinov2', name, pretrained=True)
        dinov2_model.eval()
        self.models['image_cond_model'] = dinov2_model
        transform = transforms.Compose([
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ])
        self.image_cond_model_transform = transform

    def preprocess_image(self, input: Image.Image, resolution=518) -> Image.Image:
        """
        Preprocess the input image using BiRefNet for background removal.
        Includes padding to maintain aspect ratio when resizing to 518x518.
        """
        # if has alpha channel, use it directly
        has_alpha = False
        if input.mode == 'RGBA':
            alpha = np.array(input)[:, :, -1]
            if not np.all(alpha == 255):
                has_alpha = True
        
        if has_alpha:
            output = input
        else:
            input = input.convert('RGB')
            max_size = max(input.size)
            scale = min(1, 1024 / max_size)
            if scale < 1:
                input = input.resize((int(input.width * scale), int(input.height * scale)), Image.Resampling.LANCZOS)
            
            # Load BiRefNet model if not already loaded
            if getattr(self, 'birefnet_model', None) is None:
                self._lazy_load_birefnet()
            
            # Get mask using BiRefNet
            mask = self._get_birefnet_mask(input)
            
            # Convert input to RGBA and apply mask
            input_rgba = input.convert('RGBA')
            input_array = np.array(input_rgba)
            input_array[:, :, 3] = mask * 255  # Apply mask to alpha channel
            output = Image.fromarray(input_array)

        # Process the output image
        output_np = np.array(output)
        alpha = output_np[:, :, 3]
        
        # Find bounding box of non-transparent pixels
        bbox = np.argwhere(alpha > 0.8 * 255)
        if len(bbox) == 0:  # Handle case where no foreground is detected
            return input.convert('RGB')
        
        bbox = np.min(bbox[:, 1]), np.min(bbox[:, 0]), np.max(bbox[:, 1]), np.max(bbox[:, 0])
        center = (bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2
        size = max(bbox[2] - bbox[0], bbox[3] - bbox[1])
        size = int(size * 1.2)
        
        # Calculate and apply crop bbox
        bbox = (
            int(center[0] - size // 2),
            int(center[1] - size // 2),
            int(center[0] + size // 2),
            int(center[1] + size // 2)
        )
        
        # Ensure bbox is within image bounds
        bbox = (
            max(0, bbox[0]),
            max(0, bbox[1]),
            min(output.width, bbox[2]),
            min(output.height, bbox[3])
        )
        
        output = output.crop(bbox)
        
        # Add padding to maintain aspect ratio
        width, height = output.size
        if width > height:
            new_height = width
            padding = (width - height) // 2
            padded_output = Image.new('RGBA', (width, new_height), (0, 0, 0, 0))
            padded_output.paste(output, (0, padding))
        else:
            new_width = height
            padding = (height - width) // 2
            padded_output = Image.new('RGBA', (new_width, height), (0, 0, 0, 0))
            padded_output.paste(output, (padding, 0))
        
        # Resize padded image to target size
        padded_output = padded_output.resize((resolution, resolution), Image.Resampling.LANCZOS)
        
        # Final processing
        output = np.array(padded_output).astype(np.float32) / 255
        output = np.dstack((
            output[:, :, :3] * output[:, :, 3:4],  # RGB channels premultiplied by alpha
            output[:, :, 3]                         # Original alpha channel
        ))
        output = Image.fromarray((output * 255).astype(np.uint8), mode='RGBA')
        
        return output

    def _lazy_load_birefnet(self):
        """Lazy loading of the BiRefNet model"""
        from transformers import AutoImageProcessor, Mask2FormerForUniversalSegmentation, AutoModelForImageSegmentation
        self.birefnet_model = AutoModelForImageSegmentation.from_pretrained(
            'weights/BiRefNet',
            trust_remote_code=True
        ).to(self.device)
        self.birefnet_model.eval()

    def _get_birefnet_mask(self, image: Image.Image) -> np.ndarray:
        """Get object mask using BiRefNet"""
        image_size = (1024, 1024)
        transform_image = transforms.Compose([
            transforms.Resize(image_size),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ])
        
        input_images = transform_image(image).unsqueeze(0).to(self.device)
        
        with torch.no_grad():
            preds = self.birefnet_model(input_images)[-1].sigmoid().cpu()
        
        pred = preds[0].squeeze()
        pred_pil = transforms.ToPILImage()(pred)
        mask = pred_pil.resize(image.size)
        mask_np = np.array(mask)

        return (mask_np > 128).astype(np.uint8)

    @torch.no_grad()
    def encode_image(self, image: Union[torch.Tensor, list[Image.Image]]) -> torch.Tensor:
        """
        Encode the image.

        Args:
            image (Union[torch.Tensor, list[Image.Image]]): The image to encode

        Returns:
            torch.Tensor: The encoded features.
        """
        if isinstance(image, torch.Tensor):
            assert image.ndim == 4, "Image tensor should be batched (B, C, H, W)"
        elif isinstance(image, list):
            assert all(isinstance(i, Image.Image) for i in image), "Image list should be list of PIL images"
            image = [i.resize((518, 518), Image.LANCZOS) for i in image]
            image = [np.array(i.convert('RGB')).astype(np.float32) / 255 for i in image]
            image = [torch.from_numpy(i).permute(2, 0, 1).float() for i in image]
            image = torch.stack(image).to(self.device)
        else:
            raise ValueError(f"Unsupported type of image: {type(image)}")
        
        image = self.image_cond_model_transform(image).to(self.device)
        features = self.models['image_cond_model'](image, is_training=True)['x_prenorm']
        patchtokens = F.layer_norm(features, features.shape[-1:])
        return patchtokens
        
    def get_cond(self, image: Union[torch.Tensor, list[Image.Image]]) -> dict:
        """
        Get the conditioning information for the model.

        Args:
            image (Union[torch.Tensor, list[Image.Image]]): The image prompts.

        Returns:
            dict: The conditioning information
        """
        cond = self.encode_image(image)
        neg_cond = torch.zeros_like(cond)
        return {
            'cond': cond,
            'neg_cond': neg_cond,
        }

    def sample_sparse_structure(
        self,
        cond: dict,
        num_samples: int = 1,
        sampler_params: dict = {},
    ) -> torch.Tensor:
        """
        Sample sparse structures with the given conditioning.
        
        Args:
            cond (dict): The conditioning information.
            num_samples (int): The number of samples to generate.
            sampler_params (dict): Additional parameters for the sampler.
        """
        # Sample occupancy latent
        flow_model = self.models['sparse_structure_flow_model']
        reso = flow_model.resolution
        noise = torch.randn(num_samples, flow_model.in_channels, reso, reso, reso).to(self.device)
        sampler_params = {**self.sparse_structure_sampler_params, **sampler_params}
        z_s = self.sparse_structure_sampler.sample(
            flow_model,
            noise,
            **cond,
            **sampler_params,
            verbose=True
        )["samples"]
        
        # Decode occupancy latent
        decoder = self.models['sparse_structure_decoder']
        coords = torch.argwhere(decoder(z_s)>0)[:, [0, 2, 3, 4]].int()

        return coords

    def decode_slat(
        self,
        slat: sp.SparseTensor,
        formats: List[str] = ['mesh',],
    ) -> dict:
        """
        Decode the structured latent.

        Args:
            slat (sp.SparseTensor): The structured latent.
            formats (List[str]): The formats to decode the structured latent to.

        Returns:
            dict: The decoded structured latent.
        """
        ret = {}
        if 'mesh' in formats:
            ret['mesh'] = self.models['slat_decoder_mesh'](slat)
        return ret
    
    def sample_slat(
        self,
        cond: dict,
        coords: torch.Tensor,
        sampler_params: dict = {},
    ) -> sp.SparseTensor:
        """
        Sample structured latent with the given conditioning.
        
        Args:
            cond (dict): The conditioning information.
            coords (torch.Tensor): The coordinates of the sparse structure.
            sampler_params (dict): Additional parameters for the sampler.
        """
        # Sample structured latent
        flow_model = self.models['slat_flow_model']
        noise = sp.SparseTensor(
            feats=torch.randn(coords.shape[0], flow_model.in_channels).to(self.device),
            coords=coords,
        )
        sampler_params = {**self.slat_sampler_params, **sampler_params}
        slat = self.slat_sampler.sample(
            flow_model,
            noise,
            **cond,
            **sampler_params,
            verbose=True
        )["samples"]

        std = torch.tensor(self.slat_normalization['std'])[None].to(slat.device)
        mean = torch.tensor(self.slat_normalization['mean'])[None].to(slat.device)
        slat = slat * std + mean
        
        return slat

    @torch.no_grad()
    def run(
        self,
        image: Image.Image,
        num_samples: int = 1,
        seed: int = 42,
        sparse_structure_sampler_params: dict = {},
        slat_sampler_params: dict = {},
        formats: List[str] = ['mesh',],
        preprocess_image: bool = True,
    ) -> dict:
        """
        Run the pipeline.

        Args:
            image (Image.Image): The image prompt.
            num_samples (int): The number of samples to generate.
            sparse_structure_sampler_params (dict): Additional parameters for the sparse structure sampler.
            slat_sampler_params (dict): Additional parameters for the structured latent sampler.
            preprocess_image (bool): Whether to preprocess the image.
        """
        if preprocess_image:
            image = self.preprocess_image(image)
        cond = self.get_cond([image])
        torch.manual_seed(seed)
        coords = self.sample_sparse_structure(cond, num_samples, sparse_structure_sampler_params)
        slat = self.sample_slat(cond, coords, slat_sampler_params)
        return self.decode_slat(slat, formats)

    @contextmanager
    def inject_sampler_multi_image(
        self,
        sampler_name: str,
        num_images: int,
        num_steps: int,
        mode: Literal['stochastic', 'multidiffusion'] = 'stochastic',
    ):
        """
        Inject a sampler with multiple images as condition.
        
        Args:
            sampler_name (str): The name of the sampler to inject.
            num_images (int): The number of images to condition on.
            num_steps (int): The number of steps to run the sampler for.
        """
        sampler = getattr(self, sampler_name)
        setattr(sampler, f'_old_inference_model', sampler._inference_model)

        if mode == 'stochastic':
            if num_images > num_steps:
                print(f"\033[93mWarning: number of conditioning images is greater than number of steps for {sampler_name}. "
                    "This may lead to performance degradation.\033[0m")

            cond_indices = (np.arange(num_steps) % num_images).tolist()
            def _new_inference_model(self, model, x_t, t, cond, **kwargs):
                cond_idx = cond_indices.pop(0)
                cond_i = cond[cond_idx:cond_idx+1]
                return self._old_inference_model(model, x_t, t, cond=cond_i, **kwargs)
        
        elif mode =='multidiffusion':
            from .samplers import FlowEulerSampler
            def _new_inference_model(self, model, x_t, t, cond, neg_cond, cfg_strength, cfg_interval, **kwargs):
                if cfg_interval[0] <= t <= cfg_interval[1]:
                    preds = []
                    for i in range(len(cond)):
                        preds.append(FlowEulerSampler._inference_model(self, model, x_t, t, cond[i:i+1], **kwargs))
                    pred = sum(preds) / len(preds)
                    neg_pred = FlowEulerSampler._inference_model(self, model, x_t, t, neg_cond, **kwargs)
                    return (1 + cfg_strength) * pred - cfg_strength * neg_pred
                else:
                    preds = []
                    for i in range(len(cond)):
                        preds.append(FlowEulerSampler._inference_model(self, model, x_t, t, cond[i:i+1], **kwargs))
                    pred = sum(preds) / len(preds)
                    return pred
            
        else:
            raise ValueError(f"Unsupported mode: {mode}")
            
        sampler._inference_model = _new_inference_model.__get__(sampler, type(sampler))

        yield

        sampler._inference_model = sampler._old_inference_model
        delattr(sampler, f'_old_inference_model')

    @torch.no_grad()
    def run_multi_image(
        self,
        images: List[Image.Image],
        num_samples: int = 1,
        seed: int = 42,
        sparse_structure_sampler_params: dict = {},
        slat_sampler_params: dict = {},
        formats: List[str] = ['mesh', 'radiance_field'],
        preprocess_image: bool = True,
        mode: Literal['stochastic', 'multidiffusion'] = 'stochastic',
    ) -> dict:
        """
        Run the pipeline with multiple images as condition

        Args:
            images (List[Image.Image]): The multi-view images of the assets
            num_samples (int): The number of samples to generate.
            sparse_structure_sampler_params (dict): Additional parameters for the sparse structure sampler.
            slat_sampler_params (dict): Additional parameters for the structured latent sampler.
            preprocess_image (bool): Whether to preprocess the image.
        """
        if preprocess_image:
            images = [self.preprocess_image(image) for image in images]
        cond = self.get_cond(images)
        cond['neg_cond'] = cond['neg_cond'][:1]
        torch.manual_seed(seed)
        ss_steps = {**self.sparse_structure_sampler_params, **sparse_structure_sampler_params}.get('steps')
        with self.inject_sampler_multi_image('sparse_structure_sampler', len(images), ss_steps, mode=mode):
            coords = self.sample_sparse_structure(cond, num_samples, sparse_structure_sampler_params)
        slat_steps = {**self.slat_sampler_params, **slat_sampler_params}.get('steps')
        with self.inject_sampler_multi_image('slat_sampler', len(images), slat_steps, mode=mode):
            slat = self.sample_slat(cond, coords, slat_sampler_params)
        return self.decode_slat(slat, formats)