apps/mv_models.py

import gradio as gr
import numpy as np
import torch
import PIL
from PIL import Image
import os
import sys
import rembg
import time
import json
import cv2
from datetime import datetime
from einops import repeat, rearrange
from omegaconf import OmegaConf
from typing import Dict, Optional, Tuple, List
from dataclasses import dataclass
from .utils import *
from huggingface_hub import hf_hub_download
import spaces

parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))


class GenMVImage(object):
    def __init__(self, device):
        self.seed = 1024
        self.guidance_scale = 7.5
        self.step = 50
        self.device = device

        from .third_party.CRM.pipelines import TwoStagePipeline
        stage1_config = OmegaConf.load(f"{parent_dir}/apps/third_party/CRM/configs/nf7_v3_SNR_rd_size_stroke.yaml").config
        stage1_sampler_config = stage1_config.sampler
        stage1_model_config = stage1_config.models
        stage1_model_config.resume = hf_hub_download(repo_id="Zhengyi/CRM", filename="pixel-diffusion.pth", repo_type="model")
        stage1_model_config.config = f"{parent_dir}/apps/third_party/CRM/" + stage1_model_config.config
        self.crm_pipeline = TwoStagePipeline(
                                    stage1_model_config,
                                    stage1_sampler_config,
                                    device=self.device,
                                    dtype=torch.float16
                                )
        self.crm_pipeline.set_seed(self.seed)

        sys.path.append(f"{parent_dir}/apps/third_party/Wonder3D")
        from diffusers import DiffusionPipeline  # only tested on diffusers[torch]==0.19.3, may have conflicts with newer versions of diffusers
        self.wonder3d_pipeline =  DiffusionPipeline.from_pretrained(
            'flamehaze1115/wonder3d-v1.0', # or use local checkpoint './ckpts'
            custom_pipeline='flamehaze1115/wonder3d-pipeline',
            torch_dtype=torch.float16
        )
        self.wonder3d_pipeline.unet.enable_xformers_memory_efficient_attention()
        self.wonder3d_pipeline.to(self.device)
        self.wonder3d_pipeline.set_progress_bar_config(disable=True)


        sys.path.append(f"{parent_dir}/apps/third_party/mvdream_diffusers")
        from .third_party.mvdream_diffusers.pipeline_mvdream import MVDreamPipeline
        self.mvdream_pipeline = MVDreamPipeline.from_pretrained(
            "ashawkey/mvdream-sd2.1-diffusers", # remote weights
            torch_dtype=torch.float16,
            trust_remote_code=True,
        )
        self.mvdream_pipeline = self.mvdream_pipeline.to(self.device)
        # self.imagedream_pipeline = MVDreamPipeline.from_pretrained(
        #         "ashawkey/imagedream-ipmv-diffusers", # remote weights
        #         torch_dtype=torch.float16,
        #         trust_remote_code=True,
        #     )
        # self.imagedream_pipeline = self.imagedream_pipeline.to(self.device)


    @spaces.GPU
    def gen_image_from_crm(self, image):
        rt_dict = self.crm_pipeline(
            image, 
            scale=self.guidance_scale, 
            step=self.step
        )
        mv_imgs = rt_dict["stage1_images"]
        return mv_imgs[5], mv_imgs[3], mv_imgs[2], mv_imgs[0]
    
    @spaces.GPU
    def gen_image_from_mvdream(self, image, text):
        if image is None:
            mv_imgs = self.mvdream_pipeline(
                text,
                negative_prompt="ugly, deformed, disfigured, poor details, bad anatomy",
                num_inference_steps=self.step,
                guidance_scale=self.guidance_scale,
                generator = torch.Generator(self.device).manual_seed(self.seed)
            )
        elif text is not None:
            image = np.array(image)
            image = image.astype(np.float32) / 255.0
            image = image[..., :3] * image[..., 3:4] + (1 - image[..., 3:4])

            mv_imgs = self.imagedream_pipeline(
                text,
                image,
                negative_prompt="ugly, deformed, disfigured, poor details, bad anatomy",
                num_inference_steps=self.step,
                guidance_scale=self.guidance_scale,
                generator = torch.Generator(self.device).manual_seed(self.seed)
            )
        return mv_imgs[1], mv_imgs[2], mv_imgs[3], mv_imgs[0]
    
    @spaces.GPU
    def gen_image_from_wonder3d(self, image, crop_size):

        weight_dtype = torch.float16
        batch = prepare_data(image, crop_size)

        generator = torch.Generator(device=self.wonder3d_pipeline.unet.device).manual_seed(self.seed)
        # repeat  (2B, Nv, 3, H, W)
        imgs_in = torch.cat([batch['imgs_in']] * 2, dim=0).to(weight_dtype)

        # (2B, Nv, Nce)
        camera_embeddings = torch.cat([batch['camera_embeddings']] * 2, dim=0).to(weight_dtype)

        task_embeddings = torch.cat([batch['normal_task_embeddings'], batch['color_task_embeddings']], dim=0).to(weight_dtype)

        camera_embeddings = torch.cat([camera_embeddings, task_embeddings], dim=-1).to(weight_dtype)

        # (B*Nv, 3, H, W)
        imgs_in = rearrange(imgs_in, "Nv C H W -> (Nv) C H W")
        # (B*Nv, Nce)

        out = self.wonder3d_pipeline(
            imgs_in,
            # camera_embeddings,
            generator=generator,
            guidance_scale=self.guidance_scale,
            num_inference_steps=self.step,
            output_type='pt',
            num_images_per_prompt=1,
            **{'eta': 1.0},
        ).images

        bsz = out.shape[0] // 2
        normals_pred = out[:bsz]
        images_pred = out[bsz:]

        normals_pred = [save_image(normals_pred[i]) for i in range(bsz)]
        images_pred = [save_image(images_pred[i]) for i in range(bsz)]

        mv_imgs = images_pred
        return mv_imgs[0], mv_imgs[2], mv_imgs[4], mv_imgs[5]
        
    @spaces.GPU
    def run(self, mvimg_model, text, image, crop_size, seed, guidance_scale, step):
        self.seed = seed
        self.guidance_scale = guidance_scale
        self.step = step
        if mvimg_model.upper() == "CRM":
            return self.gen_image_from_crm(image)
        elif mvimg_model.upper() == "IMAGEDREAM":
            return self.gen_image_from_mvdream(image, None)
        elif mvimg_model.upper() == "WONDER3D":
            return self.gen_image_from_wonder3d(image, crop_size)