trellis/utils/render_utils.py

import torch
import numpy as np
from tqdm import tqdm
import utils3d
from PIL import Image

from ..renderers import OctreeRenderer, GaussianRenderer, MeshRenderer
from ..representations import Octree, Gaussian, MeshExtractResult
from ..modules import sparse as sp
from .random_utils import sphere_hammersley_sequence


def yaw_pitch_r_fov_to_extrinsics_intrinsics(yaws, pitchs, rs, fovs):
    is_list = isinstance(yaws, list)
    if not is_list:
        yaws = [yaws]
        pitchs = [pitchs]
    if not isinstance(rs, list):
        rs = [rs] * len(yaws)
    if not isinstance(fovs, list):
        fovs = [fovs] * len(yaws)
    extrinsics = []
    intrinsics = []
    for yaw, pitch, r, fov in zip(yaws, pitchs, rs, fovs):
        fov = torch.deg2rad(torch.tensor(float(fov))).cuda()
        yaw = torch.tensor(float(yaw)).cuda()
        pitch = torch.tensor(float(pitch)).cuda()
        orig = torch.tensor([
            torch.sin(yaw) * torch.cos(pitch),
            torch.cos(yaw) * torch.cos(pitch),
            torch.sin(pitch),
        ]).cuda() * r
        extr = utils3d.torch.extrinsics_look_at(orig, torch.tensor([0, 0, 0]).float().cuda(), torch.tensor([0, 0, 1]).float().cuda())
        intr = utils3d.torch.intrinsics_from_fov(fov, fov)
        extrinsics.append(extr)
        intrinsics.append(intr)
    if not is_list:
        extrinsics = extrinsics[0]
        intrinsics = intrinsics[0]
    return extrinsics, intrinsics


def render_frames(sample, extrinsics, intrinsics, options={}, colors_overwrite=None, verbose=True, **kwargs):
    if isinstance(sample, Octree):
        renderer = OctreeRenderer()
        renderer.rendering_options.resolution = options.get('resolution', 512)
        renderer.rendering_options.near = options.get('near', 0.8)
        renderer.rendering_options.far = options.get('far', 1.6)
        renderer.rendering_options.bg_color = options.get('bg_color', (0, 0, 0))
        renderer.rendering_options.ssaa = options.get('ssaa', 4)
        renderer.pipe.primitive = sample.primitive
    elif isinstance(sample, Gaussian):
        renderer = GaussianRenderer()
        renderer.rendering_options.resolution = options.get('resolution', 512)
        renderer.rendering_options.near = options.get('near', 0.8)
        renderer.rendering_options.far = options.get('far', 1.6)
        renderer.rendering_options.bg_color = options.get('bg_color', (0, 0, 0))
        renderer.rendering_options.ssaa = options.get('ssaa', 1)
        renderer.pipe.kernel_size = kwargs.get('kernel_size', 0.1)
        renderer.pipe.use_mip_gaussian = True
    elif isinstance(sample, MeshExtractResult):
        renderer = MeshRenderer()
        renderer.rendering_options.resolution = options.get('resolution', 512)
        renderer.rendering_options.near = options.get('near', 1)
        renderer.rendering_options.far = options.get('far', 100)
        renderer.rendering_options.ssaa = options.get('ssaa', 4)
    else:
        raise ValueError(f'Unsupported sample type: {type(sample)}')
    
    rets = {}
    for j, (extr, intr) in tqdm(enumerate(zip(extrinsics, intrinsics)), desc='Rendering', disable=not verbose):
        if not isinstance(sample, MeshExtractResult):
            res = renderer.render(sample, extr, intr, colors_overwrite=colors_overwrite)
            if 'color' not in rets: rets['color'] = []
            if 'depth' not in rets: rets['depth'] = []
            rets['color'].append(np.clip(res['color'].detach().cpu().numpy().transpose(1, 2, 0) * 255, 0, 255).astype(np.uint8))
            if 'percent_depth' in res:
                rets['depth'].append(res['percent_depth'].detach().cpu().numpy())
            elif 'depth' in res:
                rets['depth'].append(res['depth'].detach().cpu().numpy())
            else:
                rets['depth'].append(None)
        else:
            res = renderer.render(sample, extr, intr)
            if 'normal' not in rets: rets['normal'] = []
            rets['normal'].append(np.clip(res['normal'].detach().cpu().numpy().transpose(1, 2, 0) * 255, 0, 255).astype(np.uint8))
    return rets


def render_video(sample, resolution=512, bg_color=(0, 0, 0), num_frames=300, r=2, fov=40, **kwargs):
    # Match the model viewer camera angle: yaw=45deg, pitch=75deg
    # Convert degrees to radians for the camera calculations
    yaws = torch.linspace(0, 2 * 3.1415, num_frames)  # Full rotation around Y-axis
    pitch = torch.tensor([75.0 * 3.1415 / 180.0] * num_frames)  # Fixed pitch at 75 degrees
    yaws = yaws.tolist()
    pitch = pitch.tolist()
    # Adjust distance to match model viewer's final position (closer view)
    r = 1.5  # Closer than default 2, similar to model viewer's 100% distance
    extrinsics, intrinsics = yaw_pitch_r_fov_to_extrinsics_intrinsics(yaws, pitch, r, fov)
    return render_frames(sample, extrinsics, intrinsics, {'resolution': resolution, 'bg_color': bg_color}, **kwargs)


def render_multiview(sample, resolution=512, nviews=30):
    r = 2
    fov = 40
    cams = [sphere_hammersley_sequence(i, nviews) for i in range(nviews)]
    yaws = [cam[0] for cam in cams]
    pitchs = [cam[1] for cam in cams]
    extrinsics, intrinsics = yaw_pitch_r_fov_to_extrinsics_intrinsics(yaws, pitchs, r, fov)
    res = render_frames(sample, extrinsics, intrinsics, {'resolution': resolution, 'bg_color': (0, 0, 0)})
    return res['color'], extrinsics, intrinsics


def render_snapshot(samples, resolution=512, bg_color=(0, 0, 0), offset=(-16 / 180 * np.pi, 20 / 180 * np.pi), r=10, fov=8, **kwargs):
    yaw = [0, np.pi/2, np.pi, 3*np.pi/2]
    yaw_offset = offset[0]
    yaw = [y + yaw_offset for y in yaw]
    pitch = [offset[1] for _ in range(4)]
    extrinsics, intrinsics = yaw_pitch_r_fov_to_extrinsics_intrinsics(yaw, pitch, r, fov)
    return render_frames(samples, extrinsics, intrinsics, {'resolution': resolution, 'bg_color': bg_color}, **kwargs)