Upload

app.py

@@ -0,0 +1,152 @@
+import gradio as gr
+# from gradio_litmodel3d import LitModel3D
+
+import os
+from typing import *
+import imageio
+import uuid
+from PIL import Image
+from trellis.pipelines import TrellisImageTo3DPipeline
+from trellis.utils import render_utils, postprocessing_utils
+
+
+def preprocess_image(image: Image.Image) -> Image.Image:
+    """
+    Preprocess the input image.
+
+    Args:
+        image (Image.Image): The input image.
+
+    Returns:
+        Image.Image: The preprocessed image.
+    """
+    return pipeline.preprocess_image(image)
+
+
+def image_to_3d(image: Image.Image) -> Tuple[dict, str]:
+    """
+    Convert an image to a 3D model.
+
+    Args:
+        image (Image.Image): The input image.
+
+    Returns:
+        dict: The information of the generated 3D model.
+        str: The path to the video of the 3D model.
+    """
+    outputs = pipeline(image, formats=["gaussian", "mesh"], preprocess_image=False)
+    video = render_utils.render_video(outputs['gaussian'][0])['color']
+    model_id = uuid.uuid4()
+    video_path = f"/tmp/Trellis-demo/{model_id}.mp4"
+    os.makedirs(os.path.dirname(video_path), exist_ok=True)
+    imageio.mimsave(video_path, video, fps=30)
+    model = {'gaussian': outputs['gaussian'][0], 'mesh': outputs['mesh'][0], 'model_id': model_id}
+    return model, video_path
+
+
+def extract_glb(model: dict, mesh_simplify: float, texture_size: int) -> Tuple[str, str]:
+    """
+    Extract a GLB file from the 3D model.
+
+    Args:
+        model (dict): The generated 3D model.
+        mesh_simplify (float): The mesh simplification factor.
+        texture_size (int): The texture resolution.
+
+    Returns:
+        str: The path to the extracted GLB file.
+    """
+    glb = postprocessing_utils.to_glb(model['gaussian'], model['mesh'], simplify=mesh_simplify, texture_size=texture_size)
+    glb_path = f"/tmp/Trellis-demo/{model['model_id']}.glb"
+    glb.export(glb_path)
+    return glb_path, glb_path
+
+
+def activate_button() -> gr.Button:
+    return gr.Button(interactive=True)
+
+
+def deactivate_button() -> gr.Button:
+    return gr.Button(interactive=False)
+
+
+with gr.Blocks() as demo:
+    with gr.Row():
+        with gr.Column():
+            image_prompt = gr.Image(label="Image Prompt", image_mode="RGBA", type="pil", height=300)
+            generate_btn = gr.Button("Generate", interactive=False)
+
+            mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)
+            texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
+            extract_glb_btn = gr.Button("Extract GLB", interactive=False)
+
+        with gr.Column():
+            video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
+            model_output = gr.Model3D(label="Extracted GLB", height=300)
+            download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
+
+    # Example images at the bottom of the page
+    with gr.Row():
+        examples = gr.Examples(
+            examples=[
+                f'assets/example_image/{image}'
+                for image in os.listdir("assets/example_image")
+            ],
+            inputs=[image_prompt],
+            fn=lambda image: (preprocess_image(image), gr.Button(interactive=True)),
+            outputs=[image_prompt, generate_btn],
+            run_on_click=True,
+            examples_per_page=64,
+        )
+
+    model = gr.State()
+
+    # Handlers
+    image_prompt.upload(
+        preprocess_image,
+        inputs=[image_prompt],
+        outputs=[image_prompt],
+    ).then(
+        activate_button,
+        outputs=[generate_btn],
+    )
+
+    image_prompt.clear(
+        deactivate_button,
+        outputs=[generate_btn],
+    )
+
+    generate_btn.click(
+        image_to_3d,
+        inputs=[image_prompt],
+        outputs=[model, video_output],
+    ).then(
+        activate_button,
+        outputs=[extract_glb_btn],
+    )
+
+    video_output.clear(
+        deactivate_button,
+        outputs=[extract_glb_btn],
+    )
+
+    extract_glb_btn.click(
+        extract_glb,
+        inputs=[model, mesh_simplify, texture_size],
+        outputs=[model_output, download_glb],
+    ).then(
+        activate_button,
+        outputs=[download_glb],
+    )
+
+    model_output.clear(
+        deactivate_button,
+        outputs=[download_glb],
+    )
+    
+
+# Launch the Gradio app
+if __name__ == "__main__":
+    pipeline = TrellisImageTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-image-large")
+    pipeline.cuda()
+    demo.launch()

requirements.txt

@@ -0,0 +1,28 @@
+--extra-index-url https://download.pytorch.org/whl/cu118
+--find-links https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.4.0_cu121.html
+
+
+torch==2.4.0
+torchvision==0.19.0
+pillow==10.4.0
+imageio==2.36.1
+imageio-ffmpeg==0.5.1
+tqdm==4.67.1
+easydict==1.13
+opencv-python-headless==4.10.0.84
+scipy==1.14.1
+rembg==2.0.60
+onnxruntime==1.20.1
+trimesh==4.5.3
+xatlas==0.0.9
+pyvista==0.44.2
+pymeshfix==0.17.0
+igraph==0.11.8
+git+https://github.com/EasternJournalist/utils3d.git@9a4eb15e4021b67b12c460c7057d642626897ec8
+xformers==0.0.27.post2+cu118
+flash-attn==2.7.0.post2
+kaolin==0.17.0
+spconv-cu118==2.3.6
+transformers==4.46.3
+wheels/nvdiffrast-0.3.3-py3-none-any.whl
+wheels/diff_gaussian_rasterization-0.0.0-cp310-cp310-linux_x86_64.whl

trellis/__init__.py

@@ -0,0 +1,6 @@
+from . import models
+from . import modules
+from . import pipelines
+from . import renderers
+from . import representations
+from . import utils

trellis/models/__init__.py

@@ -0,0 +1,70 @@
+import importlib
+
+__attributes = {
+    'SparseStructureEncoder': 'sparse_structure_vae',
+    'SparseStructureDecoder': 'sparse_structure_vae',
+    'SparseStructureFlowModel': 'sparse_structure_flow',
+    'SLatEncoder': 'structured_latent_vae',
+    'SLatGaussianDecoder': 'structured_latent_vae',
+    'SLatRadianceFieldDecoder': 'structured_latent_vae',
+    'SLatMeshDecoder': 'structured_latent_vae',
+    'SLatFlowModel': 'structured_latent_flow',
+}
+
+__submodules = []
+
+__all__ = list(__attributes.keys()) + __submodules
+
+def __getattr__(name):
+    if name not in globals():
+        if name in __attributes:
+            module_name = __attributes[name]
+            module = importlib.import_module(f".{module_name}", __name__)
+            globals()[name] = getattr(module, name)
+        elif name in __submodules:
+            module = importlib.import_module(f".{name}", __name__)
+            globals()[name] = module
+        else:
+            raise AttributeError(f"module {__name__} has no attribute {name}")
+    return globals()[name]
+
+
+def from_pretrained(path: str, **kwargs):
+    """
+    Load a model from a pretrained checkpoint.
+
+    Args:
+        path: The path to the checkpoint. Can be either local path or a Hugging Face model name.
+              NOTE: config file and model file should take the name f'{path}.json' and f'{path}.safetensors' respectively.
+        **kwargs: Additional arguments for the model constructor.
+    """
+    import os
+    import json
+    from safetensors.torch import load_file
+    is_local = os.path.exists(f"{path}.json") and os.path.exists(f"{path}.safetensors")
+
+    if is_local:
+        config_file = f"{path}.json"
+        model_file = f"{path}.safetensors"
+    else:
+        from huggingface_hub import hf_hub_download
+        path_parts = path.split('/')
+        repo_id = f'{path_parts[0]}/{path_parts[1]}'
+        model_name = '/'.join(path_parts[2:])
+        config_file = hf_hub_download(repo_id, f"{model_name}.json")
+        model_file = hf_hub_download(repo_id, f"{model_name}.safetensors")
+
+    with open(config_file, 'r') as f:
+        config = json.load(f)
+    model = __getattr__(config['name'])(**config['args'], **kwargs)
+    model.load_state_dict(load_file(model_file))
+
+    return model
+
+
+# For Pylance
+if __name__ == '__main__':
+    from .sparse_structure_vae import SparseStructureEncoder, SparseStructureDecoder
+    from .sparse_structure_flow import SparseStructureFlowModel
+    from .structured_latent_vae import SLatEncoder, SLatGaussianDecoder, SLatRadianceFieldDecoder, SLatMeshDecoder
+    from .structured_latent_flow import SLatFlowModel

trellis/models/sparse_structure_flow.py

@@ -0,0 +1,200 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ..modules.utils import convert_module_to_f16, convert_module_to_f32
+from ..modules.transformer import AbsolutePositionEmbedder, ModulatedTransformerCrossBlock
+from ..modules.spatial import patchify, unpatchify
+
+
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        """
+        Create sinusoidal timestep embeddings.
+
+        Args:
+            t: a 1-D Tensor of N indices, one per batch element.
+                These may be fractional.
+            dim: the dimension of the output.
+            max_period: controls the minimum frequency of the embeddings.
+
+        Returns:
+            an (N, D) Tensor of positional embeddings.
+        """
+        # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
+        half = dim // 2
+        freqs = torch.exp(
+            -np.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=t.device)
+        args = t[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, t):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+class SparseStructureFlowModel(nn.Module):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        model_channels: int,
+        cond_channels: int,
+        out_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        patch_size: int = 2,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        share_mod: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+    ):
+        super().__init__()
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.cond_channels = cond_channels
+        self.out_channels = out_channels
+        self.num_blocks = num_blocks
+        self.num_heads = num_heads or model_channels // num_head_channels
+        self.mlp_ratio = mlp_ratio
+        self.patch_size = patch_size
+        self.pe_mode = pe_mode
+        self.use_fp16 = use_fp16
+        self.use_checkpoint = use_checkpoint
+        self.share_mod = share_mod
+        self.qk_rms_norm = qk_rms_norm
+        self.qk_rms_norm_cross = qk_rms_norm_cross
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.t_embedder = TimestepEmbedder(model_channels)
+        if share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(model_channels, 6 * model_channels, bias=True)
+            )
+
+        if pe_mode == "ape":
+            pos_embedder = AbsolutePositionEmbedder(model_channels, 3)
+            coords = torch.meshgrid(*[torch.arange(res, device=self.device) for res in [resolution // patch_size] * 3], indexing='ij')
+            coords = torch.stack(coords, dim=-1).reshape(-1, 3)
+            pos_emb = pos_embedder(coords)
+            self.register_buffer("pos_emb", pos_emb)
+
+        self.input_layer = nn.Linear(in_channels * patch_size**3, model_channels)
+            
+        self.blocks = nn.ModuleList([
+            ModulatedTransformerCrossBlock(
+                model_channels,
+                cond_channels,
+                num_heads=self.num_heads,
+                mlp_ratio=self.mlp_ratio,
+                attn_mode='full',
+                use_checkpoint=self.use_checkpoint,
+                use_rope=(pe_mode == "rope"),
+                share_mod=share_mod,
+                qk_rms_norm=self.qk_rms_norm,
+                qk_rms_norm_cross=self.qk_rms_norm_cross,
+            )
+            for _ in range(num_blocks)
+        ])
+
+        self.out_layer = nn.Linear(model_channels, out_channels * patch_size**3)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.blocks.apply(convert_module_to_f32)
+
+    def initialize_weights(self) -> None:
+        # Initialize transformer layers:
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+        self.apply(_basic_init)
+
+        # Initialize timestep embedding MLP:
+        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
+
+        # Zero-out adaLN modulation layers in DiT blocks:
+        if self.share_mod:
+            nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
+        else:
+            for block in self.blocks:
+                nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+                nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor, cond: torch.Tensor) -> torch.Tensor:
+        assert [*x.shape] == [x.shape[0], self.in_channels, *[self.resolution] * 3], \
+                f"Input shape mismatch, got {x.shape}, expected {[x.shape[0], self.in_channels, *[self.resolution] * 3]}"
+
+        h = patchify(x, self.patch_size)
+        h = h.view(*h.shape[:2], -1).permute(0, 2, 1).contiguous()
+
+        h = self.input_layer(h)
+        h = h + self.pos_emb[None]
+        t_emb = self.t_embedder(t)
+        if self.share_mod:
+            t_emb = self.adaLN_modulation(t_emb)
+        t_emb = t_emb.type(self.dtype)
+        h = h.type(self.dtype)
+        cond = cond.type(self.dtype)
+        for block in self.blocks:
+            h = block(h, t_emb, cond)
+        h = h.type(x.dtype)
+        h = F.layer_norm(h, h.shape[-1:])
+        h = self.out_layer(h)
+
+        h = h.permute(0, 2, 1).view(h.shape[0], h.shape[2], *[self.resolution // self.patch_size] * 3)
+        h = unpatchify(h, self.patch_size).contiguous()
+
+        return h

trellis/models/sparse_structure_vae.py

@@ -0,0 +1,306 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from ..modules.norm import GroupNorm32, ChannelLayerNorm32
+from ..modules.spatial import pixel_shuffle_3d
+from ..modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
+
+
+def norm_layer(norm_type: str, *args, **kwargs) -> nn.Module:
+    """
+    Return a normalization layer.
+    """
+    if norm_type == "group":
+        return GroupNorm32(32, *args, **kwargs)
+    elif norm_type == "layer":
+        return ChannelLayerNorm32(*args, **kwargs)
+    else:
+        raise ValueError(f"Invalid norm type {norm_type}")
+
+
+class ResBlock3d(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        out_channels: Optional[int] = None,
+        norm_type: Literal["group", "layer"] = "layer",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+
+        self.norm1 = norm_layer(norm_type, channels)
+        self.norm2 = norm_layer(norm_type, self.out_channels)
+        self.conv1 = nn.Conv3d(channels, self.out_channels, 3, padding=1)
+        self.conv2 = zero_module(nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1))
+        self.skip_connection = nn.Conv3d(channels, self.out_channels, 1) if channels != self.out_channels else nn.Identity()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = self.norm1(x)
+        h = F.silu(h)
+        h = self.conv1(h)
+        h = self.norm2(h)
+        h = F.silu(h)
+        h = self.conv2(h)
+        h = h + self.skip_connection(x)
+        return h
+
+
+class DownsampleBlock3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        mode: Literal["conv", "avgpool"] = "conv",
+    ):
+        assert mode in ["conv", "avgpool"], f"Invalid mode {mode}"
+
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        if mode == "conv":
+            self.conv = nn.Conv3d(in_channels, out_channels, 2, stride=2)
+        elif mode == "avgpool":
+            assert in_channels == out_channels, "Pooling mode requires in_channels to be equal to out_channels"
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if hasattr(self, "conv"):
+            return self.conv(x)
+        else:
+            return F.avg_pool3d(x, 2)
+
+
+class UpsampleBlock3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        mode: Literal["conv", "nearest"] = "conv",
+    ):
+        assert mode in ["conv", "nearest"], f"Invalid mode {mode}"
+
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        if mode == "conv":
+            self.conv = nn.Conv3d(in_channels, out_channels*8, 3, padding=1)
+        elif mode == "nearest":
+            assert in_channels == out_channels, "Nearest mode requires in_channels to be equal to out_channels"
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if hasattr(self, "conv"):
+            x = self.conv(x)
+            return pixel_shuffle_3d(x, 2)
+        else:
+            return F.interpolate(x, scale_factor=2, mode="nearest")
+        
+
+class SparseStructureEncoder(nn.Module):
+    """
+    Encoder for Sparse Structure (\mathcal{E}_S in the paper Sec. 3.3).
+    
+    Args:
+        in_channels (int): Channels of the input.
+        latent_channels (int): Channels of the latent representation.
+        num_res_blocks (int): Number of residual blocks at each resolution.
+        channels (List[int]): Channels of the encoder blocks.
+        num_res_blocks_middle (int): Number of residual blocks in the middle.
+        norm_type (Literal["group", "layer"]): Type of normalization layer.
+        use_fp16 (bool): Whether to use FP16.
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        latent_channels: int,
+        num_res_blocks: int,
+        channels: List[int],
+        num_res_blocks_middle: int = 2,
+        norm_type: Literal["group", "layer"] = "layer",
+        use_fp16: bool = False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.latent_channels = latent_channels
+        self.num_res_blocks = num_res_blocks
+        self.channels = channels
+        self.num_res_blocks_middle = num_res_blocks_middle
+        self.norm_type = norm_type
+        self.use_fp16 = use_fp16
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.input_layer = nn.Conv3d(in_channels, channels[0], 3, padding=1)
+
+        self.blocks = nn.ModuleList([])
+        for i, ch in enumerate(channels):
+            self.blocks.extend([
+                ResBlock3d(ch, ch)
+                for _ in range(num_res_blocks)
+            ])
+            if i < len(channels) - 1:
+                self.blocks.append(
+                    DownsampleBlock3d(ch, channels[i+1])
+                )
+        
+        self.middle_block = nn.Sequential(*[
+            ResBlock3d(channels[-1], channels[-1])
+            for _ in range(num_res_blocks_middle)
+        ])
+
+        self.out_layer = nn.Sequential(
+            norm_layer(norm_type, channels[-1]),
+            nn.SiLU(),
+            nn.Conv3d(channels[-1], latent_channels*2, 3, padding=1)
+        )
+
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.use_fp16 = True
+        self.dtype = torch.float16
+        self.blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.use_fp16 = False
+        self.dtype = torch.float32
+        self.blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+
+    def forward(self, x: torch.Tensor, sample_posterior: bool = False, return_raw: bool = False) -> torch.Tensor:
+        h = self.input_layer(x)
+        h = h.type(self.dtype)
+
+        for block in self.blocks:
+            h = block(h)
+        h = self.middle_block(h)
+
+        h = h.type(x.dtype)
+        h = self.out_layer(h)
+
+        mean, logvar = h.chunk(2, dim=1)
+
+        if sample_posterior:
+            std = torch.exp(0.5 * logvar)
+            z = mean + std * torch.randn_like(std)
+        else:
+            z = mean
+            
+        if return_raw:
+            return z, mean, logvar
+        return z
+        
+
+class SparseStructureDecoder(nn.Module):
+    """
+    Decoder for Sparse Structure (\mathcal{D}_S in the paper Sec. 3.3).
+    
+    Args:
+        out_channels (int): Channels of the output.
+        latent_channels (int): Channels of the latent representation.
+        num_res_blocks (int): Number of residual blocks at each resolution.
+        channels (List[int]): Channels of the decoder blocks.
+        num_res_blocks_middle (int): Number of residual blocks in the middle.
+        norm_type (Literal["group", "layer"]): Type of normalization layer.
+        use_fp16 (bool): Whether to use FP16.
+    """ 
+    def __init__(
+        self,
+        out_channels: int,
+        latent_channels: int,
+        num_res_blocks: int,
+        channels: List[int],
+        num_res_blocks_middle: int = 2,
+        norm_type: Literal["group", "layer"] = "layer",
+        use_fp16: bool = False,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.latent_channels = latent_channels
+        self.num_res_blocks = num_res_blocks
+        self.channels = channels
+        self.num_res_blocks_middle = num_res_blocks_middle
+        self.norm_type = norm_type
+        self.use_fp16 = use_fp16
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.input_layer = nn.Conv3d(latent_channels, channels[0], 3, padding=1)
+
+        self.middle_block = nn.Sequential(*[
+            ResBlock3d(channels[0], channels[0])
+            for _ in range(num_res_blocks_middle)
+        ])
+
+        self.blocks = nn.ModuleList([])
+        for i, ch in enumerate(channels):
+            self.blocks.extend([
+                ResBlock3d(ch, ch)
+                for _ in range(num_res_blocks)
+            ])
+            if i < len(channels) - 1:
+                self.blocks.append(
+                    UpsampleBlock3d(ch, channels[i+1])
+                )
+
+        self.out_layer = nn.Sequential(
+            norm_layer(norm_type, channels[-1]),
+            nn.SiLU(),
+            nn.Conv3d(channels[-1], out_channels, 3, padding=1)
+        )
+
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+    
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.use_fp16 = True
+        self.dtype = torch.float16
+        self.blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.use_fp16 = False
+        self.dtype = torch.float32
+        self.blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = self.input_layer(x)
+        
+        h = h.type(self.dtype)
+                
+        h = self.middle_block(h)
+        for block in self.blocks:
+            h = block(h)
+
+        h = h.type(x.dtype)
+        h = self.out_layer(h)
+        return h

trellis/models/structured_latent_flow.py

@@ -0,0 +1,262 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ..modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
+from ..modules.transformer import AbsolutePositionEmbedder
+from ..modules.norm import LayerNorm32
+from ..modules import sparse as sp
+from ..modules.sparse.transformer import ModulatedSparseTransformerCrossBlock
+from .sparse_structure_flow import TimestepEmbedder
+
+
+class SparseResBlock3d(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        emb_channels: int,
+        out_channels: Optional[int] = None,
+        downsample: bool = False,
+        upsample: bool = False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.out_channels = out_channels or channels
+        self.downsample = downsample
+        self.upsample = upsample
+        
+        assert not (downsample and upsample), "Cannot downsample and upsample at the same time"
+
+        self.norm1 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
+        self.norm2 = LayerNorm32(self.out_channels, elementwise_affine=False, eps=1e-6)
+        self.conv1 = sp.SparseConv3d(channels, self.out_channels, 3)
+        self.conv2 = zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3))
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(emb_channels, 2 * self.out_channels, bias=True),
+        )
+        self.skip_connection = sp.SparseLinear(channels, self.out_channels) if channels != self.out_channels else nn.Identity()
+        self.updown = None
+        if self.downsample:
+            self.updown = sp.SparseDownsample(2)
+        elif self.upsample:
+            self.updown = sp.SparseUpsample(2)
+
+    def _updown(self, x: sp.SparseTensor) -> sp.SparseTensor:
+        if self.updown is not None:
+            x = self.updown(x)
+        return x
+
+    def forward(self, x: sp.SparseTensor, emb: torch.Tensor) -> sp.SparseTensor:
+        emb_out = self.emb_layers(emb).type(x.dtype)
+        scale, shift = torch.chunk(emb_out, 2, dim=1)
+
+        x = self._updown(x)
+        h = x.replace(self.norm1(x.feats))
+        h = h.replace(F.silu(h.feats))
+        h = self.conv1(h)
+        h = h.replace(self.norm2(h.feats)) * (1 + scale) + shift
+        h = h.replace(F.silu(h.feats))
+        h = self.conv2(h)
+        h = h + self.skip_connection(x)
+
+        return h
+    
+
+class SLatFlowModel(nn.Module):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        model_channels: int,
+        cond_channels: int,
+        out_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        patch_size: int = 2,
+        num_io_res_blocks: int = 2,
+        io_block_channels: List[int] = None,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        use_skip_connection: bool = True,
+        share_mod: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+    ):
+        super().__init__()
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.cond_channels = cond_channels
+        self.out_channels = out_channels
+        self.num_blocks = num_blocks
+        self.num_heads = num_heads or model_channels // num_head_channels
+        self.mlp_ratio = mlp_ratio
+        self.patch_size = patch_size
+        self.num_io_res_blocks = num_io_res_blocks
+        self.io_block_channels = io_block_channels
+        self.pe_mode = pe_mode
+        self.use_fp16 = use_fp16
+        self.use_checkpoint = use_checkpoint
+        self.use_skip_connection = use_skip_connection
+        self.share_mod = share_mod
+        self.qk_rms_norm = qk_rms_norm
+        self.qk_rms_norm_cross = qk_rms_norm_cross
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        assert int(np.log2(patch_size)) == np.log2(patch_size), "Patch size must be a power of 2"
+        assert np.log2(patch_size) == len(io_block_channels), "Number of IO ResBlocks must match the number of stages"
+
+        self.t_embedder = TimestepEmbedder(model_channels)
+        if share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(model_channels, 6 * model_channels, bias=True)
+            )
+
+        if pe_mode == "ape":
+            self.pos_embedder = AbsolutePositionEmbedder(model_channels)
+
+        self.input_layer = sp.SparseLinear(in_channels, io_block_channels[0])
+        self.input_blocks = nn.ModuleList([])
+        for chs, next_chs in zip(io_block_channels, io_block_channels[1:] + [model_channels]):
+            self.input_blocks.extend([
+                SparseResBlock3d(
+                    chs,
+                    model_channels,
+                    out_channels=chs,
+                )
+                for _ in range(num_io_res_blocks-1)
+            ])
+            self.input_blocks.append(
+                SparseResBlock3d(
+                    chs,
+                    model_channels,
+                    out_channels=next_chs,
+                    downsample=True,
+                )
+            )
+            
+        self.blocks = nn.ModuleList([
+            ModulatedSparseTransformerCrossBlock(
+                model_channels,
+                cond_channels,
+                num_heads=self.num_heads,
+                mlp_ratio=self.mlp_ratio,
+                attn_mode='full',
+                use_checkpoint=self.use_checkpoint,
+                use_rope=(pe_mode == "rope"),
+                share_mod=self.share_mod,
+                qk_rms_norm=self.qk_rms_norm,
+                qk_rms_norm_cross=self.qk_rms_norm_cross,
+            )
+            for _ in range(num_blocks)
+        ])
+
+        self.out_blocks = nn.ModuleList([])
+        for chs, prev_chs in zip(reversed(io_block_channels), [model_channels] + list(reversed(io_block_channels[1:]))):
+            self.out_blocks.append(
+                SparseResBlock3d(
+                    prev_chs * 2 if self.use_skip_connection else prev_chs,
+                    model_channels,
+                    out_channels=chs,
+                    upsample=True,
+                )
+            )
+            self.out_blocks.extend([
+                SparseResBlock3d(
+                    chs * 2 if self.use_skip_connection else chs,
+                    model_channels,
+                    out_channels=chs,
+                )
+                for _ in range(num_io_res_blocks-1)
+            ])
+        self.out_layer = sp.SparseLinear(io_block_channels[0], out_channels)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.blocks.apply(convert_module_to_f16)
+        self.out_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.blocks.apply(convert_module_to_f32)
+        self.out_blocks.apply(convert_module_to_f32)
+
+    def initialize_weights(self) -> None:
+        # Initialize transformer layers:
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+        self.apply(_basic_init)
+
+        # Initialize timestep embedding MLP:
+        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
+
+        # Zero-out adaLN modulation layers in DiT blocks:
+        if self.share_mod:
+            nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
+        else:
+            for block in self.blocks:
+                nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+                nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def forward(self, x: sp.SparseTensor, t: torch.Tensor, cond: torch.Tensor) -> sp.SparseTensor:
+        h = self.input_layer(x).type(self.dtype)
+        t_emb = self.t_embedder(t)
+        if self.share_mod:
+            t_emb = self.adaLN_modulation(t_emb)
+        t_emb = t_emb.type(self.dtype)
+        cond = cond.type(self.dtype)
+
+        skips = []
+        # pack with input blocks
+        for block in self.input_blocks:
+            h = block(h, t_emb)
+            skips.append(h.feats)
+        
+        if self.pe_mode == "ape":
+            h = h + self.pos_embedder(h.coords[:, 1:]).type(self.dtype)
+        for block in self.blocks:
+            h = block(h, t_emb, cond)
+
+        # unpack with output blocks
+        for block, skip in zip(self.out_blocks, reversed(skips)):
+            if self.use_skip_connection:
+                h = block(h.replace(torch.cat([h.feats, skip], dim=1)), t_emb)
+            else:
+                h = block(h, t_emb)
+
+        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
+        h = self.out_layer(h.type(x.dtype))
+        return h

trellis/models/structured_latent_vae/__init__.py

@@ -0,0 +1,4 @@
+from .encoder import SLatEncoder
+from .decoder_gs import SLatGaussianDecoder
+from .decoder_rf import SLatRadianceFieldDecoder
+from .decoder_mesh import SLatMeshDecoder

trellis/models/structured_latent_vae/base.py

@@ -0,0 +1,117 @@
+from typing import *
+import torch
+import torch.nn as nn
+from ...modules.utils import convert_module_to_f16, convert_module_to_f32
+from ...modules import sparse as sp
+from ...modules.transformer import AbsolutePositionEmbedder
+from ...modules.sparse.transformer import SparseTransformerBlock
+
+
+def block_attn_config(self):
+    """
+    Return the attention configuration of the model.
+    """
+    for i in range(self.num_blocks):
+        if self.attn_mode == "shift_window":
+            yield "serialized", self.window_size, 0, (16 * (i % 2),) * 3, sp.SerializeMode.Z_ORDER
+        elif self.attn_mode == "shift_sequence":
+            yield "serialized", self.window_size, self.window_size // 2 * (i % 2), (0, 0, 0), sp.SerializeMode.Z_ORDER
+        elif self.attn_mode == "shift_order":
+            yield "serialized", self.window_size, 0, (0, 0, 0), sp.SerializeModes[i % 4]
+        elif self.attn_mode == "full":
+            yield "full", None, None, None, None
+        elif self.attn_mode == "swin":
+            yield "windowed", self.window_size, None, self.window_size // 2 * (i % 2), None
+
+
+class SparseTransformerBase(nn.Module):
+    """
+    Sparse Transformer without output layers.
+    Serve as the base class for encoder and decoder.
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        model_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        window_size: Optional[int] = None,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.num_blocks = num_blocks
+        self.window_size = window_size
+        self.num_heads = num_heads or model_channels // num_head_channels
+        self.mlp_ratio = mlp_ratio
+        self.attn_mode = attn_mode
+        self.pe_mode = pe_mode
+        self.use_fp16 = use_fp16
+        self.use_checkpoint = use_checkpoint
+        self.qk_rms_norm = qk_rms_norm
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        if pe_mode == "ape":
+            self.pos_embedder = AbsolutePositionEmbedder(model_channels)
+
+        self.input_layer = sp.SparseLinear(in_channels, model_channels)
+        self.blocks = nn.ModuleList([
+            SparseTransformerBlock(
+                model_channels,
+                num_heads=self.num_heads,
+                mlp_ratio=self.mlp_ratio,
+                attn_mode=attn_mode,
+                window_size=window_size,
+                shift_sequence=shift_sequence,
+                shift_window=shift_window,
+                serialize_mode=serialize_mode,
+                use_checkpoint=self.use_checkpoint,
+                use_rope=(pe_mode == "rope"),
+                qk_rms_norm=self.qk_rms_norm,
+            )
+            for attn_mode, window_size, shift_sequence, shift_window, serialize_mode in block_attn_config(self)
+        ])
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.blocks.apply(convert_module_to_f32)
+
+    def initialize_weights(self) -> None:
+        # Initialize transformer layers:
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+        self.apply(_basic_init)
+
+    def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
+        h = self.input_layer(x)
+        if self.pe_mode == "ape":
+            h = h + self.pos_embedder(x.coords[:, 1:])
+        h = h.type(self.dtype)
+        for block in self.blocks:
+            h = block(h)
+        return h