scripts/convert_mvdream_to_diffusers.py

# Modified from https://github.com/huggingface/diffusers/blob/bc691231360a4cbc7d19a58742ebb8ed0f05e027/scripts/convert_original_stable_diffusion_to_diffusers.py

import argparse
import torch
import sys

sys.path.insert(0, '../')

from transformers import (
    CLIPImageProcessor,
    CLIPVisionModelWithProjection,
)

from diffusers.models import (
    AutoencoderKL,
    UNet2DConditionModel,
)
from diffusers.schedulers import DDIMScheduler
from diffusers.utils import logging

from accelerate import init_empty_weights
from accelerate.utils import set_module_tensor_to_device
from rich import print, print_json
from models import MultiViewUNetModel, MultiViewUNetWrapperModel
from pipeline_mvdream import MVDreamStableDiffusionPipeline
from transformers import CLIPTokenizer, CLIPTextModel

logger = logging.get_logger(__name__)

# def create_unet_diffusers_config(original_config, image_size: int, controlnet=False):
#     """
#     Creates a config for the diffusers based on the config of the LDM model.
#     """
#     if controlnet:
#         unet_params = original_config.model.params.control_stage_config.params
#     else:
#         if "unet_config" in original_config.model.params and original_config.model.params.unet_config is not None:
#             unet_params = original_config.model.params.unet_config.params
#         else:
#             unet_params = original_config.model.params.network_config.params

#     vae_params = original_config.model.params.first_stage_config.params.ddconfig

#     block_out_channels = [unet_params.model_channels * mult for mult in unet_params.channel_mult]

#     down_block_types = []
#     resolution = 1
#     for i in range(len(block_out_channels)):
#         block_type = "CrossAttnDownBlock2D" if resolution in unet_params.attention_resolutions else "DownBlock2D"
#         down_block_types.append(block_type)
#         if i != len(block_out_channels) - 1:
#             resolution *= 2

#     up_block_types = []
#     for i in range(len(block_out_channels)):
#         block_type = "CrossAttnUpBlock2D" if resolution in unet_params.attention_resolutions else "UpBlock2D"
#         up_block_types.append(block_type)
#         resolution //= 2

#     if unet_params.transformer_depth is not None:
#         transformer_layers_per_block = (
#             unet_params.transformer_depth
#             if isinstance(unet_params.transformer_depth, int)
#             else list(unet_params.transformer_depth)
#         )
#     else:
#         transformer_layers_per_block = 1

#     vae_scale_factor = 2 ** (len(vae_params.ch_mult) - 1)

#     head_dim = unet_params.num_heads if "num_heads" in unet_params else None
#     use_linear_projection = (
#         unet_params.use_linear_in_transformer if "use_linear_in_transformer" in unet_params else False
#     )
#     if use_linear_projection:
#         # stable diffusion 2-base-512 and 2-768
#         if head_dim is None:
#             head_dim_mult = unet_params.model_channels // unet_params.num_head_channels
#             head_dim = [head_dim_mult * c for c in list(unet_params.channel_mult)]

#     class_embed_type = None
#     addition_embed_type = None
#     addition_time_embed_dim = None
#     projection_class_embeddings_input_dim = None
#     context_dim = None

#     if unet_params.context_dim is not None:
#         context_dim = (
#             unet_params.context_dim if isinstance(unet_params.context_dim, int) else unet_params.context_dim[0]
#         )

#     if "num_classes" in unet_params:
#         if unet_params.num_classes == "sequential":
#             if context_dim in [2048, 1280]:
#                 # SDXL
#                 addition_embed_type = "text_time"
#                 addition_time_embed_dim = 256
#             else:
#                 class_embed_type = "projection"
#             assert "adm_in_channels" in unet_params
#             projection_class_embeddings_input_dim = unet_params.adm_in_channels
#         else:
#             raise NotImplementedError(f"Unknown conditional unet num_classes config: {unet_params.num_classes}")

#     config = {
#         "sample_size": image_size // vae_scale_factor,
#         "in_channels": unet_params.in_channels,
#         "down_block_types": tuple(down_block_types),
#         "block_out_channels": tuple(block_out_channels),
#         "layers_per_block": unet_params.num_res_blocks,
#         "cross_attention_dim": context_dim,
#         "attention_head_dim": head_dim,
#         "use_linear_projection": use_linear_projection,
#         "class_embed_type": class_embed_type,
#         "addition_embed_type": addition_embed_type,
#         "addition_time_embed_dim": addition_time_embed_dim,
#         "projection_class_embeddings_input_dim": projection_class_embeddings_input_dim,
#         "transformer_layers_per_block": transformer_layers_per_block,
#     }

#     if controlnet:
#         config["conditioning_channels"] = unet_params.hint_channels
#     else:
#         config["out_channels"] = unet_params.out_channels
#         config["up_block_types"] = tuple(up_block_types)

#     return config


def assign_to_checkpoint(paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None):
    """
    This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits
    attention layers, and takes into account additional replacements that may arise.
    Assigns the weights to the new checkpoint.
    """
    assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."

    # Splits the attention layers into three variables.
    if attention_paths_to_split is not None:
        for path, path_map in attention_paths_to_split.items():
            old_tensor = old_checkpoint[path]
            channels = old_tensor.shape[0] // 3

            target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)

            assert config is not None
            num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3

            old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])
            query, key, value = old_tensor.split(channels // num_heads, dim=1)

            checkpoint[path_map["query"]] = query.reshape(target_shape)
            checkpoint[path_map["key"]] = key.reshape(target_shape)
            checkpoint[path_map["value"]] = value.reshape(target_shape)

    for path in paths:
        new_path = path["new"]

        # These have already been assigned
        if attention_paths_to_split is not None and new_path in attention_paths_to_split:
            continue

        # Global renaming happens here
        new_path = new_path.replace("middle_block.0", "mid_block.resnets.0")
        new_path = new_path.replace("middle_block.1", "mid_block.attentions.0")
        new_path = new_path.replace("middle_block.2", "mid_block.resnets.1")

        if additional_replacements is not None:
            for replacement in additional_replacements:
                new_path = new_path.replace(replacement["old"], replacement["new"])

        # proj_attn.weight has to be converted from conv 1D to linear
        is_attn_weight = "proj_attn.weight" in new_path or ("attentions" in new_path and "to_" in new_path)
        shape = old_checkpoint[path["old"]].shape
        if is_attn_weight and len(shape) == 3:
            checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0]
        elif is_attn_weight and len(shape) == 4:
            checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0, 0]
        else:
            checkpoint[new_path] = old_checkpoint[path["old"]]


def shave_segments(path, n_shave_prefix_segments=1):
    """
    Removes segments. Positive values shave the first segments, negative shave the last segments.
    """
    if n_shave_prefix_segments >= 0:
        return ".".join(path.split(".")[n_shave_prefix_segments:])
    else:
        return ".".join(path.split(".")[:n_shave_prefix_segments])


def renew_resnet_paths(old_list, n_shave_prefix_segments=0):
    """
    Updates paths inside resnets to the new naming scheme (local renaming)
    """
    mapping = []
    for old_item in old_list:
        new_item = old_item.replace("in_layers.0", "norm1")
        new_item = new_item.replace("in_layers.2", "conv1")

        new_item = new_item.replace("out_layers.0", "norm2")
        new_item = new_item.replace("out_layers.3", "conv2")

        new_item = new_item.replace("emb_layers.1", "time_emb_proj")
        new_item = new_item.replace("skip_connection", "conv_shortcut")

        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)

        mapping.append({"old": old_item, "new": new_item})

    return mapping


def renew_attention_paths(old_list, n_shave_prefix_segments=0):
    """
    Updates paths inside attentions to the new naming scheme (local renaming)
    """
    mapping = []
    for old_item in old_list:
        new_item = old_item

        #         new_item = new_item.replace('norm.weight', 'group_norm.weight')
        #         new_item = new_item.replace('norm.bias', 'group_norm.bias')

        #         new_item = new_item.replace('proj_out.weight', 'proj_attn.weight')
        #         new_item = new_item.replace('proj_out.bias', 'proj_attn.bias')

        #         new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)

        mapping.append({"old": old_item, "new": new_item})

    return mapping


# def convert_ldm_unet_checkpoint(
#     checkpoint, config, path=None, extract_ema=False, controlnet=False, skip_extract_state_dict=False
# ):
#     """
#     Takes a state dict and a config, and returns a converted checkpoint.
#     """

#     if skip_extract_state_dict:
#         unet_state_dict = checkpoint
#     else:
#         # extract state_dict for UNet
#         unet_state_dict = {}
#         keys = list(checkpoint.keys())

#         if controlnet:
#             unet_key = "control_model."
#         else:
#             unet_key = "model.diffusion_model."

#         # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
#         if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema:
#             logger.warning(f"Checkpoint {path} has both EMA and non-EMA weights.")
#             logger.warning(
#                 "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
#                 " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
#             )
#             for key in keys:
#                 if key.startswith("model.diffusion_model"):
#                     flat_ema_key = "model_ema." + "".join(key.split(".")[1:])
#                     unet_state_dict[key.replace(unet_key, "")] = checkpoint[flat_ema_key]
#         else:
#             if sum(k.startswith("model_ema") for k in keys) > 100:
#                 logger.warning(
#                     "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
#                     " weights (usually better for inference), please make sure to add the `--extract_ema` flag."
#                 )

#             for key in keys:
#                 if key.startswith(unet_key):
#                     unet_state_dict[key.replace(unet_key, "")] = checkpoint[key]

#     new_checkpoint = {}

#     new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"]
#     new_checkpoint["time_embedding.linear_1.bias"] = unet_state_dict["time_embed.0.bias"]
#     new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"]
#     new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"]

#     if config["class_embed_type"] is None:
#         # No parameters to port
#         ...
#     elif config["class_embed_type"] == "timestep" or config["class_embed_type"] == "projection":
#         new_checkpoint["class_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"]
#         new_checkpoint["class_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"]
#         new_checkpoint["class_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"]
#         new_checkpoint["class_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]
#     else:
#         raise NotImplementedError(f"Not implemented `class_embed_type`: {config['class_embed_type']}")

#     if config["addition_embed_type"] == "text_time":
#         new_checkpoint["add_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"]
#         new_checkpoint["add_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"]
#         new_checkpoint["add_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"]
#         new_checkpoint["add_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]

#     new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
#     new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]

#     if not controlnet:
#         new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"]
#         new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"]
#         new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"]
#         new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"]

#     # Retrieves the keys for the input blocks only
#     num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer})
#     input_blocks = {
#         layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key]
#         for layer_id in range(num_input_blocks)
#     }

#     # Retrieves the keys for the middle blocks only
#     num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer})
#     middle_blocks = {
#         layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key]
#         for layer_id in range(num_middle_blocks)
#     }

#     # Retrieves the keys for the output blocks only
#     num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer})
#     output_blocks = {
#         layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key]
#         for layer_id in range(num_output_blocks)
#     }

#     for i in range(1, num_input_blocks):
#         block_id = (i - 1) // (config["layers_per_block"] + 1)
#         layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)

#         resnets = [
#             key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
#         ]
#         attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]

#         if f"input_blocks.{i}.0.op.weight" in unet_state_dict:
#             new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop(
#                 f"input_blocks.{i}.0.op.weight"
#             )
#             new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(
#                 f"input_blocks.{i}.0.op.bias"
#             )

#         paths = renew_resnet_paths(resnets)
#         meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"}
#         assign_to_checkpoint(
#             paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
#         )

#         if len(attentions):
#             paths = renew_attention_paths(attentions)
#             meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"}
#             assign_to_checkpoint(
#                 paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
#             )

#     resnet_0 = middle_blocks[0]
#     attentions = middle_blocks[1]
#     resnet_1 = middle_blocks[2]

#     resnet_0_paths = renew_resnet_paths(resnet_0)
#     assign_to_checkpoint(resnet_0_paths, new_checkpoint, unet_state_dict, config=config)

#     resnet_1_paths = renew_resnet_paths(resnet_1)
#     assign_to_checkpoint(resnet_1_paths, new_checkpoint, unet_state_dict, config=config)

#     attentions_paths = renew_attention_paths(attentions)
#     meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"}
#     assign_to_checkpoint(
#         attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
#     )

#     for i in range(num_output_blocks):
#         block_id = i // (config["layers_per_block"] + 1)
#         layer_in_block_id = i % (config["layers_per_block"] + 1)
#         output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]]
#         output_block_list = {}

#         for layer in output_block_layers:
#             layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1)
#             if layer_id in output_block_list:
#                 output_block_list[layer_id].append(layer_name)
#             else:
#                 output_block_list[layer_id] = [layer_name]

#         if len(output_block_list) > 1:
#             resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key]
#             attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key]

#             resnet_0_paths = renew_resnet_paths(resnets)
#             paths = renew_resnet_paths(resnets)

#             meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"}
#             assign_to_checkpoint(
#                 paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
#             )

#             output_block_list = {k: sorted(v) for k, v in output_block_list.items()}
#             if ["conv.bias", "conv.weight"] in output_block_list.values():
#                 index = list(output_block_list.values()).index(["conv.bias", "conv.weight"])
#                 new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
#                     f"output_blocks.{i}.{index}.conv.weight"
#                 ]
#                 new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
#                     f"output_blocks.{i}.{index}.conv.bias"
#                 ]

#                 # Clear attentions as they have been attributed above.
#                 if len(attentions) == 2:
#                     attentions = []

#             if len(attentions):
#                 paths = renew_attention_paths(attentions)
#                 meta_path = {
#                     "old": f"output_blocks.{i}.1",
#                     "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}",
#                 }
#                 assign_to_checkpoint(
#                     paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
#                 )
#         else:
#             resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1)
#             for path in resnet_0_paths:
#                 old_path = ".".join(["output_blocks", str(i), path["old"]])
#                 new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]])

#                 new_checkpoint[new_path] = unet_state_dict[old_path]

#     if controlnet:
#         # conditioning embedding

#         orig_index = 0

#         new_checkpoint["controlnet_cond_embedding.conv_in.weight"] = unet_state_dict.pop(
#             f"input_hint_block.{orig_index}.weight"
#         )
#         new_checkpoint["controlnet_cond_embedding.conv_in.bias"] = unet_state_dict.pop(
#             f"input_hint_block.{orig_index}.bias"
#         )

#         orig_index += 2

#         diffusers_index = 0

#         while diffusers_index < 6:
#             new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_index}.weight"] = unet_state_dict.pop(
#                 f"input_hint_block.{orig_index}.weight"
#             )
#             new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_index}.bias"] = unet_state_dict.pop(
#                 f"input_hint_block.{orig_index}.bias"
#             )
#             diffusers_index += 1
#             orig_index += 2

#         new_checkpoint["controlnet_cond_embedding.conv_out.weight"] = unet_state_dict.pop(
#             f"input_hint_block.{orig_index}.weight"
#         )
#         new_checkpoint["controlnet_cond_embedding.conv_out.bias"] = unet_state_dict.pop(
#             f"input_hint_block.{orig_index}.bias"
#         )

#         # down blocks
#         for i in range(num_input_blocks):
#             new_checkpoint[f"controlnet_down_blocks.{i}.weight"] = unet_state_dict.pop(f"zero_convs.{i}.0.weight")
#             new_checkpoint[f"controlnet_down_blocks.{i}.bias"] = unet_state_dict.pop(f"zero_convs.{i}.0.bias")

#         # mid block
#         new_checkpoint["controlnet_mid_block.weight"] = unet_state_dict.pop("middle_block_out.0.weight")
#         new_checkpoint["controlnet_mid_block.bias"] = unet_state_dict.pop("middle_block_out.0.bias")

#     return new_checkpoint


def create_vae_diffusers_config(original_config, image_size: int):
    """
    Creates a config for the diffusers based on the config of the LDM model.
    """
    vae_params = original_config.model.params.first_stage_config.params.ddconfig
    _ = original_config.model.params.first_stage_config.params.embed_dim

    block_out_channels = [vae_params.ch * mult for mult in vae_params.ch_mult]
    down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels)
    up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels)

    config = {
        "sample_size": image_size,
        "in_channels": vae_params.in_channels,
        "out_channels": vae_params.out_ch,
        "down_block_types": tuple(down_block_types),
        "up_block_types": tuple(up_block_types),
        "block_out_channels": tuple(block_out_channels),
        "latent_channels": vae_params.z_channels,
        "layers_per_block": vae_params.num_res_blocks,
    }
    return config


def convert_ldm_vae_checkpoint(checkpoint, config):
    # extract state dict for VAE
    vae_state_dict = {}
    vae_key = "first_stage_model."
    keys = list(checkpoint.keys())
    for key in keys:
        if key.startswith(vae_key):
            vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key)

    new_checkpoint = {}

    new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"]
    new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"]
    new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"]
    new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"]
    new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"]
    new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"]

    new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"]
    new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"]
    new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"]
    new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"]
    new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"]
    new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"]

    new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"]
    new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"]
    new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"]
    new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"]

    # Retrieves the keys for the encoder down blocks only
    num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer})
    down_blocks = {layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)}

    # Retrieves the keys for the decoder up blocks only
    num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer})
    up_blocks = {layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)}

    for i in range(num_down_blocks):
        resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]

        if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(f"encoder.down.{i}.downsample.conv.weight")
            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop(f"encoder.down.{i}.downsample.conv.bias")

        paths = renew_vae_resnet_paths(resnets)
        meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"}
        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)

    mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
    num_mid_res_blocks = 2
    for i in range(1, num_mid_res_blocks + 1):
        resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]

        paths = renew_vae_resnet_paths(resnets)
        meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)

    mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
    paths = renew_vae_attention_paths(mid_attentions)
    meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
    assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
    conv_attn_to_linear(new_checkpoint)

    for i in range(num_up_blocks):
        block_id = num_up_blocks - 1 - i
        resnets = [key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key]

        if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[f"decoder.up.{block_id}.upsample.conv.weight"]
            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[f"decoder.up.{block_id}.upsample.conv.bias"]

        paths = renew_vae_resnet_paths(resnets)
        meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"}
        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)

    mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
    num_mid_res_blocks = 2
    for i in range(1, num_mid_res_blocks + 1):
        resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]

        paths = renew_vae_resnet_paths(resnets)
        meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
        assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)

    mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key]
    paths = renew_vae_attention_paths(mid_attentions)
    meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
    assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
    conv_attn_to_linear(new_checkpoint)
    return new_checkpoint


def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0):
    """
    Updates paths inside resnets to the new naming scheme (local renaming)
    """
    mapping = []
    for old_item in old_list:
        new_item = old_item

        new_item = new_item.replace("nin_shortcut", "conv_shortcut")
        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)

        mapping.append({"old": old_item, "new": new_item})

    return mapping


def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0):
    """
    Updates paths inside attentions to the new naming scheme (local renaming)
    """
    mapping = []
    for old_item in old_list:
        new_item = old_item

        new_item = new_item.replace("norm.weight", "group_norm.weight")
        new_item = new_item.replace("norm.bias", "group_norm.bias")

        new_item = new_item.replace("q.weight", "to_q.weight")
        new_item = new_item.replace("q.bias", "to_q.bias")

        new_item = new_item.replace("k.weight", "to_k.weight")
        new_item = new_item.replace("k.bias", "to_k.bias")

        new_item = new_item.replace("v.weight", "to_v.weight")
        new_item = new_item.replace("v.bias", "to_v.bias")

        new_item = new_item.replace("proj_out.weight", "to_out.0.weight")
        new_item = new_item.replace("proj_out.bias", "to_out.0.bias")

        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)

        mapping.append({"old": old_item, "new": new_item})

    return mapping


def conv_attn_to_linear(checkpoint):
    keys = list(checkpoint.keys())
    attn_keys = ["query.weight", "key.weight", "value.weight"]
    for key in keys:
        if ".".join(key.split(".")[-2:]) in attn_keys:
            if checkpoint[key].ndim > 2:
                checkpoint[key] = checkpoint[key][:, :, 0, 0]
        elif "proj_attn.weight" in key:
            if checkpoint[key].ndim > 2:
                checkpoint[key] = checkpoint[key][:, :, 0]


def convert_from_original_mvdream_ckpt(checkpoint_path, original_config_file, device):
    checkpoint = torch.load(checkpoint_path, map_location=device)
    # print(f"Checkpoint: {checkpoint.keys()}")
    torch.cuda.empty_cache()

    from omegaconf import OmegaConf

    original_config = OmegaConf.load(original_config_file)
    # print(f"Original Config: {original_config}")
    prediction_type = "epsilon"
    image_size = 256
    num_train_timesteps = getattr(original_config.model.params, "timesteps", None) or 1000
    beta_start = getattr(original_config.model.params, "linear_start", None) or 0.02
    beta_end = getattr(original_config.model.params, "linear_end", None) or 0.085
    scheduler = DDIMScheduler(
        beta_end=beta_end,
        beta_schedule="scaled_linear",
        beta_start=beta_start,
        num_train_timesteps=num_train_timesteps,
        steps_offset=1,
        clip_sample=False,
        set_alpha_to_one=False,
        prediction_type=prediction_type,
    )
    scheduler.register_to_config(clip_sample=False)

    # Convert the UNet2DConditionModel model.
    # upcast_attention = None
    # unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
    # unet_config["upcast_attention"] = upcast_attention
    # with init_empty_weights():
    #     unet = UNet2DConditionModel(**unet_config)
    # converted_unet_checkpoint = convert_ldm_unet_checkpoint(
    #     checkpoint, unet_config, path=None, extract_ema=extract_ema
    # )
    # print(f"Unet Config: {original_config.model.params.unet_config.params}")
    unet: MultiViewUNetWrapperModel = MultiViewUNetWrapperModel(**original_config.model.params.unet_config.params)
    # print(f"Unet State Dict: {unet.state_dict().keys()}")
    unet.load_state_dict({key.replace("model.diffusion_model.", "unet."): value for key, value in checkpoint.items() if key.replace("model.diffusion_model.", "unet.") in unet.state_dict()})
    for param_name, param in unet.state_dict().items():
        set_module_tensor_to_device(unet, param_name, "cuda:0", value=param)

    # Convert the VAE model.
    vae_config = create_vae_diffusers_config(original_config, image_size=image_size)
    converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)

    if ("model" in original_config and "params" in original_config.model and "scale_factor" in original_config.model.params):
        vae_scaling_factor = original_config.model.params.scale_factor
    else:
        vae_scaling_factor = 0.18215 # default SD scaling factor

    vae_config["scaling_factor"] = vae_scaling_factor

    with init_empty_weights():
        vae = AutoencoderKL(**vae_config)

    tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
    text_encoder: CLIPTextModel = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14").to(device=torch.device("cuda:0")) # type: ignore

    for param_name, param in converted_vae_checkpoint.items():
        set_module_tensor_to_device(vae, param_name, "cuda:0", value=param)

    pipe = MVDreamStableDiffusionPipeline(
        vae=vae,
        unet=unet,
        tokenizer=tokenizer,
        text_encoder=text_encoder,
        scheduler=scheduler,
    )

    return pipe


if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    parser.add_argument("--checkpoint_path", default=None, type=str, required=True, help="Path to the checkpoint to convert.")
    parser.add_argument(
        "--original_config_file",
        default=None,
        type=str,
        help="The YAML config file corresponding to the original architecture.",
    )
    parser.add_argument(
        "--to_safetensors",
        action="store_true",
        help="Whether to store pipeline in safetensors format or not.",
    )
    parser.add_argument("--half", action="store_true", help="Save weights in half precision.")
    parser.add_argument("--test", help="Whether to test inference after convertion.")
    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")
    parser.add_argument("--device", type=str, help="Device to use (e.g. cpu, cuda:0, cuda:1, etc.)")
    args = parser.parse_args()

    pipe = convert_from_original_mvdream_ckpt(
        checkpoint_path=args.checkpoint_path,
        original_config_file=args.original_config_file,
        device=args.device,
    )

    if args.half:
        pipe.to(torch_dtype=torch.float16)

    if args.test:
        images = pipe(
            prompt="Head of Hatsune Miku",
            negative_prompt="painting, bad quality, flat",
            output_type="pil",
            guidance_scale=7.5,
            num_inference_steps=50,
        )
        for i, image in enumerate(images):
            image.save(f"image_{i}.png") # type: ignore

    pipe.save_pretrained(args.dump_path, safe_serialization=args.to_safetensors)