# from .sd3_vae_encoder import SD3VAEEncoder, SDVAEEncoderStateDictConverter # from .sd3_vae_decoder import SD3VAEDecoder, SDVAEDecoderStateDictConverter import torch # from .sd_unet import ResnetBlock, DownSampler # from .sd_vae_encoder import VAEAttentionBlock, SDVAEEncoderStateDictConverter # from .tiler import TileWorker from einops import rearrange from models.attention import Attention class ResnetBlock(torch.nn.Module): def __init__(self, in_channels, out_channels, temb_channels=None, groups=32, eps=1e-5): super().__init__() self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True) self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) if temb_channels is not None: self.time_emb_proj = torch.nn.Linear(temb_channels, out_channels) self.norm2 = torch.nn.GroupNorm(num_groups=groups, num_channels=out_channels, eps=eps, affine=True) self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1) self.nonlinearity = torch.nn.SiLU() self.conv_shortcut = None if in_channels != out_channels: self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0, bias=True) def forward(self, hidden_states, time_emb, text_emb, res_stack, **kwargs): x = hidden_states x = self.norm1(x) x = self.nonlinearity(x) x = self.conv1(x) if time_emb is not None: emb = self.nonlinearity(time_emb) emb = self.time_emb_proj(emb)[:, :, None, None] x = x + emb x = self.norm2(x) x = self.nonlinearity(x) x = self.conv2(x) if self.conv_shortcut is not None: hidden_states = self.conv_shortcut(hidden_states) hidden_states = hidden_states + x return hidden_states, time_emb, text_emb, res_stack class DownSampler(torch.nn.Module): def __init__(self, channels, padding=1, extra_padding=False): super().__init__() self.conv = torch.nn.Conv2d(channels, channels, 3, stride=2, padding=padding) self.extra_padding = extra_padding def forward(self, hidden_states, time_emb, text_emb, res_stack, **kwargs): if self.extra_padding: hidden_states = torch.nn.functional.pad(hidden_states, (0, 1, 0, 1), mode="constant", value=0) hidden_states = self.conv(hidden_states) return hidden_states, time_emb, text_emb, res_stack class UpSampler(torch.nn.Module): def __init__(self, channels): super().__init__() self.conv = torch.nn.Conv2d(channels, channels, 3, padding=1) def forward(self, hidden_states, time_emb, text_emb, res_stack, **kwargs): hidden_states = torch.nn.functional.interpolate(hidden_states, scale_factor=2.0, mode="nearest") hidden_states = self.conv(hidden_states) return hidden_states, time_emb, text_emb, res_stack class VAEAttentionBlock(torch.nn.Module): def __init__(self, num_attention_heads, attention_head_dim, in_channels, num_layers=1, norm_num_groups=32, eps=1e-5): super().__init__() inner_dim = num_attention_heads * attention_head_dim self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=eps, affine=True) self.transformer_blocks = torch.nn.ModuleList([ Attention( inner_dim, num_attention_heads, attention_head_dim, bias_q=True, bias_kv=True, bias_out=True ) for d in range(num_layers) ]) def forward(self, hidden_states, time_emb, text_emb, res_stack): batch, _, height, width = hidden_states.shape residual = hidden_states hidden_states = self.norm(hidden_states) inner_dim = hidden_states.shape[1] hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) for block in self.transformer_blocks: hidden_states = block(hidden_states) hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() hidden_states = hidden_states + residual return hidden_states, time_emb, text_emb, res_stack class SDVAEEncoderStateDictConverter: def __init__(self): pass def from_diffusers(self, state_dict): # architecture block_types = [ 'ResnetBlock', 'ResnetBlock', 'DownSampler', 'ResnetBlock', 'ResnetBlock', 'DownSampler', 'ResnetBlock', 'ResnetBlock', 'DownSampler', 'ResnetBlock', 'ResnetBlock', 'ResnetBlock', 'VAEAttentionBlock', 'ResnetBlock' ] # Rename each parameter local_rename_dict = { "quant_conv": "quant_conv", "encoder.conv_in": "conv_in", "encoder.mid_block.attentions.0.group_norm": "blocks.12.norm", "encoder.mid_block.attentions.0.to_q": "blocks.12.transformer_blocks.0.to_q", "encoder.mid_block.attentions.0.to_k": "blocks.12.transformer_blocks.0.to_k", "encoder.mid_block.attentions.0.to_v": "blocks.12.transformer_blocks.0.to_v", "encoder.mid_block.attentions.0.to_out.0": "blocks.12.transformer_blocks.0.to_out", "encoder.mid_block.resnets.0.norm1": "blocks.11.norm1", "encoder.mid_block.resnets.0.conv1": "blocks.11.conv1", "encoder.mid_block.resnets.0.norm2": "blocks.11.norm2", "encoder.mid_block.resnets.0.conv2": "blocks.11.conv2", "encoder.mid_block.resnets.1.norm1": "blocks.13.norm1", "encoder.mid_block.resnets.1.conv1": "blocks.13.conv1", "encoder.mid_block.resnets.1.norm2": "blocks.13.norm2", "encoder.mid_block.resnets.1.conv2": "blocks.13.conv2", "encoder.conv_norm_out": "conv_norm_out", "encoder.conv_out": "conv_out", } name_list = sorted([name for name in state_dict]) rename_dict = {} block_id = {"ResnetBlock": -1, "DownSampler": -1, "UpSampler": -1} last_block_type_with_id = {"ResnetBlock": "", "DownSampler": "", "UpSampler": ""} for name in name_list: names = name.split(".") name_prefix = ".".join(names[:-1]) if name_prefix in local_rename_dict: rename_dict[name] = local_rename_dict[name_prefix] + "." + names[-1] elif name.startswith("encoder.down_blocks"): block_type = {"resnets": "ResnetBlock", "downsamplers": "DownSampler", "upsamplers": "UpSampler"}[names[3]] block_type_with_id = ".".join(names[:5]) if block_type_with_id != last_block_type_with_id[block_type]: block_id[block_type] += 1 last_block_type_with_id[block_type] = block_type_with_id while block_id[block_type] < len(block_types) and block_types[block_id[block_type]] != block_type: block_id[block_type] += 1 block_type_with_id = ".".join(names[:5]) names = ["blocks", str(block_id[block_type])] + names[5:] rename_dict[name] = ".".join(names) # Convert state_dict state_dict_ = {} for name, param in state_dict.items(): if name in rename_dict: state_dict_[rename_dict[name]] = param return state_dict_ def from_civitai(self, state_dict): rename_dict = { "first_stage_model.encoder.conv_in.bias": "conv_in.bias", "first_stage_model.encoder.conv_in.weight": "conv_in.weight", "first_stage_model.encoder.conv_out.bias": "conv_out.bias", "first_stage_model.encoder.conv_out.weight": "conv_out.weight", "first_stage_model.encoder.down.0.block.0.conv1.bias": "blocks.0.conv1.bias", "first_stage_model.encoder.down.0.block.0.conv1.weight": "blocks.0.conv1.weight", "first_stage_model.encoder.down.0.block.0.conv2.bias": "blocks.0.conv2.bias", "first_stage_model.encoder.down.0.block.0.conv2.weight": "blocks.0.conv2.weight", "first_stage_model.encoder.down.0.block.0.norm1.bias": "blocks.0.norm1.bias", "first_stage_model.encoder.down.0.block.0.norm1.weight": "blocks.0.norm1.weight", "first_stage_model.encoder.down.0.block.0.norm2.bias": "blocks.0.norm2.bias", "first_stage_model.encoder.down.0.block.0.norm2.weight": "blocks.0.norm2.weight", "first_stage_model.encoder.down.0.block.1.conv1.bias": "blocks.1.conv1.bias", "first_stage_model.encoder.down.0.block.1.conv1.weight": "blocks.1.conv1.weight", "first_stage_model.encoder.down.0.block.1.conv2.bias": "blocks.1.conv2.bias", "first_stage_model.encoder.down.0.block.1.conv2.weight": "blocks.1.conv2.weight", "first_stage_model.encoder.down.0.block.1.norm1.bias": "blocks.1.norm1.bias", "first_stage_model.encoder.down.0.block.1.norm1.weight": "blocks.1.norm1.weight", "first_stage_model.encoder.down.0.block.1.norm2.bias": "blocks.1.norm2.bias", "first_stage_model.encoder.down.0.block.1.norm2.weight": "blocks.1.norm2.weight", "first_stage_model.encoder.down.0.downsample.conv.bias": "blocks.2.conv.bias", "first_stage_model.encoder.down.0.downsample.conv.weight": "blocks.2.conv.weight", "first_stage_model.encoder.down.1.block.0.conv1.bias": "blocks.3.conv1.bias", "first_stage_model.encoder.down.1.block.0.conv1.weight": "blocks.3.conv1.weight", "first_stage_model.encoder.down.1.block.0.conv2.bias": "blocks.3.conv2.bias", "first_stage_model.encoder.down.1.block.0.conv2.weight": "blocks.3.conv2.weight", "first_stage_model.encoder.down.1.block.0.nin_shortcut.bias": "blocks.3.conv_shortcut.bias", "first_stage_model.encoder.down.1.block.0.nin_shortcut.weight": "blocks.3.conv_shortcut.weight", "first_stage_model.encoder.down.1.block.0.norm1.bias": "blocks.3.norm1.bias", "first_stage_model.encoder.down.1.block.0.norm1.weight": "blocks.3.norm1.weight", "first_stage_model.encoder.down.1.block.0.norm2.bias": "blocks.3.norm2.bias", "first_stage_model.encoder.down.1.block.0.norm2.weight": "blocks.3.norm2.weight", "first_stage_model.encoder.down.1.block.1.conv1.bias": "blocks.4.conv1.bias", "first_stage_model.encoder.down.1.block.1.conv1.weight": "blocks.4.conv1.weight", "first_stage_model.encoder.down.1.block.1.conv2.bias": "blocks.4.conv2.bias", "first_stage_model.encoder.down.1.block.1.conv2.weight": "blocks.4.conv2.weight", "first_stage_model.encoder.down.1.block.1.norm1.bias": "blocks.4.norm1.bias", "first_stage_model.encoder.down.1.block.1.norm1.weight": "blocks.4.norm1.weight", "first_stage_model.encoder.down.1.block.1.norm2.bias": "blocks.4.norm2.bias", "first_stage_model.encoder.down.1.block.1.norm2.weight": "blocks.4.norm2.weight", "first_stage_model.encoder.down.1.downsample.conv.bias": "blocks.5.conv.bias", "first_stage_model.encoder.down.1.downsample.conv.weight": "blocks.5.conv.weight", "first_stage_model.encoder.down.2.block.0.conv1.bias": "blocks.6.conv1.bias", "first_stage_model.encoder.down.2.block.0.conv1.weight": "blocks.6.conv1.weight", "first_stage_model.encoder.down.2.block.0.conv2.bias": "blocks.6.conv2.bias", "first_stage_model.encoder.down.2.block.0.conv2.weight": "blocks.6.conv2.weight", "first_stage_model.encoder.down.2.block.0.nin_shortcut.bias": "blocks.6.conv_shortcut.bias", "first_stage_model.encoder.down.2.block.0.nin_shortcut.weight": "blocks.6.conv_shortcut.weight", "first_stage_model.encoder.down.2.block.0.norm1.bias": "blocks.6.norm1.bias", "first_stage_model.encoder.down.2.block.0.norm1.weight": "blocks.6.norm1.weight", "first_stage_model.encoder.down.2.block.0.norm2.bias": "blocks.6.norm2.bias", "first_stage_model.encoder.down.2.block.0.norm2.weight": "blocks.6.norm2.weight", "first_stage_model.encoder.down.2.block.1.conv1.bias": "blocks.7.conv1.bias", "first_stage_model.encoder.down.2.block.1.conv1.weight": "blocks.7.conv1.weight", "first_stage_model.encoder.down.2.block.1.conv2.bias": "blocks.7.conv2.bias", "first_stage_model.encoder.down.2.block.1.conv2.weight": "blocks.7.conv2.weight", "first_stage_model.encoder.down.2.block.1.norm1.bias": "blocks.7.norm1.bias", "first_stage_model.encoder.down.2.block.1.norm1.weight": "blocks.7.norm1.weight", "first_stage_model.encoder.down.2.block.1.norm2.bias": "blocks.7.norm2.bias", "first_stage_model.encoder.down.2.block.1.norm2.weight": "blocks.7.norm2.weight", "first_stage_model.encoder.down.2.downsample.conv.bias": "blocks.8.conv.bias", "first_stage_model.encoder.down.2.downsample.conv.weight": "blocks.8.conv.weight", "first_stage_model.encoder.down.3.block.0.conv1.bias": "blocks.9.conv1.bias", "first_stage_model.encoder.down.3.block.0.conv1.weight": "blocks.9.conv1.weight", "first_stage_model.encoder.down.3.block.0.conv2.bias": "blocks.9.conv2.bias", "first_stage_model.encoder.down.3.block.0.conv2.weight": "blocks.9.conv2.weight", "first_stage_model.encoder.down.3.block.0.norm1.bias": "blocks.9.norm1.bias", "first_stage_model.encoder.down.3.block.0.norm1.weight": "blocks.9.norm1.weight", "first_stage_model.encoder.down.3.block.0.norm2.bias": "blocks.9.norm2.bias", "first_stage_model.encoder.down.3.block.0.norm2.weight": "blocks.9.norm2.weight", "first_stage_model.encoder.down.3.block.1.conv1.bias": "blocks.10.conv1.bias", "first_stage_model.encoder.down.3.block.1.conv1.weight": "blocks.10.conv1.weight", "first_stage_model.encoder.down.3.block.1.conv2.bias": "blocks.10.conv2.bias", "first_stage_model.encoder.down.3.block.1.conv2.weight": "blocks.10.conv2.weight", "first_stage_model.encoder.down.3.block.1.norm1.bias": "blocks.10.norm1.bias", "first_stage_model.encoder.down.3.block.1.norm1.weight": "blocks.10.norm1.weight", "first_stage_model.encoder.down.3.block.1.norm2.bias": "blocks.10.norm2.bias", "first_stage_model.encoder.down.3.block.1.norm2.weight": "blocks.10.norm2.weight", "first_stage_model.encoder.mid.attn_1.k.bias": "blocks.12.transformer_blocks.0.to_k.bias", "first_stage_model.encoder.mid.attn_1.k.weight": "blocks.12.transformer_blocks.0.to_k.weight", "first_stage_model.encoder.mid.attn_1.norm.bias": "blocks.12.norm.bias", "first_stage_model.encoder.mid.attn_1.norm.weight": "blocks.12.norm.weight", "first_stage_model.encoder.mid.attn_1.proj_out.bias": "blocks.12.transformer_blocks.0.to_out.bias", "first_stage_model.encoder.mid.attn_1.proj_out.weight": "blocks.12.transformer_blocks.0.to_out.weight", "first_stage_model.encoder.mid.attn_1.q.bias": "blocks.12.transformer_blocks.0.to_q.bias", "first_stage_model.encoder.mid.attn_1.q.weight": "blocks.12.transformer_blocks.0.to_q.weight", "first_stage_model.encoder.mid.attn_1.v.bias": "blocks.12.transformer_blocks.0.to_v.bias", "first_stage_model.encoder.mid.attn_1.v.weight": "blocks.12.transformer_blocks.0.to_v.weight", "first_stage_model.encoder.mid.block_1.conv1.bias": "blocks.11.conv1.bias", "first_stage_model.encoder.mid.block_1.conv1.weight": "blocks.11.conv1.weight", "first_stage_model.encoder.mid.block_1.conv2.bias": "blocks.11.conv2.bias", "first_stage_model.encoder.mid.block_1.conv2.weight": "blocks.11.conv2.weight", "first_stage_model.encoder.mid.block_1.norm1.bias": "blocks.11.norm1.bias", "first_stage_model.encoder.mid.block_1.norm1.weight": "blocks.11.norm1.weight", "first_stage_model.encoder.mid.block_1.norm2.bias": "blocks.11.norm2.bias", "first_stage_model.encoder.mid.block_1.norm2.weight": "blocks.11.norm2.weight", "first_stage_model.encoder.mid.block_2.conv1.bias": "blocks.13.conv1.bias", "first_stage_model.encoder.mid.block_2.conv1.weight": "blocks.13.conv1.weight", "first_stage_model.encoder.mid.block_2.conv2.bias": "blocks.13.conv2.bias", "first_stage_model.encoder.mid.block_2.conv2.weight": "blocks.13.conv2.weight", "first_stage_model.encoder.mid.block_2.norm1.bias": "blocks.13.norm1.bias", "first_stage_model.encoder.mid.block_2.norm1.weight": "blocks.13.norm1.weight", "first_stage_model.encoder.mid.block_2.norm2.bias": "blocks.13.norm2.bias", "first_stage_model.encoder.mid.block_2.norm2.weight": "blocks.13.norm2.weight", "first_stage_model.encoder.norm_out.bias": "conv_norm_out.bias", "first_stage_model.encoder.norm_out.weight": "conv_norm_out.weight", "first_stage_model.quant_conv.bias": "quant_conv.bias", "first_stage_model.quant_conv.weight": "quant_conv.weight", } state_dict_ = {} for name in state_dict: if name in rename_dict: param = state_dict[name] if "transformer_blocks" in rename_dict[name]: param = param.squeeze() state_dict_[rename_dict[name]] = param return state_dict_ class SD3VAEEncoder(torch.nn.Module): def __init__(self): super().__init__() self.scaling_factor = 1.5305 # Different from SD 1.x self.shift_factor = 0.0609 # Different from SD 1.x self.conv_in = torch.nn.Conv2d(3, 128, kernel_size=3, padding=1) self.blocks = torch.nn.ModuleList([ # DownEncoderBlock2D ResnetBlock(128, 128, eps=1e-6), ResnetBlock(128, 128, eps=1e-6), DownSampler(128, padding=0, extra_padding=True), # DownEncoderBlock2D ResnetBlock(128, 256, eps=1e-6), ResnetBlock(256, 256, eps=1e-6), DownSampler(256, padding=0, extra_padding=True), # DownEncoderBlock2D ResnetBlock(256, 512, eps=1e-6), ResnetBlock(512, 512, eps=1e-6), DownSampler(512, padding=0, extra_padding=True), # DownEncoderBlock2D ResnetBlock(512, 512, eps=1e-6), ResnetBlock(512, 512, eps=1e-6), # UNetMidBlock2D ResnetBlock(512, 512, eps=1e-6), VAEAttentionBlock(1, 512, 512, 1, eps=1e-6), ResnetBlock(512, 512, eps=1e-6), ]) self.conv_norm_out = torch.nn.GroupNorm(num_channels=512, num_groups=32, eps=1e-6) self.conv_act = torch.nn.SiLU() self.conv_out = torch.nn.Conv2d(512, 32, kernel_size=3, padding=1) def tiled_forward(self, sample, tile_size=64, tile_stride=32): hidden_states = TileWorker().tiled_forward( # type: ignore lambda x: self.forward(x), sample, tile_size, tile_stride, tile_device=sample.device, tile_dtype=sample.dtype ) return hidden_states def forward(self, sample, tiled=False, tile_size=64, tile_stride=32, **kwargs): # For VAE Decoder, we do not need to apply the tiler on each layer. if tiled: return self.tiled_forward(sample, tile_size=tile_size, tile_stride=tile_stride) # 1. pre-process hidden_states = self.conv_in(sample) time_emb = None text_emb = None res_stack = None # 2. blocks for i, block in enumerate(self.blocks): hidden_states, time_emb, text_emb, res_stack = block(hidden_states, time_emb, text_emb, res_stack) # 3. output hidden_states = self.conv_norm_out(hidden_states) hidden_states = self.conv_act(hidden_states) hidden_states = self.conv_out(hidden_states) hidden_states = hidden_states[:, :16] hidden_states = (hidden_states - self.shift_factor) * self.scaling_factor return hidden_states def encode_video(self, sample, batch_size=8): B = sample.shape[0] hidden_states = [] for i in range(0, sample.shape[2], batch_size): j = min(i + batch_size, sample.shape[2]) sample_batch = rearrange(sample[:,:,i:j], "B C T H W -> (B T) C H W") hidden_states_batch = self(sample_batch) hidden_states_batch = rearrange(hidden_states_batch, "(B T) C H W -> B C T H W", B=B) hidden_states.append(hidden_states_batch) hidden_states = torch.concat(hidden_states, dim=2) return hidden_states @staticmethod def state_dict_converter(): return SDVAEEncoderStateDictConverter() class SDVAEDecoderStateDictConverter: def __init__(self): pass def from_diffusers(self, state_dict): # architecture block_types = [ 'ResnetBlock', 'VAEAttentionBlock', 'ResnetBlock', 'ResnetBlock', 'ResnetBlock', 'ResnetBlock', 'UpSampler', 'ResnetBlock', 'ResnetBlock', 'ResnetBlock', 'UpSampler', 'ResnetBlock', 'ResnetBlock', 'ResnetBlock', 'UpSampler', 'ResnetBlock', 'ResnetBlock', 'ResnetBlock' ] # Rename each parameter local_rename_dict = { "post_quant_conv": "post_quant_conv", "decoder.conv_in": "conv_in", "decoder.mid_block.attentions.0.group_norm": "blocks.1.norm", "decoder.mid_block.attentions.0.to_q": "blocks.1.transformer_blocks.0.to_q", "decoder.mid_block.attentions.0.to_k": "blocks.1.transformer_blocks.0.to_k", "decoder.mid_block.attentions.0.to_v": "blocks.1.transformer_blocks.0.to_v", "decoder.mid_block.attentions.0.to_out.0": "blocks.1.transformer_blocks.0.to_out", "decoder.mid_block.resnets.0.norm1": "blocks.0.norm1", "decoder.mid_block.resnets.0.conv1": "blocks.0.conv1", "decoder.mid_block.resnets.0.norm2": "blocks.0.norm2", "decoder.mid_block.resnets.0.conv2": "blocks.0.conv2", "decoder.mid_block.resnets.1.norm1": "blocks.2.norm1", "decoder.mid_block.resnets.1.conv1": "blocks.2.conv1", "decoder.mid_block.resnets.1.norm2": "blocks.2.norm2", "decoder.mid_block.resnets.1.conv2": "blocks.2.conv2", "decoder.conv_norm_out": "conv_norm_out", "decoder.conv_out": "conv_out", } name_list = sorted([name for name in state_dict]) rename_dict = {} block_id = {"ResnetBlock": 2, "DownSampler": 2, "UpSampler": 2} last_block_type_with_id = {"ResnetBlock": "", "DownSampler": "", "UpSampler": ""} for name in name_list: names = name.split(".") name_prefix = ".".join(names[:-1]) if name_prefix in local_rename_dict: rename_dict[name] = local_rename_dict[name_prefix] + "." + names[-1] elif name.startswith("decoder.up_blocks"): block_type = {"resnets": "ResnetBlock", "downsamplers": "DownSampler", "upsamplers": "UpSampler"}[names[3]] block_type_with_id = ".".join(names[:5]) if block_type_with_id != last_block_type_with_id[block_type]: block_id[block_type] += 1 last_block_type_with_id[block_type] = block_type_with_id while block_id[block_type] < len(block_types) and block_types[block_id[block_type]] != block_type: block_id[block_type] += 1 block_type_with_id = ".".join(names[:5]) names = ["blocks", str(block_id[block_type])] + names[5:] rename_dict[name] = ".".join(names) # Convert state_dict state_dict_ = {} for name, param in state_dict.items(): if name in rename_dict: state_dict_[rename_dict[name]] = param return state_dict_ def from_civitai(self, state_dict): rename_dict = { "first_stage_model.decoder.conv_in.bias": "conv_in.bias", "first_stage_model.decoder.conv_in.weight": "conv_in.weight", "first_stage_model.decoder.conv_out.bias": "conv_out.bias", "first_stage_model.decoder.conv_out.weight": "conv_out.weight", "first_stage_model.decoder.mid.attn_1.k.bias": "blocks.1.transformer_blocks.0.to_k.bias", "first_stage_model.decoder.mid.attn_1.k.weight": "blocks.1.transformer_blocks.0.to_k.weight", "first_stage_model.decoder.mid.attn_1.norm.bias": "blocks.1.norm.bias", "first_stage_model.decoder.mid.attn_1.norm.weight": "blocks.1.norm.weight", "first_stage_model.decoder.mid.attn_1.proj_out.bias": "blocks.1.transformer_blocks.0.to_out.bias", "first_stage_model.decoder.mid.attn_1.proj_out.weight": "blocks.1.transformer_blocks.0.to_out.weight", "first_stage_model.decoder.mid.attn_1.q.bias": "blocks.1.transformer_blocks.0.to_q.bias", "first_stage_model.decoder.mid.attn_1.q.weight": "blocks.1.transformer_blocks.0.to_q.weight", "first_stage_model.decoder.mid.attn_1.v.bias": "blocks.1.transformer_blocks.0.to_v.bias", "first_stage_model.decoder.mid.attn_1.v.weight": "blocks.1.transformer_blocks.0.to_v.weight", "first_stage_model.decoder.mid.block_1.conv1.bias": "blocks.0.conv1.bias", "first_stage_model.decoder.mid.block_1.conv1.weight": "blocks.0.conv1.weight", "first_stage_model.decoder.mid.block_1.conv2.bias": "blocks.0.conv2.bias", "first_stage_model.decoder.mid.block_1.conv2.weight": "blocks.0.conv2.weight", "first_stage_model.decoder.mid.block_1.norm1.bias": "blocks.0.norm1.bias", "first_stage_model.decoder.mid.block_1.norm1.weight": "blocks.0.norm1.weight", "first_stage_model.decoder.mid.block_1.norm2.bias": "blocks.0.norm2.bias", "first_stage_model.decoder.mid.block_1.norm2.weight": "blocks.0.norm2.weight", "first_stage_model.decoder.mid.block_2.conv1.bias": "blocks.2.conv1.bias", "first_stage_model.decoder.mid.block_2.conv1.weight": "blocks.2.conv1.weight", "first_stage_model.decoder.mid.block_2.conv2.bias": "blocks.2.conv2.bias", "first_stage_model.decoder.mid.block_2.conv2.weight": "blocks.2.conv2.weight", "first_stage_model.decoder.mid.block_2.norm1.bias": "blocks.2.norm1.bias", "first_stage_model.decoder.mid.block_2.norm1.weight": "blocks.2.norm1.weight", "first_stage_model.decoder.mid.block_2.norm2.bias": "blocks.2.norm2.bias", "first_stage_model.decoder.mid.block_2.norm2.weight": "blocks.2.norm2.weight", "first_stage_model.decoder.norm_out.bias": "conv_norm_out.bias", "first_stage_model.decoder.norm_out.weight": "conv_norm_out.weight", "first_stage_model.decoder.up.0.block.0.conv1.bias": "blocks.15.conv1.bias", "first_stage_model.decoder.up.0.block.0.conv1.weight": "blocks.15.conv1.weight", "first_stage_model.decoder.up.0.block.0.conv2.bias": "blocks.15.conv2.bias", "first_stage_model.decoder.up.0.block.0.conv2.weight": "blocks.15.conv2.weight", "first_stage_model.decoder.up.0.block.0.nin_shortcut.bias": "blocks.15.conv_shortcut.bias", "first_stage_model.decoder.up.0.block.0.nin_shortcut.weight": "blocks.15.conv_shortcut.weight", "first_stage_model.decoder.up.0.block.0.norm1.bias": "blocks.15.norm1.bias", "first_stage_model.decoder.up.0.block.0.norm1.weight": "blocks.15.norm1.weight", "first_stage_model.decoder.up.0.block.0.norm2.bias": "blocks.15.norm2.bias", "first_stage_model.decoder.up.0.block.0.norm2.weight": "blocks.15.norm2.weight", "first_stage_model.decoder.up.0.block.1.conv1.bias": "blocks.16.conv1.bias", "first_stage_model.decoder.up.0.block.1.conv1.weight": "blocks.16.conv1.weight", "first_stage_model.decoder.up.0.block.1.conv2.bias": "blocks.16.conv2.bias", "first_stage_model.decoder.up.0.block.1.conv2.weight": "blocks.16.conv2.weight", "first_stage_model.decoder.up.0.block.1.norm1.bias": "blocks.16.norm1.bias", "first_stage_model.decoder.up.0.block.1.norm1.weight": "blocks.16.norm1.weight", "first_stage_model.decoder.up.0.block.1.norm2.bias": "blocks.16.norm2.bias", "first_stage_model.decoder.up.0.block.1.norm2.weight": "blocks.16.norm2.weight", "first_stage_model.decoder.up.0.block.2.conv1.bias": "blocks.17.conv1.bias", "first_stage_model.decoder.up.0.block.2.conv1.weight": "blocks.17.conv1.weight", "first_stage_model.decoder.up.0.block.2.conv2.bias": "blocks.17.conv2.bias", "first_stage_model.decoder.up.0.block.2.conv2.weight": "blocks.17.conv2.weight", "first_stage_model.decoder.up.0.block.2.norm1.bias": "blocks.17.norm1.bias", "first_stage_model.decoder.up.0.block.2.norm1.weight": "blocks.17.norm1.weight", "first_stage_model.decoder.up.0.block.2.norm2.bias": "blocks.17.norm2.bias", "first_stage_model.decoder.up.0.block.2.norm2.weight": "blocks.17.norm2.weight", "first_stage_model.decoder.up.1.block.0.conv1.bias": "blocks.11.conv1.bias", "first_stage_model.decoder.up.1.block.0.conv1.weight": "blocks.11.conv1.weight", "first_stage_model.decoder.up.1.block.0.conv2.bias": "blocks.11.conv2.bias", "first_stage_model.decoder.up.1.block.0.conv2.weight": "blocks.11.conv2.weight", "first_stage_model.decoder.up.1.block.0.nin_shortcut.bias": "blocks.11.conv_shortcut.bias", "first_stage_model.decoder.up.1.block.0.nin_shortcut.weight": "blocks.11.conv_shortcut.weight", "first_stage_model.decoder.up.1.block.0.norm1.bias": "blocks.11.norm1.bias", "first_stage_model.decoder.up.1.block.0.norm1.weight": "blocks.11.norm1.weight", "first_stage_model.decoder.up.1.block.0.norm2.bias": "blocks.11.norm2.bias", "first_stage_model.decoder.up.1.block.0.norm2.weight": "blocks.11.norm2.weight", "first_stage_model.decoder.up.1.block.1.conv1.bias": "blocks.12.conv1.bias", "first_stage_model.decoder.up.1.block.1.conv1.weight": "blocks.12.conv1.weight", "first_stage_model.decoder.up.1.block.1.conv2.bias": "blocks.12.conv2.bias", "first_stage_model.decoder.up.1.block.1.conv2.weight": "blocks.12.conv2.weight", "first_stage_model.decoder.up.1.block.1.norm1.bias": "blocks.12.norm1.bias", "first_stage_model.decoder.up.1.block.1.norm1.weight": "blocks.12.norm1.weight", "first_stage_model.decoder.up.1.block.1.norm2.bias": "blocks.12.norm2.bias", "first_stage_model.decoder.up.1.block.1.norm2.weight": "blocks.12.norm2.weight", "first_stage_model.decoder.up.1.block.2.conv1.bias": "blocks.13.conv1.bias", "first_stage_model.decoder.up.1.block.2.conv1.weight": "blocks.13.conv1.weight", "first_stage_model.decoder.up.1.block.2.conv2.bias": "blocks.13.conv2.bias", "first_stage_model.decoder.up.1.block.2.conv2.weight": "blocks.13.conv2.weight", "first_stage_model.decoder.up.1.block.2.norm1.bias": "blocks.13.norm1.bias", "first_stage_model.decoder.up.1.block.2.norm1.weight": "blocks.13.norm1.weight", "first_stage_model.decoder.up.1.block.2.norm2.bias": "blocks.13.norm2.bias", "first_stage_model.decoder.up.1.block.2.norm2.weight": "blocks.13.norm2.weight", "first_stage_model.decoder.up.1.upsample.conv.bias": "blocks.14.conv.bias", "first_stage_model.decoder.up.1.upsample.conv.weight": "blocks.14.conv.weight", "first_stage_model.decoder.up.2.block.0.conv1.bias": "blocks.7.conv1.bias", "first_stage_model.decoder.up.2.block.0.conv1.weight": "blocks.7.conv1.weight", "first_stage_model.decoder.up.2.block.0.conv2.bias": "blocks.7.conv2.bias", "first_stage_model.decoder.up.2.block.0.conv2.weight": "blocks.7.conv2.weight", "first_stage_model.decoder.up.2.block.0.norm1.bias": "blocks.7.norm1.bias", "first_stage_model.decoder.up.2.block.0.norm1.weight": "blocks.7.norm1.weight", "first_stage_model.decoder.up.2.block.0.norm2.bias": "blocks.7.norm2.bias", "first_stage_model.decoder.up.2.block.0.norm2.weight": "blocks.7.norm2.weight", "first_stage_model.decoder.up.2.block.1.conv1.bias": "blocks.8.conv1.bias", "first_stage_model.decoder.up.2.block.1.conv1.weight": "blocks.8.conv1.weight", "first_stage_model.decoder.up.2.block.1.conv2.bias": "blocks.8.conv2.bias", "first_stage_model.decoder.up.2.block.1.conv2.weight": "blocks.8.conv2.weight", "first_stage_model.decoder.up.2.block.1.norm1.bias": "blocks.8.norm1.bias", "first_stage_model.decoder.up.2.block.1.norm1.weight": "blocks.8.norm1.weight", "first_stage_model.decoder.up.2.block.1.norm2.bias": "blocks.8.norm2.bias", "first_stage_model.decoder.up.2.block.1.norm2.weight": "blocks.8.norm2.weight", "first_stage_model.decoder.up.2.block.2.conv1.bias": "blocks.9.conv1.bias", "first_stage_model.decoder.up.2.block.2.conv1.weight": "blocks.9.conv1.weight", "first_stage_model.decoder.up.2.block.2.conv2.bias": "blocks.9.conv2.bias", "first_stage_model.decoder.up.2.block.2.conv2.weight": "blocks.9.conv2.weight", "first_stage_model.decoder.up.2.block.2.norm1.bias": "blocks.9.norm1.bias", "first_stage_model.decoder.up.2.block.2.norm1.weight": "blocks.9.norm1.weight", "first_stage_model.decoder.up.2.block.2.norm2.bias": "blocks.9.norm2.bias", "first_stage_model.decoder.up.2.block.2.norm2.weight": "blocks.9.norm2.weight", "first_stage_model.decoder.up.2.upsample.conv.bias": "blocks.10.conv.bias", "first_stage_model.decoder.up.2.upsample.conv.weight": "blocks.10.conv.weight", "first_stage_model.decoder.up.3.block.0.conv1.bias": "blocks.3.conv1.bias", "first_stage_model.decoder.up.3.block.0.conv1.weight": "blocks.3.conv1.weight", "first_stage_model.decoder.up.3.block.0.conv2.bias": "blocks.3.conv2.bias", "first_stage_model.decoder.up.3.block.0.conv2.weight": "blocks.3.conv2.weight", "first_stage_model.decoder.up.3.block.0.norm1.bias": "blocks.3.norm1.bias", "first_stage_model.decoder.up.3.block.0.norm1.weight": "blocks.3.norm1.weight", "first_stage_model.decoder.up.3.block.0.norm2.bias": "blocks.3.norm2.bias", "first_stage_model.decoder.up.3.block.0.norm2.weight": "blocks.3.norm2.weight", "first_stage_model.decoder.up.3.block.1.conv1.bias": "blocks.4.conv1.bias", "first_stage_model.decoder.up.3.block.1.conv1.weight": "blocks.4.conv1.weight", "first_stage_model.decoder.up.3.block.1.conv2.bias": "blocks.4.conv2.bias", "first_stage_model.decoder.up.3.block.1.conv2.weight": "blocks.4.conv2.weight", "first_stage_model.decoder.up.3.block.1.norm1.bias": "blocks.4.norm1.bias", "first_stage_model.decoder.up.3.block.1.norm1.weight": "blocks.4.norm1.weight", "first_stage_model.decoder.up.3.block.1.norm2.bias": "blocks.4.norm2.bias", "first_stage_model.decoder.up.3.block.1.norm2.weight": "blocks.4.norm2.weight", "first_stage_model.decoder.up.3.block.2.conv1.bias": "blocks.5.conv1.bias", "first_stage_model.decoder.up.3.block.2.conv1.weight": "blocks.5.conv1.weight", "first_stage_model.decoder.up.3.block.2.conv2.bias": "blocks.5.conv2.bias", "first_stage_model.decoder.up.3.block.2.conv2.weight": "blocks.5.conv2.weight", "first_stage_model.decoder.up.3.block.2.norm1.bias": "blocks.5.norm1.bias", "first_stage_model.decoder.up.3.block.2.norm1.weight": "blocks.5.norm1.weight", "first_stage_model.decoder.up.3.block.2.norm2.bias": "blocks.5.norm2.bias", "first_stage_model.decoder.up.3.block.2.norm2.weight": "blocks.5.norm2.weight", "first_stage_model.decoder.up.3.upsample.conv.bias": "blocks.6.conv.bias", "first_stage_model.decoder.up.3.upsample.conv.weight": "blocks.6.conv.weight", "first_stage_model.post_quant_conv.bias": "post_quant_conv.bias", "first_stage_model.post_quant_conv.weight": "post_quant_conv.weight", } state_dict_ = {} for name in state_dict: if name in rename_dict: param = state_dict[name] if "transformer_blocks" in rename_dict[name]: param = param.squeeze() state_dict_[rename_dict[name]] = param return state_dict_ class SD3VAEDecoder(torch.nn.Module): def __init__(self): super().__init__() self.scaling_factor = 1.5305 # Different from SD 1.x self.shift_factor = 0.0609 # Different from SD 1.x self.conv_in = torch.nn.Conv2d(16, 512, kernel_size=3, padding=1) # Different from SD 1.x self.blocks = torch.nn.ModuleList([ # UNetMidBlock2D ResnetBlock(512, 512, eps=1e-6), VAEAttentionBlock(1, 512, 512, 1, eps=1e-6), ResnetBlock(512, 512, eps=1e-6), # UpDecoderBlock2D ResnetBlock(512, 512, eps=1e-6), ResnetBlock(512, 512, eps=1e-6), ResnetBlock(512, 512, eps=1e-6), UpSampler(512), # UpDecoderBlock2D ResnetBlock(512, 512, eps=1e-6), ResnetBlock(512, 512, eps=1e-6), ResnetBlock(512, 512, eps=1e-6), UpSampler(512), # UpDecoderBlock2D ResnetBlock(512, 256, eps=1e-6), ResnetBlock(256, 256, eps=1e-6), ResnetBlock(256, 256, eps=1e-6), UpSampler(256), # UpDecoderBlock2D ResnetBlock(256, 128, eps=1e-6), ResnetBlock(128, 128, eps=1e-6), ResnetBlock(128, 128, eps=1e-6), ]) self.conv_norm_out = torch.nn.GroupNorm(num_channels=128, num_groups=32, eps=1e-6) self.conv_act = torch.nn.SiLU() self.conv_out = torch.nn.Conv2d(128, 3, kernel_size=3, padding=1) def tiled_forward(self, sample, tile_size=64, tile_stride=32): hidden_states = TileWorker().tiled_forward( # type: ignore lambda x: self.forward(x), sample, tile_size, tile_stride, tile_device=sample.device, tile_dtype=sample.dtype ) return hidden_states def forward(self, sample, tiled=False, tile_size=64, tile_stride=32, **kwargs): # For VAE Decoder, we do not need to apply the tiler on each layer. if tiled: return self.tiled_forward(sample, tile_size=tile_size, tile_stride=tile_stride) # 1. pre-process hidden_states = sample / self.scaling_factor + self.shift_factor hidden_states = self.conv_in(hidden_states) time_emb = None text_emb = None res_stack = None # 2. blocks for i, block in enumerate(self.blocks): hidden_states, time_emb, text_emb, res_stack = block(hidden_states, time_emb, text_emb, res_stack) # 3. output hidden_states = self.conv_norm_out(hidden_states) hidden_states = self.conv_act(hidden_states) hidden_states = self.conv_out(hidden_states) return hidden_states @staticmethod def state_dict_converter(): return SDVAEDecoderStateDictConverter() class FluxVAEEncoder(SD3VAEEncoder): def __init__(self): super().__init__() self.scaling_factor = 0.3611 self.shift_factor = 0.1159 @staticmethod def state_dict_converter(): return FluxVAEEncoderStateDictConverter() class FluxVAEDecoder(SD3VAEDecoder): def __init__(self): super().__init__() self.scaling_factor = 0.3611 self.shift_factor = 0.1159 @staticmethod def state_dict_converter(): return FluxVAEDecoderStateDictConverter() class FluxVAEEncoderStateDictConverter(SDVAEEncoderStateDictConverter): def __init__(self): pass def from_civitai(self, state_dict): rename_dict = { "encoder.conv_in.bias": "conv_in.bias", "encoder.conv_in.weight": "conv_in.weight", "encoder.conv_out.bias": "conv_out.bias", "encoder.conv_out.weight": "conv_out.weight", "encoder.down.0.block.0.conv1.bias": "blocks.0.conv1.bias", "encoder.down.0.block.0.conv1.weight": "blocks.0.conv1.weight", "encoder.down.0.block.0.conv2.bias": "blocks.0.conv2.bias", "encoder.down.0.block.0.conv2.weight": "blocks.0.conv2.weight", "encoder.down.0.block.0.norm1.bias": "blocks.0.norm1.bias", "encoder.down.0.block.0.norm1.weight": "blocks.0.norm1.weight", "encoder.down.0.block.0.norm2.bias": "blocks.0.norm2.bias", "encoder.down.0.block.0.norm2.weight": "blocks.0.norm2.weight", "encoder.down.0.block.1.conv1.bias": "blocks.1.conv1.bias", "encoder.down.0.block.1.conv1.weight": "blocks.1.conv1.weight", "encoder.down.0.block.1.conv2.bias": "blocks.1.conv2.bias", "encoder.down.0.block.1.conv2.weight": "blocks.1.conv2.weight", "encoder.down.0.block.1.norm1.bias": "blocks.1.norm1.bias", "encoder.down.0.block.1.norm1.weight": "blocks.1.norm1.weight", "encoder.down.0.block.1.norm2.bias": "blocks.1.norm2.bias", "encoder.down.0.block.1.norm2.weight": "blocks.1.norm2.weight", "encoder.down.0.downsample.conv.bias": "blocks.2.conv.bias", "encoder.down.0.downsample.conv.weight": "blocks.2.conv.weight", "encoder.down.1.block.0.conv1.bias": "blocks.3.conv1.bias", "encoder.down.1.block.0.conv1.weight": "blocks.3.conv1.weight", "encoder.down.1.block.0.conv2.bias": "blocks.3.conv2.bias", "encoder.down.1.block.0.conv2.weight": "blocks.3.conv2.weight", "encoder.down.1.block.0.nin_shortcut.bias": "blocks.3.conv_shortcut.bias", "encoder.down.1.block.0.nin_shortcut.weight": "blocks.3.conv_shortcut.weight", "encoder.down.1.block.0.norm1.bias": "blocks.3.norm1.bias", "encoder.down.1.block.0.norm1.weight": "blocks.3.norm1.weight", "encoder.down.1.block.0.norm2.bias": "blocks.3.norm2.bias", "encoder.down.1.block.0.norm2.weight": "blocks.3.norm2.weight", "encoder.down.1.block.1.conv1.bias": "blocks.4.conv1.bias", "encoder.down.1.block.1.conv1.weight": "blocks.4.conv1.weight", "encoder.down.1.block.1.conv2.bias": "blocks.4.conv2.bias", "encoder.down.1.block.1.conv2.weight": "blocks.4.conv2.weight", "encoder.down.1.block.1.norm1.bias": "blocks.4.norm1.bias", "encoder.down.1.block.1.norm1.weight": "blocks.4.norm1.weight", "encoder.down.1.block.1.norm2.bias": "blocks.4.norm2.bias", "encoder.down.1.block.1.norm2.weight": "blocks.4.norm2.weight", "encoder.down.1.downsample.conv.bias": "blocks.5.conv.bias", "encoder.down.1.downsample.conv.weight": "blocks.5.conv.weight", "encoder.down.2.block.0.conv1.bias": "blocks.6.conv1.bias", "encoder.down.2.block.0.conv1.weight": "blocks.6.conv1.weight", "encoder.down.2.block.0.conv2.bias": "blocks.6.conv2.bias", "encoder.down.2.block.0.conv2.weight": "blocks.6.conv2.weight", "encoder.down.2.block.0.nin_shortcut.bias": "blocks.6.conv_shortcut.bias", "encoder.down.2.block.0.nin_shortcut.weight": "blocks.6.conv_shortcut.weight", "encoder.down.2.block.0.norm1.bias": "blocks.6.norm1.bias", "encoder.down.2.block.0.norm1.weight": "blocks.6.norm1.weight", "encoder.down.2.block.0.norm2.bias": "blocks.6.norm2.bias", "encoder.down.2.block.0.norm2.weight": "blocks.6.norm2.weight", "encoder.down.2.block.1.conv1.bias": "blocks.7.conv1.bias", "encoder.down.2.block.1.conv1.weight": "blocks.7.conv1.weight", "encoder.down.2.block.1.conv2.bias": "blocks.7.conv2.bias", "encoder.down.2.block.1.conv2.weight": "blocks.7.conv2.weight", "encoder.down.2.block.1.norm1.bias": "blocks.7.norm1.bias", "encoder.down.2.block.1.norm1.weight": "blocks.7.norm1.weight", "encoder.down.2.block.1.norm2.bias": "blocks.7.norm2.bias", "encoder.down.2.block.1.norm2.weight": "blocks.7.norm2.weight", "encoder.down.2.downsample.conv.bias": "blocks.8.conv.bias", "encoder.down.2.downsample.conv.weight": "blocks.8.conv.weight", "encoder.down.3.block.0.conv1.bias": "blocks.9.conv1.bias", "encoder.down.3.block.0.conv1.weight": "blocks.9.conv1.weight", "encoder.down.3.block.0.conv2.bias": "blocks.9.conv2.bias", "encoder.down.3.block.0.conv2.weight": "blocks.9.conv2.weight", "encoder.down.3.block.0.norm1.bias": "blocks.9.norm1.bias", "encoder.down.3.block.0.norm1.weight": "blocks.9.norm1.weight", "encoder.down.3.block.0.norm2.bias": "blocks.9.norm2.bias", "encoder.down.3.block.0.norm2.weight": "blocks.9.norm2.weight", "encoder.down.3.block.1.conv1.bias": "blocks.10.conv1.bias", "encoder.down.3.block.1.conv1.weight": "blocks.10.conv1.weight", "encoder.down.3.block.1.conv2.bias": "blocks.10.conv2.bias", "encoder.down.3.block.1.conv2.weight": "blocks.10.conv2.weight", "encoder.down.3.block.1.norm1.bias": "blocks.10.norm1.bias", "encoder.down.3.block.1.norm1.weight": "blocks.10.norm1.weight", "encoder.down.3.block.1.norm2.bias": "blocks.10.norm2.bias", "encoder.down.3.block.1.norm2.weight": "blocks.10.norm2.weight", "encoder.mid.attn_1.k.bias": "blocks.12.transformer_blocks.0.to_k.bias", "encoder.mid.attn_1.k.weight": "blocks.12.transformer_blocks.0.to_k.weight", "encoder.mid.attn_1.norm.bias": "blocks.12.norm.bias", "encoder.mid.attn_1.norm.weight": "blocks.12.norm.weight", "encoder.mid.attn_1.proj_out.bias": "blocks.12.transformer_blocks.0.to_out.bias", "encoder.mid.attn_1.proj_out.weight": "blocks.12.transformer_blocks.0.to_out.weight", "encoder.mid.attn_1.q.bias": "blocks.12.transformer_blocks.0.to_q.bias", "encoder.mid.attn_1.q.weight": "blocks.12.transformer_blocks.0.to_q.weight", "encoder.mid.attn_1.v.bias": "blocks.12.transformer_blocks.0.to_v.bias", "encoder.mid.attn_1.v.weight": "blocks.12.transformer_blocks.0.to_v.weight", "encoder.mid.block_1.conv1.bias": "blocks.11.conv1.bias", "encoder.mid.block_1.conv1.weight": "blocks.11.conv1.weight", "encoder.mid.block_1.conv2.bias": "blocks.11.conv2.bias", "encoder.mid.block_1.conv2.weight": "blocks.11.conv2.weight", "encoder.mid.block_1.norm1.bias": "blocks.11.norm1.bias", "encoder.mid.block_1.norm1.weight": "blocks.11.norm1.weight", "encoder.mid.block_1.norm2.bias": "blocks.11.norm2.bias", "encoder.mid.block_1.norm2.weight": "blocks.11.norm2.weight", "encoder.mid.block_2.conv1.bias": "blocks.13.conv1.bias", "encoder.mid.block_2.conv1.weight": "blocks.13.conv1.weight", "encoder.mid.block_2.conv2.bias": "blocks.13.conv2.bias", "encoder.mid.block_2.conv2.weight": "blocks.13.conv2.weight", "encoder.mid.block_2.norm1.bias": "blocks.13.norm1.bias", "encoder.mid.block_2.norm1.weight": "blocks.13.norm1.weight", "encoder.mid.block_2.norm2.bias": "blocks.13.norm2.bias", "encoder.mid.block_2.norm2.weight": "blocks.13.norm2.weight", "encoder.norm_out.bias": "conv_norm_out.bias", "encoder.norm_out.weight": "conv_norm_out.weight", } state_dict_ = {} for name in state_dict: if name in rename_dict: param = state_dict[name] if "transformer_blocks" in rename_dict[name]: param = param.squeeze() state_dict_[rename_dict[name]] = param return state_dict_ class FluxVAEDecoderStateDictConverter(SDVAEDecoderStateDictConverter): def __init__(self): pass def from_civitai(self, state_dict): rename_dict = { "decoder.conv_in.bias": "conv_in.bias", "decoder.conv_in.weight": "conv_in.weight", "decoder.conv_out.bias": "conv_out.bias", "decoder.conv_out.weight": "conv_out.weight", "decoder.mid.attn_1.k.bias": "blocks.1.transformer_blocks.0.to_k.bias", "decoder.mid.attn_1.k.weight": "blocks.1.transformer_blocks.0.to_k.weight", "decoder.mid.attn_1.norm.bias": "blocks.1.norm.bias", "decoder.mid.attn_1.norm.weight": "blocks.1.norm.weight", "decoder.mid.attn_1.proj_out.bias": "blocks.1.transformer_blocks.0.to_out.bias", "decoder.mid.attn_1.proj_out.weight": "blocks.1.transformer_blocks.0.to_out.weight", "decoder.mid.attn_1.q.bias": "blocks.1.transformer_blocks.0.to_q.bias", "decoder.mid.attn_1.q.weight": "blocks.1.transformer_blocks.0.to_q.weight", "decoder.mid.attn_1.v.bias": "blocks.1.transformer_blocks.0.to_v.bias", "decoder.mid.attn_1.v.weight": "blocks.1.transformer_blocks.0.to_v.weight", "decoder.mid.block_1.conv1.bias": "blocks.0.conv1.bias", "decoder.mid.block_1.conv1.weight": "blocks.0.conv1.weight", "decoder.mid.block_1.conv2.bias": "blocks.0.conv2.bias", "decoder.mid.block_1.conv2.weight": "blocks.0.conv2.weight", "decoder.mid.block_1.norm1.bias": "blocks.0.norm1.bias", "decoder.mid.block_1.norm1.weight": "blocks.0.norm1.weight", "decoder.mid.block_1.norm2.bias": "blocks.0.norm2.bias", "decoder.mid.block_1.norm2.weight": "blocks.0.norm2.weight", "decoder.mid.block_2.conv1.bias": "blocks.2.conv1.bias", "decoder.mid.block_2.conv1.weight": "blocks.2.conv1.weight", "decoder.mid.block_2.conv2.bias": "blocks.2.conv2.bias", "decoder.mid.block_2.conv2.weight": "blocks.2.conv2.weight", "decoder.mid.block_2.norm1.bias": "blocks.2.norm1.bias", "decoder.mid.block_2.norm1.weight": "blocks.2.norm1.weight", "decoder.mid.block_2.norm2.bias": "blocks.2.norm2.bias", "decoder.mid.block_2.norm2.weight": "blocks.2.norm2.weight", "decoder.norm_out.bias": "conv_norm_out.bias", "decoder.norm_out.weight": "conv_norm_out.weight", "decoder.up.0.block.0.conv1.bias": "blocks.15.conv1.bias", "decoder.up.0.block.0.conv1.weight": "blocks.15.conv1.weight", "decoder.up.0.block.0.conv2.bias": "blocks.15.conv2.bias", "decoder.up.0.block.0.conv2.weight": "blocks.15.conv2.weight", "decoder.up.0.block.0.nin_shortcut.bias": "blocks.15.conv_shortcut.bias", "decoder.up.0.block.0.nin_shortcut.weight": "blocks.15.conv_shortcut.weight", "decoder.up.0.block.0.norm1.bias": "blocks.15.norm1.bias", "decoder.up.0.block.0.norm1.weight": "blocks.15.norm1.weight", "decoder.up.0.block.0.norm2.bias": "blocks.15.norm2.bias", "decoder.up.0.block.0.norm2.weight": "blocks.15.norm2.weight", "decoder.up.0.block.1.conv1.bias": "blocks.16.conv1.bias", "decoder.up.0.block.1.conv1.weight": "blocks.16.conv1.weight", "decoder.up.0.block.1.conv2.bias": "blocks.16.conv2.bias", "decoder.up.0.block.1.conv2.weight": "blocks.16.conv2.weight", "decoder.up.0.block.1.norm1.bias": "blocks.16.norm1.bias", "decoder.up.0.block.1.norm1.weight": "blocks.16.norm1.weight", "decoder.up.0.block.1.norm2.bias": "blocks.16.norm2.bias", "decoder.up.0.block.1.norm2.weight": "blocks.16.norm2.weight", "decoder.up.0.block.2.conv1.bias": "blocks.17.conv1.bias", "decoder.up.0.block.2.conv1.weight": "blocks.17.conv1.weight", "decoder.up.0.block.2.conv2.bias": "blocks.17.conv2.bias", "decoder.up.0.block.2.conv2.weight": "blocks.17.conv2.weight", "decoder.up.0.block.2.norm1.bias": "blocks.17.norm1.bias", "decoder.up.0.block.2.norm1.weight": "blocks.17.norm1.weight", "decoder.up.0.block.2.norm2.bias": "blocks.17.norm2.bias", "decoder.up.0.block.2.norm2.weight": "blocks.17.norm2.weight", "decoder.up.1.block.0.conv1.bias": "blocks.11.conv1.bias", "decoder.up.1.block.0.conv1.weight": "blocks.11.conv1.weight", "decoder.up.1.block.0.conv2.bias": "blocks.11.conv2.bias", "decoder.up.1.block.0.conv2.weight": "blocks.11.conv2.weight", "decoder.up.1.block.0.nin_shortcut.bias": "blocks.11.conv_shortcut.bias", "decoder.up.1.block.0.nin_shortcut.weight": "blocks.11.conv_shortcut.weight", "decoder.up.1.block.0.norm1.bias": "blocks.11.norm1.bias", "decoder.up.1.block.0.norm1.weight": "blocks.11.norm1.weight", "decoder.up.1.block.0.norm2.bias": "blocks.11.norm2.bias", "decoder.up.1.block.0.norm2.weight": "blocks.11.norm2.weight", "decoder.up.1.block.1.conv1.bias": "blocks.12.conv1.bias", "decoder.up.1.block.1.conv1.weight": "blocks.12.conv1.weight", "decoder.up.1.block.1.conv2.bias": "blocks.12.conv2.bias", "decoder.up.1.block.1.conv2.weight": "blocks.12.conv2.weight", "decoder.up.1.block.1.norm1.bias": "blocks.12.norm1.bias", "decoder.up.1.block.1.norm1.weight": "blocks.12.norm1.weight", "decoder.up.1.block.1.norm2.bias": "blocks.12.norm2.bias", "decoder.up.1.block.1.norm2.weight": "blocks.12.norm2.weight", "decoder.up.1.block.2.conv1.bias": "blocks.13.conv1.bias", "decoder.up.1.block.2.conv1.weight": "blocks.13.conv1.weight", "decoder.up.1.block.2.conv2.bias": "blocks.13.conv2.bias", "decoder.up.1.block.2.conv2.weight": "blocks.13.conv2.weight", "decoder.up.1.block.2.norm1.bias": "blocks.13.norm1.bias", "decoder.up.1.block.2.norm1.weight": "blocks.13.norm1.weight", "decoder.up.1.block.2.norm2.bias": "blocks.13.norm2.bias", "decoder.up.1.block.2.norm2.weight": "blocks.13.norm2.weight", "decoder.up.1.upsample.conv.bias": "blocks.14.conv.bias", "decoder.up.1.upsample.conv.weight": "blocks.14.conv.weight", "decoder.up.2.block.0.conv1.bias": "blocks.7.conv1.bias", "decoder.up.2.block.0.conv1.weight": "blocks.7.conv1.weight", "decoder.up.2.block.0.conv2.bias": "blocks.7.conv2.bias", "decoder.up.2.block.0.conv2.weight": "blocks.7.conv2.weight", "decoder.up.2.block.0.norm1.bias": "blocks.7.norm1.bias", "decoder.up.2.block.0.norm1.weight": "blocks.7.norm1.weight", "decoder.up.2.block.0.norm2.bias": "blocks.7.norm2.bias", "decoder.up.2.block.0.norm2.weight": "blocks.7.norm2.weight", "decoder.up.2.block.1.conv1.bias": "blocks.8.conv1.bias", "decoder.up.2.block.1.conv1.weight": "blocks.8.conv1.weight", "decoder.up.2.block.1.conv2.bias": "blocks.8.conv2.bias", "decoder.up.2.block.1.conv2.weight": "blocks.8.conv2.weight", "decoder.up.2.block.1.norm1.bias": "blocks.8.norm1.bias", "decoder.up.2.block.1.norm1.weight": "blocks.8.norm1.weight", "decoder.up.2.block.1.norm2.bias": "blocks.8.norm2.bias", "decoder.up.2.block.1.norm2.weight": "blocks.8.norm2.weight", "decoder.up.2.block.2.conv1.bias": "blocks.9.conv1.bias", "decoder.up.2.block.2.conv1.weight": "blocks.9.conv1.weight", "decoder.up.2.block.2.conv2.bias": "blocks.9.conv2.bias", "decoder.up.2.block.2.conv2.weight": "blocks.9.conv2.weight", "decoder.up.2.block.2.norm1.bias": "blocks.9.norm1.bias", "decoder.up.2.block.2.norm1.weight": "blocks.9.norm1.weight", "decoder.up.2.block.2.norm2.bias": "blocks.9.norm2.bias", "decoder.up.2.block.2.norm2.weight": "blocks.9.norm2.weight", "decoder.up.2.upsample.conv.bias": "blocks.10.conv.bias", "decoder.up.2.upsample.conv.weight": "blocks.10.conv.weight", "decoder.up.3.block.0.conv1.bias": "blocks.3.conv1.bias", "decoder.up.3.block.0.conv1.weight": "blocks.3.conv1.weight", "decoder.up.3.block.0.conv2.bias": "blocks.3.conv2.bias", "decoder.up.3.block.0.conv2.weight": "blocks.3.conv2.weight", "decoder.up.3.block.0.norm1.bias": "blocks.3.norm1.bias", "decoder.up.3.block.0.norm1.weight": "blocks.3.norm1.weight", "decoder.up.3.block.0.norm2.bias": "blocks.3.norm2.bias", "decoder.up.3.block.0.norm2.weight": "blocks.3.norm2.weight", "decoder.up.3.block.1.conv1.bias": "blocks.4.conv1.bias", "decoder.up.3.block.1.conv1.weight": "blocks.4.conv1.weight", "decoder.up.3.block.1.conv2.bias": "blocks.4.conv2.bias", "decoder.up.3.block.1.conv2.weight": "blocks.4.conv2.weight", "decoder.up.3.block.1.norm1.bias": "blocks.4.norm1.bias", "decoder.up.3.block.1.norm1.weight": "blocks.4.norm1.weight", "decoder.up.3.block.1.norm2.bias": "blocks.4.norm2.bias", "decoder.up.3.block.1.norm2.weight": "blocks.4.norm2.weight", "decoder.up.3.block.2.conv1.bias": "blocks.5.conv1.bias", "decoder.up.3.block.2.conv1.weight": "blocks.5.conv1.weight", "decoder.up.3.block.2.conv2.bias": "blocks.5.conv2.bias", "decoder.up.3.block.2.conv2.weight": "blocks.5.conv2.weight", "decoder.up.3.block.2.norm1.bias": "blocks.5.norm1.bias", "decoder.up.3.block.2.norm1.weight": "blocks.5.norm1.weight", "decoder.up.3.block.2.norm2.bias": "blocks.5.norm2.bias", "decoder.up.3.block.2.norm2.weight": "blocks.5.norm2.weight", "decoder.up.3.upsample.conv.bias": "blocks.6.conv.bias", "decoder.up.3.upsample.conv.weight": "blocks.6.conv.weight", } state_dict_ = {} for name in state_dict: if name in rename_dict: param = state_dict[name] if "transformer_blocks" in rename_dict[name]: param = param.squeeze() state_dict_[rename_dict[name]] = param return state_dict_