import torch.nn as nn import torch import cv2 import numpy as np import safetensors.torch as sf from accelerate.logging import get_logger logger = get_logger(__name__, log_level="INFO") from tqdm import tqdm from typing import Optional, Tuple from diffusers.configuration_utils import ConfigMixin, register_to_config from diffusers.models.modeling_utils import ModelMixin from diffusers.models.unets.unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution import torchvision def zero_module(module): """ Zero out the parameters of a module and return it. """ for p in module.parameters(): p.detach().zero_() return module class LatentTransparencyOffsetEncoder(torch.nn.Module): def __init__(self, latent_c=4, *args, **kwargs): super().__init__(*args, **kwargs) self.blocks = torch.nn.Sequential( torch.nn.Conv2d(4, 32, kernel_size=3, padding=1, stride=1), nn.SiLU(), torch.nn.Conv2d(32, 32, kernel_size=3, padding=1, stride=1), nn.SiLU(), torch.nn.Conv2d(32, 64, kernel_size=3, padding=1, stride=2), nn.SiLU(), torch.nn.Conv2d(64, 64, kernel_size=3, padding=1, stride=1), nn.SiLU(), torch.nn.Conv2d(64, 128, kernel_size=3, padding=1, stride=2), nn.SiLU(), torch.nn.Conv2d(128, 128, kernel_size=3, padding=1, stride=1), nn.SiLU(), torch.nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2), nn.SiLU(), torch.nn.Conv2d(256, 256, kernel_size=3, padding=1, stride=1), nn.SiLU(), zero_module(torch.nn.Conv2d(256, latent_c, kernel_size=3, padding=1, stride=1)), ) def __call__(self, x): return self.blocks(x) # 1024 * 1024 * 3 -> 16 * 16 * 512 -> 1024 * 1024 * 3 class UNet1024(ModelMixin, ConfigMixin): @register_to_config def __init__( self, in_channels: int = 3, out_channels: int = 3, down_block_types: Tuple[str] = ("DownBlock2D", "DownBlock2D", "DownBlock2D", "DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D"), up_block_types: Tuple[str] = ("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D", "UpBlock2D", "UpBlock2D", "UpBlock2D"), block_out_channels: Tuple[int] = (32, 32, 64, 128, 256, 512, 512), layers_per_block: int = 2, mid_block_scale_factor: float = 1, downsample_padding: int = 1, downsample_type: str = "conv", upsample_type: str = "conv", dropout: float = 0.0, act_fn: str = "silu", attention_head_dim: Optional[int] = 8, norm_num_groups: int = 4, norm_eps: float = 1e-5, latent_c: int = 4, ): super().__init__() # input self.conv_in = nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)) self.latent_conv_in = zero_module(nn.Conv2d(latent_c, block_out_channels[2], kernel_size=1)) self.down_blocks = nn.ModuleList([]) self.mid_block = None self.up_blocks = nn.ModuleList([]) # down output_channel = block_out_channels[0] for i, down_block_type in enumerate(down_block_types): input_channel = output_channel output_channel = block_out_channels[i] is_final_block = i == len(block_out_channels) - 1 down_block = get_down_block( down_block_type, num_layers=layers_per_block, in_channels=input_channel, out_channels=output_channel, temb_channels=None, add_downsample=not is_final_block, resnet_eps=norm_eps, resnet_act_fn=act_fn, resnet_groups=norm_num_groups, attention_head_dim=attention_head_dim if attention_head_dim is not None else output_channel, downsample_padding=downsample_padding, resnet_time_scale_shift="default", downsample_type=downsample_type, dropout=dropout, ) self.down_blocks.append(down_block) # mid self.mid_block = UNetMidBlock2D( in_channels=block_out_channels[-1], temb_channels=None, dropout=dropout, resnet_eps=norm_eps, resnet_act_fn=act_fn, output_scale_factor=mid_block_scale_factor, resnet_time_scale_shift="default", attention_head_dim=attention_head_dim if attention_head_dim is not None else block_out_channels[-1], resnet_groups=norm_num_groups, attn_groups=None, add_attention=True, ) # up reversed_block_out_channels = list(reversed(block_out_channels)) output_channel = reversed_block_out_channels[0] for i, up_block_type in enumerate(up_block_types): prev_output_channel = output_channel output_channel = reversed_block_out_channels[i] input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] is_final_block = i == len(block_out_channels) - 1 up_block = get_up_block( up_block_type, num_layers=layers_per_block + 1, in_channels=input_channel, out_channels=output_channel, prev_output_channel=prev_output_channel, temb_channels=None, add_upsample=not is_final_block, resnet_eps=norm_eps, resnet_act_fn=act_fn, resnet_groups=norm_num_groups, attention_head_dim=attention_head_dim if attention_head_dim is not None else output_channel, resnet_time_scale_shift="default", upsample_type=upsample_type, dropout=dropout, ) self.up_blocks.append(up_block) prev_output_channel = output_channel # out self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps) self.conv_act = nn.SiLU() self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1) def forward(self, x, latent): sample_latent = self.latent_conv_in(latent) sample = self.conv_in(x) emb = None down_block_res_samples = (sample,) for i, downsample_block in enumerate(self.down_blocks): if i == 3: sample = sample + sample_latent sample, res_samples = downsample_block(hidden_states=sample, temb=emb) down_block_res_samples += res_samples sample = self.mid_block(sample, emb) for upsample_block in self.up_blocks: res_samples = down_block_res_samples[-len(upsample_block.resnets) :] down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] sample = upsample_block(sample, res_samples, emb) sample = self.conv_norm_out(sample) sample = self.conv_act(sample) sample = self.conv_out(sample) return sample def checkerboard(shape): return np.indices(shape).sum(axis=0) % 2 def build_alpha_pyramid(color, alpha, dk=1.2): # Written by lvmin at Stanford # Massive iterative Gaussian filters are mathematically consistent to pyramid. pyramid = [] current_premultiplied_color = color * alpha current_alpha = alpha while True: pyramid.append((current_premultiplied_color, current_alpha)) H, W, C = current_alpha.shape if min(H, W) == 1: break current_premultiplied_color = cv2.resize(current_premultiplied_color, (int(W / dk), int(H / dk)), interpolation=cv2.INTER_AREA) current_alpha = cv2.resize(current_alpha, (int(W / dk), int(H / dk)), interpolation=cv2.INTER_AREA)[:, :, None] return pyramid[::-1] def pad_rgb(np_rgba_hwc_uint8): # Written by lvmin at Stanford # Massive iterative Gaussian filters are mathematically consistent to pyramid. np_rgba_hwc = np_rgba_hwc_uint8.astype(np.float32) #/ 255.0 pyramid = build_alpha_pyramid(color=np_rgba_hwc[..., :3], alpha=np_rgba_hwc[..., 3:]) top_c, top_a = pyramid[0] fg = np.sum(top_c, axis=(0, 1), keepdims=True) / np.sum(top_a, axis=(0, 1), keepdims=True).clip(1e-8, 1e32) for layer_c, layer_a in pyramid: layer_h, layer_w, _ = layer_c.shape fg = cv2.resize(fg, (layer_w, layer_h), interpolation=cv2.INTER_LINEAR) fg = layer_c + fg * (1.0 - layer_a) return fg def dist_sample_deterministic(dist: DiagonalGaussianDistribution, perturbation: torch.Tensor): # Modified from diffusers.models.autoencoders.vae.DiagonalGaussianDistribution.sample() x = dist.mean + dist.std * perturbation.to(dist.std) return x class TransparentVAE(torch.nn.Module): def __init__(self, sd_vae, dtype=torch.float16, encoder_file=None, decoder_file=None, alpha=300.0, latent_c=16, *args, **kwargs): super().__init__(*args, **kwargs) self.dtype = dtype self.sd_vae = sd_vae self.sd_vae.to(dtype=self.dtype) self.sd_vae.requires_grad_(False) self.encoder = LatentTransparencyOffsetEncoder(latent_c=latent_c) if encoder_file is not None: temp = sf.load_file(encoder_file) # del temp['blocks.16.weight'] # del temp['blocks.16.bias'] self.encoder.load_state_dict(temp, strict=True) del temp self.encoder.to(dtype=self.dtype) self.alpha = alpha self.decoder = UNet1024(in_channels=3, out_channels=4, latent_c=latent_c) if decoder_file is not None: temp = sf.load_file(decoder_file) # del temp['latent_conv_in.weight'] # del temp['latent_conv_in.bias'] self.decoder.load_state_dict(temp, strict=True) del temp self.decoder.to(dtype=self.dtype) self.latent_c = latent_c def sd_decode(self, latent): return self.sd_vae.decode(latent) def decode(self, latent, aug=True): origin_pixel = self.sd_vae.decode(latent).sample origin_pixel = (origin_pixel * 0.5 + 0.5) if not aug: y = self.decoder(origin_pixel.to(self.dtype), latent.to(self.dtype)) return origin_pixel, y list_y = [] for i in range(int(latent.shape[0])): y = self.estimate_augmented(origin_pixel[i:i + 1].to(self.dtype), latent[i:i + 1].to(self.dtype)) list_y.append(y) y = torch.concat(list_y, dim=0) return origin_pixel, y def encode(self, img_rgba, img_rgb, padded_img_rgb, use_offset=True): a_bchw_01 = img_rgba[:, 3:, :, :] vae_feed = img_rgb.to(device=self.sd_vae.device, dtype=self.sd_vae.dtype) latent_dist = self.sd_vae.encode(vae_feed).latent_dist offset_feed = torch.cat([padded_img_rgb, a_bchw_01], dim=1).to(device=self.sd_vae.device, dtype=self.dtype) offset = self.encoder(offset_feed) * self.alpha if use_offset: latent = dist_sample_deterministic(dist=latent_dist, perturbation=offset) latent = self.sd_vae.config.scaling_factor * (latent - self.sd_vae.config.shift_factor) else: latent = latent_dist.sample() latent = self.sd_vae.config.scaling_factor * (latent - self.sd_vae.config.shift_factor) return latent def forward(self, img_rgba, img_rgb, padded_img_rgb, use_offset=True): return self.decode(self.encode(img_rgba, img_rgb, padded_img_rgb, use_offset)) @property def device(self): return next(self.parameters()).device @torch.no_grad() def estimate_augmented(self, pixel, latent): args = [ [False, 0], [False, 1], [False, 2], [False, 3], [True, 0], [True, 1], [True, 2], [True, 3], ] result = [] for flip, rok in tqdm(args): feed_pixel = pixel.clone() feed_latent = latent.clone() if flip: feed_pixel = torch.flip(feed_pixel, dims=(3,)) feed_latent = torch.flip(feed_latent, dims=(3,)) feed_pixel = torch.rot90(feed_pixel, k=rok, dims=(2, 3)) feed_latent = torch.rot90(feed_latent, k=rok, dims=(2, 3)) eps = self.decoder(feed_pixel, feed_latent).clip(0, 1) eps = torch.rot90(eps, k=-rok, dims=(2, 3)) if flip: eps = torch.flip(eps, dims=(3,)) result += [eps] result = torch.stack(result, dim=0) median = torch.median(result, dim=0).values return median class TransparentVAEDecoder(torch.nn.Module): def __init__(self, filename, dtype=torch.float16, *args, **kwargs): super().__init__(*args, **kwargs) sd = sf.load_file(filename) model = UNet1024(in_channels=3, out_channels=4) model.load_state_dict(sd, strict=True) model.to(dtype=dtype) model.eval() self.model = model self.dtype = dtype return @torch.no_grad() def estimate_single_pass(self, pixel, latent): y = self.model(pixel, latent) return y @torch.no_grad() def estimate_augmented(self, pixel, latent): args = [ [False, 0], [False, 1], [False, 2], [False, 3], [True, 0], [True, 1], [True, 2], [True, 3], ] result = [] for flip, rok in tqdm(args): feed_pixel = pixel.clone() feed_latent = latent.clone() if flip: feed_pixel = torch.flip(feed_pixel, dims=(3,)) feed_latent = torch.flip(feed_latent, dims=(3,)) feed_pixel = torch.rot90(feed_pixel, k=rok, dims=(2, 3)) feed_latent = torch.rot90(feed_latent, k=rok, dims=(2, 3)) eps = self.estimate_single_pass(feed_pixel, feed_latent).clip(0, 1) eps = torch.rot90(eps, k=-rok, dims=(2, 3)) if flip: eps = torch.flip(eps, dims=(3,)) result += [eps] result = torch.stack(result, dim=0) median = torch.median(result, dim=0).values return median @torch.no_grad() def forward(self, sd_vae, latent): pixel = sd_vae.decode(latent).sample pixel = (pixel * 0.5 + 0.5).clip(0, 1).to(self.dtype) latent = latent.to(self.dtype) result_list = [] vis_list = [] for i in range(int(latent.shape[0])): y = self.estimate_augmented(pixel[i:i + 1], latent[i:i + 1]) y = y.clip(0, 1).movedim(1, -1) alpha = y[..., :1] fg = y[..., 1:] B, H, W, C = fg.shape cb = checkerboard(shape=(H // 64, W // 64)) cb = cv2.resize(cb, (W, H), interpolation=cv2.INTER_NEAREST) cb = (0.5 + (cb - 0.5) * 0.1)[None, ..., None] cb = torch.from_numpy(cb).to(fg) vis = (fg * alpha + cb * (1 - alpha))[0] vis = (vis * 255.0).detach().cpu().float().numpy().clip(0, 255).astype(np.uint8) vis_list.append(vis) png = torch.cat([fg, alpha], dim=3)[0] png = (png * 255.0).detach().cpu().float().numpy().clip(0, 255).astype(np.uint8) result_list.append(png) return result_list, vis_list class TransparentVAEEncoder(torch.nn.Module): def __init__(self, filename, dtype=torch.float16, alpha=300.0, *args, **kwargs): super().__init__(*args, **kwargs) sd = sf.load_file(filename) self.dtype = dtype model = LatentTransparencyOffsetEncoder() model.load_state_dict(sd, strict=True) model.to(dtype=self.dtype) model.eval() self.model = model # similar to LoRA's alpha to avoid initial zero-initialized outputs being too small self.alpha = alpha return @torch.no_grad() def forward(self, sd_vae, list_of_np_rgba_hwc_uint8, use_offset=True): list_of_np_rgb_padded = [pad_rgb(x) for x in list_of_np_rgba_hwc_uint8] rgb_padded_bchw_01 = torch.from_numpy(np.stack(list_of_np_rgb_padded, axis=0)).float().movedim(-1, 1) rgba_bchw_01 = torch.from_numpy(np.stack(list_of_np_rgba_hwc_uint8, axis=0)).float().movedim(-1, 1) / 255.0 rgb_bchw_01 = rgba_bchw_01[:, :3, :, :] a_bchw_01 = rgba_bchw_01[:, 3:, :, :] vae_feed = (rgb_bchw_01 * 2.0 - 1.0) * a_bchw_01 vae_feed = vae_feed.to(device=sd_vae.device, dtype=sd_vae.dtype) latent_dist = sd_vae.encode(vae_feed).latent_dist offset_feed = torch.cat([a_bchw_01, rgb_padded_bchw_01], dim=1).to(device=sd_vae.device, dtype=self.dtype) offset = self.model(offset_feed) * self.alpha if use_offset: latent = dist_sample_deterministic(dist=latent_dist, perturbation=offset) else: latent = latent_dist.sample() return latent