そもそも Stable Diffusion とは

2022年頃から出てきた超スーパーハイパーウルトラめちゃくちゃ優秀な画像生成AI君です。

どういうふうに使うかというと、基本的には生成したい画像のキャプションとなる文章を打ち込むことで、それに準ずる画像がしゅぽっと出てくるって感じですね。

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models, arXiv:2112.10752 [cs.CV].

裏で何が行われているのか

ここからは少し難しい話をしましょう。

深層学習

さて、 Stable Diffusion を話す上で避けて通れないものとして、深層学習(Deep Learning)というものがあります。

ニューラルネットワーク(NN)という人間の神経構造(ニューロン)を模倣したネットワークを形成することで、ヒューリスティックな手法ながら多くの分野で採択されており、またその多くで従来の機械学習モデルより高性能であるとの報告が後を絶たないですね。

この NN は Python のライブラリで PyTorch や TensorFlow などでサポートされており、誰でも気軽に構築することができます(ん？)。

今回は、筆者自身が PyTorch 大好きマンなのでこちらを使って考えていきたいと思います(とは言っても全然使いこなせないのですが)。

拡散モデル(Diffusion Model)

拡散モデルは、 NN ベースの生成系のモデルになります。

元を辿ると2015年の Jascha Sohl-Dickstein と愉快な仲間たちが出した "Deep Unsupervised Learning using Nonequilibrium Thermodynamics" で提唱されたものだそうです。

出発点は非平衡熱力学の分野だったんですね！

どんなものかというのをざっくり説明すると、

まず完璧な状態のデータを用意する
T ステップに分けて徐々にノイズを加えていく
T ステップ後から逆順にノイズを除去するように学習させ、生成する空間を見つける

といった流れです。

NVIDIA Developer. Improving Diffusion Models as an Alternative To GANs, Part 1.

Transformer

これは自然言語処理の分野から出てきて、文字通り世界を変えた革新的なモデルですね。

2017年、 Google 社のチームが発表した "Attention Is All You Need" という論文で提唱されたもので、これが出る前と出た後では自然言語処理の分野の発展が著しいのはもちろん、画像処理や時系列解析といった他分野でも活躍するという話も上がっています。

自己注意機構(Self-Attention Mechanism)と呼ばれる入力データの各々にその他の部分との関係を計算することでデータを文脈単位で機械が理解するということを可能にしています。これが Transformer の中核となる要素です！

また、位置エンコーディング(Positional Encoding)なるもので、各データのシークエンスの順序情報を理解します。

マルチヘッド自己注意(Multi-Head Self-Attention)によって複数の視点から文脈を理解し、偏った考え方をしなくなります。

位置単位順伝播ネットワーク(Position-Wise Feedforward Network)で、各データの表現を更新します。

そして、エンコーダ(Encoder)で入力の情報を処理し、デコーダ(Decoder)で出力の形に持っていくということをします。

これらの構成要素を持って学習させることで、今までにない革命的なモデルが誕生しました！

Stable Diffusion ではテキストから画像を生成するため、この Transformer の技術も取り入れています。

何がしたいん？

ここからの目標は、 Stable Diffusion のモデルのアーキテクチャを一つひとつ理解して、~~紛い物~~オリジナルモデルを作りたいなと思っています！

そのためにはまず、よく使われているモデルのアーキテクチャを調べなければいけないんですよね。で、どうするか。

Hugging Face のドキュメントにある Stability AI が公式で出している Stable Diffusion のモデルがあるので、こいつのアーキテクチャを覗いてみましょう！

UNet2DConditionModel(
  (conv_in): Conv2d(4, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (time_proj): Timesteps()
  (time_embedding): TimestepEmbedding(
    (linear_1): LoRACompatibleLinear(in_features=320, out_features=1280, bias=True)
    (act): SiLU()
    (linear_2): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
  )
  (down_blocks): ModuleList(
    (0): CrossAttnDownBlock2D(
      (attentions): ModuleList(
        (0-1): 2 x Transformer2DModel(
          (norm): GroupNorm(32, 320, eps=1e-06, affine=True)
          (proj_in): LoRACompatibleLinear(in_features=320, out_features=320, bias=True)
          (transformer_blocks): ModuleList(
            (0): BasicTransformerBlock(
              (norm1): LayerNorm((320,), eps=1e-05, elementwise_affine=True)
              (attn1): Attention(
                (to_q): LoRACompatibleLinear(in_features=320, out_features=320, bias=False)
                (to_k): LoRACompatibleLinear(in_features=320, out_features=320, bias=False)
                (to_v): LoRACompatibleLinear(in_features=320, out_features=320, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=320, out_features=320, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm2): LayerNorm((320,), eps=1e-05, elementwise_affine=True)
              (attn2): Attention(
                (to_q): LoRACompatibleLinear(in_features=320, out_features=320, bias=False)
                (to_k): LoRACompatibleLinear(in_features=1024, out_features=320, bias=False)
                (to_v): LoRACompatibleLinear(in_features=1024, out_features=320, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=320, out_features=320, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm3): LayerNorm((320,), eps=1e-05, elementwise_affine=True)
              (ff): FeedForward(
                (net): ModuleList(
                  (0): GEGLU(
                    (proj): LoRACompatibleLinear(in_features=320, out_features=2560, bias=True)
                  )
                  (1): Dropout(p=0.0, inplace=False)
                  (2): LoRACompatibleLinear(in_features=1280, out_features=320, bias=True)
                )
              )
            )
          )
          (proj_out): LoRACompatibleLinear(in_features=320, out_features=320, bias=True)
        )
      )
      (resnets): ModuleList(
        (0-1): 2 x ResnetBlock2D(
          (norm1): GroupNorm(32, 320, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=320, bias=True)
          (norm2): GroupNorm(32, 320, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
        )
      )
      (downsamplers): ModuleList(
        (0): Downsample2D(
          (conv): LoRACompatibleConv(320, 320, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
        )
      )
    )
    (1): CrossAttnDownBlock2D(
      (attentions): ModuleList(
        (0-1): 2 x Transformer2DModel(
          (norm): GroupNorm(32, 640, eps=1e-06, affine=True)
          (proj_in): LoRACompatibleLinear(in_features=640, out_features=640, bias=True)
          (transformer_blocks): ModuleList(
            (0): BasicTransformerBlock(
              (norm1): LayerNorm((640,), eps=1e-05, elementwise_affine=True)
              (attn1): Attention(
                (to_q): LoRACompatibleLinear(in_features=640, out_features=640, bias=False)
                (to_k): LoRACompatibleLinear(in_features=640, out_features=640, bias=False)
                (to_v): LoRACompatibleLinear(in_features=640, out_features=640, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=640, out_features=640, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm2): LayerNorm((640,), eps=1e-05, elementwise_affine=True)
              (attn2): Attention(
                (to_q): LoRACompatibleLinear(in_features=640, out_features=640, bias=False)
                (to_k): LoRACompatibleLinear(in_features=1024, out_features=640, bias=False)
                (to_v): LoRACompatibleLinear(in_features=1024, out_features=640, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=640, out_features=640, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm3): LayerNorm((640,), eps=1e-05, elementwise_affine=True)
              (ff): FeedForward(
                (net): ModuleList(
                  (0): GEGLU(
                    (proj): LoRACompatibleLinear(in_features=640, out_features=5120, bias=True)
                  )
                  (1): Dropout(p=0.0, inplace=False)
                  (2): LoRACompatibleLinear(in_features=2560, out_features=640, bias=True)
                )
              )
            )
          )
          (proj_out): LoRACompatibleLinear(in_features=640, out_features=640, bias=True)
        )
      )
      (resnets): ModuleList(
        (0): ResnetBlock2D(
          (norm1): GroupNorm(32, 320, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(320, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=640, bias=True)
          (norm2): GroupNorm(32, 640, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(320, 640, kernel_size=(1, 1), stride=(1, 1))
        )
        (1): ResnetBlock2D(
          (norm1): GroupNorm(32, 640, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=640, bias=True)
          (norm2): GroupNorm(32, 640, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
        )
      )
      (downsamplers): ModuleList(
        (0): Downsample2D(
          (conv): LoRACompatibleConv(640, 640, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
        )
      )
    )
    (2): CrossAttnDownBlock2D(
      (attentions): ModuleList(
        (0-1): 2 x Transformer2DModel(
          (norm): GroupNorm(32, 1280, eps=1e-06, affine=True)
          (proj_in): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
          (transformer_blocks): ModuleList(
            (0): BasicTransformerBlock(
              (norm1): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
              (attn1): Attention(
                (to_q): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
                (to_k): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
                (to_v): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm2): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
              (attn2): Attention(
                (to_q): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
                (to_k): LoRACompatibleLinear(in_features=1024, out_features=1280, bias=False)
                (to_v): LoRACompatibleLinear(in_features=1024, out_features=1280, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm3): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
              (ff): FeedForward(
                (net): ModuleList(
                  (0): GEGLU(
                    (proj): LoRACompatibleLinear(in_features=1280, out_features=10240, bias=True)
                  )
                  (1): Dropout(p=0.0, inplace=False)
                  (2): LoRACompatibleLinear(in_features=5120, out_features=1280, bias=True)
                )
              )
            )
          )
          (proj_out): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
        )
      )
      (resnets): ModuleList(
        (0): ResnetBlock2D(
          (norm1): GroupNorm(32, 640, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(640, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
          (norm2): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(640, 1280, kernel_size=(1, 1), stride=(1, 1))
        )
        (1): ResnetBlock2D(
          (norm1): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
          (norm2): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
        )
      )
      (downsamplers): ModuleList(
        (0): Downsample2D(
          (conv): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
        )
      )
    )
    (3): DownBlock2D(
      (resnets): ModuleList(
        (0-1): 2 x ResnetBlock2D(
          (norm1): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
          (norm2): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
        )
      )
    )
  )
  (up_blocks): ModuleList(
    (0): UpBlock2D(
      (resnets): ModuleList(
        (0-2): 3 x ResnetBlock2D(
          (norm1): GroupNorm(32, 2560, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(2560, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
          (norm2): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(2560, 1280, kernel_size=(1, 1), stride=(1, 1))
        )
      )
      (upsamplers): ModuleList(
        (0): Upsample2D(
          (conv): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        )
      )
    )
    (1): CrossAttnUpBlock2D(
      (attentions): ModuleList(
        (0-2): 3 x Transformer2DModel(
          (norm): GroupNorm(32, 1280, eps=1e-06, affine=True)
          (proj_in): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
          (transformer_blocks): ModuleList(
            (0): BasicTransformerBlock(
              (norm1): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
              (attn1): Attention(
                (to_q): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
                (to_k): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
                (to_v): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm2): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
              (attn2): Attention(
                (to_q): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
                (to_k): LoRACompatibleLinear(in_features=1024, out_features=1280, bias=False)
                (to_v): LoRACompatibleLinear(in_features=1024, out_features=1280, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm3): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
              (ff): FeedForward(
                (net): ModuleList(
                  (0): GEGLU(
                    (proj): LoRACompatibleLinear(in_features=1280, out_features=10240, bias=True)
                  )
                  (1): Dropout(p=0.0, inplace=False)
                  (2): LoRACompatibleLinear(in_features=5120, out_features=1280, bias=True)
                )
              )
            )
          )
          (proj_out): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
        )
      )
      (resnets): ModuleList(
        (0-1): 2 x ResnetBlock2D(
          (norm1): GroupNorm(32, 2560, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(2560, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
          (norm2): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(2560, 1280, kernel_size=(1, 1), stride=(1, 1))
        )
        (2): ResnetBlock2D(
          (norm1): GroupNorm(32, 1920, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(1920, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
          (norm2): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(1920, 1280, kernel_size=(1, 1), stride=(1, 1))
        )
      )
      (upsamplers): ModuleList(
        (0): Upsample2D(
          (conv): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        )
      )
    )
    (2): CrossAttnUpBlock2D(
      (attentions): ModuleList(
        (0-2): 3 x Transformer2DModel(
          (norm): GroupNorm(32, 640, eps=1e-06, affine=True)
          (proj_in): LoRACompatibleLinear(in_features=640, out_features=640, bias=True)
          (transformer_blocks): ModuleList(
            (0): BasicTransformerBlock(
              (norm1): LayerNorm((640,), eps=1e-05, elementwise_affine=True)
              (attn1): Attention(
                (to_q): LoRACompatibleLinear(in_features=640, out_features=640, bias=False)
                (to_k): LoRACompatibleLinear(in_features=640, out_features=640, bias=False)
                (to_v): LoRACompatibleLinear(in_features=640, out_features=640, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=640, out_features=640, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm2): LayerNorm((640,), eps=1e-05, elementwise_affine=True)
              (attn2): Attention(
                (to_q): LoRACompatibleLinear(in_features=640, out_features=640, bias=False)
                (to_k): LoRACompatibleLinear(in_features=1024, out_features=640, bias=False)
                (to_v): LoRACompatibleLinear(in_features=1024, out_features=640, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=640, out_features=640, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm3): LayerNorm((640,), eps=1e-05, elementwise_affine=True)
              (ff): FeedForward(
                (net): ModuleList(
                  (0): GEGLU(
                    (proj): LoRACompatibleLinear(in_features=640, out_features=5120, bias=True)
                  )
                  (1): Dropout(p=0.0, inplace=False)
                  (2): LoRACompatibleLinear(in_features=2560, out_features=640, bias=True)
                )
              )
            )
          )
          (proj_out): LoRACompatibleLinear(in_features=640, out_features=640, bias=True)
        )
      )
      (resnets): ModuleList(
        (0): ResnetBlock2D(
          (norm1): GroupNorm(32, 1920, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(1920, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=640, bias=True)
          (norm2): GroupNorm(32, 640, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(1920, 640, kernel_size=(1, 1), stride=(1, 1))
        )
        (1): ResnetBlock2D(
          (norm1): GroupNorm(32, 1280, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(1280, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=640, bias=True)
          (norm2): GroupNorm(32, 640, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(1280, 640, kernel_size=(1, 1), stride=(1, 1))
        )
        (2): ResnetBlock2D(
          (norm1): GroupNorm(32, 960, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(960, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=640, bias=True)
          (norm2): GroupNorm(32, 640, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(960, 640, kernel_size=(1, 1), stride=(1, 1))
        )
      )
      (upsamplers): ModuleList(
        (0): Upsample2D(
          (conv): LoRACompatibleConv(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        )
      )
    )
    (3): CrossAttnUpBlock2D(
      (attentions): ModuleList(
        (0-2): 3 x Transformer2DModel(
          (norm): GroupNorm(32, 320, eps=1e-06, affine=True)
          (proj_in): LoRACompatibleLinear(in_features=320, out_features=320, bias=True)
          (transformer_blocks): ModuleList(
            (0): BasicTransformerBlock(
              (norm1): LayerNorm((320,), eps=1e-05, elementwise_affine=True)
              (attn1): Attention(
                (to_q): LoRACompatibleLinear(in_features=320, out_features=320, bias=False)
                (to_k): LoRACompatibleLinear(in_features=320, out_features=320, bias=False)
                (to_v): LoRACompatibleLinear(in_features=320, out_features=320, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=320, out_features=320, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm2): LayerNorm((320,), eps=1e-05, elementwise_affine=True)
              (attn2): Attention(
                (to_q): LoRACompatibleLinear(in_features=320, out_features=320, bias=False)
                (to_k): LoRACompatibleLinear(in_features=1024, out_features=320, bias=False)
                (to_v): LoRACompatibleLinear(in_features=1024, out_features=320, bias=False)
                (to_out): ModuleList(
                  (0): LoRACompatibleLinear(in_features=320, out_features=320, bias=True)
                  (1): Dropout(p=0.0, inplace=False)
                )
              )
              (norm3): LayerNorm((320,), eps=1e-05, elementwise_affine=True)
              (ff): FeedForward(
                (net): ModuleList(
                  (0): GEGLU(
                    (proj): LoRACompatibleLinear(in_features=320, out_features=2560, bias=True)
                  )
                  (1): Dropout(p=0.0, inplace=False)
                  (2): LoRACompatibleLinear(in_features=1280, out_features=320, bias=True)
                )
              )
            )
          )
          (proj_out): LoRACompatibleLinear(in_features=320, out_features=320, bias=True)
        )
      )
      (resnets): ModuleList(
        (0): ResnetBlock2D(
          (norm1): GroupNorm(32, 960, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(960, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=320, bias=True)
          (norm2): GroupNorm(32, 320, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(960, 320, kernel_size=(1, 1), stride=(1, 1))
        )
        (1-2): 2 x ResnetBlock2D(
          (norm1): GroupNorm(32, 640, eps=1e-05, affine=True)
          (conv1): LoRACompatibleConv(640, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=320, bias=True)
          (norm2): GroupNorm(32, 320, eps=1e-05, affine=True)
          (dropout): Dropout(p=0.0, inplace=False)
          (conv2): LoRACompatibleConv(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
          (nonlinearity): SiLU()
          (conv_shortcut): LoRACompatibleConv(640, 320, kernel_size=(1, 1), stride=(1, 1))
        )
      )
    )
  )
  (mid_block): UNetMidBlock2DCrossAttn(
    (attentions): ModuleList(
      (0): Transformer2DModel(
        (norm): GroupNorm(32, 1280, eps=1e-06, affine=True)
        (proj_in): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
        (transformer_blocks): ModuleList(
          (0): BasicTransformerBlock(
            (norm1): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
            (attn1): Attention(
              (to_q): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
              (to_k): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
              (to_v): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
              (to_out): ModuleList(
                (0): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
                (1): Dropout(p=0.0, inplace=False)
              )
            )
            (norm2): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
            (attn2): Attention(
              (to_q): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=False)
              (to_k): LoRACompatibleLinear(in_features=1024, out_features=1280, bias=False)
              (to_v): LoRACompatibleLinear(in_features=1024, out_features=1280, bias=False)
              (to_out): ModuleList(
                (0): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
                (1): Dropout(p=0.0, inplace=False)
              )
            )
            (norm3): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)
            (ff): FeedForward(
              (net): ModuleList(
                (0): GEGLU(
                  (proj): LoRACompatibleLinear(in_features=1280, out_features=10240, bias=True)
                )
                (1): Dropout(p=0.0, inplace=False)
                (2): LoRACompatibleLinear(in_features=5120, out_features=1280, bias=True)
              )
            )
          )
        )
        (proj_out): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
      )
    )
    (resnets): ModuleList(
      (0-1): 2 x ResnetBlock2D(
        (norm1): GroupNorm(32, 1280, eps=1e-05, affine=True)
        (conv1): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (time_emb_proj): LoRACompatibleLinear(in_features=1280, out_features=1280, bias=True)
        (norm2): GroupNorm(32, 1280, eps=1e-05, affine=True)
        (dropout): Dropout(p=0.0, inplace=False)
        (conv2): LoRACompatibleConv(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        (nonlinearity): SiLU()
      )
    )
  )
  (conv_norm_out): GroupNorm(32, 320, eps=1e-05, affine=True)
  (conv_act): SiLU()
  (conv_out): Conv2d(320, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)

長い。

次回

秋田死す

デュエルスタンバイ！