{
  "id": "diffusion-unet",
  "name": "Diffusion UNet (Stable-Diffusion-style)",
  "description": "Image generator backbone — predicts the noise to subtract at each denoising step, conditioned on a text embedding via cross-attention. Latent-space UNet (4 channels at 64×64 = 512×512 RGB after VAE decode). Compressed for legibility; real SD has 4 down + 1 mid + 4 up stacks.",
  "icon": "🎨",
  "category": "Generative",
  "components": [
    {
      "id": "input-1",
      "type": "input",
      "name": "noisy_latent",
      "position": { "x": 200, "y": 50 },
      "params": { "shape": [4, 64, 64] },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "conv-in",
      "type": "conv2d",
      "name": "conv_in",
      "scope": "stem",
      "position": { "x": 200, "y": 220 },
      "params": { "outChannels": 320, "kernelSize": 3, "stride": 1, "padding": 1 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "down1-norm",
      "type": "groupNorm",
      "name": "down1_norm",
      "scope": "down_blocks.0",
      "position": { "x": 200, "y": 390 },
      "params": { "numGroups": 32, "numChannels": 320 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "down1-conv",
      "type": "conv2d",
      "name": "down1_conv",
      "scope": "down_blocks.0",
      "position": { "x": 200, "y": 560 },
      "params": { "outChannels": 320, "kernelSize": 3, "stride": 1, "padding": 1 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "down1-silu",
      "type": "swish",
      "name": "down1_silu",
      "scope": "down_blocks.0",
      "position": { "x": 200, "y": 730 },
      "params": {},
      "inputs": [],
      "outputs": []
    },
    {
      "id": "down1-to-tokens",
      "type": "reshape",
      "name": "to_tokens",
      "scope": "down_blocks.0",
      "position": { "x": 200, "y": 860 },
      "params": { "shape": [4096, 320] },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "down1-xattn",
      "type": "crossAttention",
      "name": "down1_text_attn",
      "scope": "down_blocks.0",
      "position": { "x": 200, "y": 900 },
      "params": { "embedDim": 320, "numHeads": 8, "kvDim": 768 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "down1-to-map",
      "type": "reshape",
      "name": "to_feature_map",
      "scope": "down_blocks.0",
      "position": { "x": 200, "y": 940 },
      "params": { "shape": [320, 64, 64] },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "downsample-1",
      "type": "conv2d",
      "name": "downsample_1",
      "scope": "down_blocks.0",
      "position": { "x": 200, "y": 1080 },
      "params": { "outChannels": 640, "kernelSize": 3, "stride": 2, "padding": 1 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "mid-norm",
      "type": "groupNorm",
      "name": "mid_norm",
      "scope": "mid_block",
      "position": { "x": 200, "y": 1260 },
      "params": { "numGroups": 32, "numChannels": 640 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "mid-to-tokens",
      "type": "reshape",
      "name": "to_tokens",
      "scope": "mid_block",
      "position": { "x": 200, "y": 1160 },
      "params": { "shape": [1024, 640] },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "mid-xattn",
      "type": "crossAttention",
      "name": "mid_text_attn",
      "scope": "mid_block",
      "position": { "x": 200, "y": 1430 },
      "params": { "embedDim": 640, "numHeads": 8, "kvDim": 768 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "mid-to-map",
      "type": "reshape",
      "name": "to_feature_map",
      "scope": "mid_block",
      "position": { "x": 200, "y": 1240 },
      "params": { "shape": [640, 32, 32] },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "upsample-1",
      "type": "upsample",
      "name": "upsample_1",
      "scope": "up_blocks.0",
      "position": { "x": 200, "y": 1610 },
      "params": { "scaleFactor": 2, "mode": "nearest" },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "up1-conv",
      "type": "conv2d",
      "name": "up1_conv",
      "scope": "up_blocks.0",
      "position": { "x": 200, "y": 1790 },
      "params": { "outChannels": 320, "kernelSize": 3, "stride": 1, "padding": 1 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "conv-out-norm",
      "type": "groupNorm",
      "name": "conv_out_norm",
      "scope": "head",
      "position": { "x": 200, "y": 2130 },
      "params": { "numGroups": 32, "numChannels": 320 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "up1-silu",
      "type": "swish",
      "name": "up1_silu",
      "scope": "up_blocks.0",
      "position": { "x": 200, "y": 1960 },
      "params": {},
      "inputs": [],
      "outputs": []
    },
    {
      "id": "conv-out",
      "type": "conv2d",
      "name": "conv_out",
      "scope": "head",
      "position": { "x": 200, "y": 2300 },
      "params": { "outChannels": 4, "kernelSize": 3, "stride": 1, "padding": 1 },
      "inputs": [],
      "outputs": []
    },
    {
      "id": "output-1",
      "type": "output",
      "name": "predicted_noise",
      "position": { "x": 200, "y": 2470 },
      "params": {},
      "inputs": [],
      "outputs": []
    }
  ],
  "connections": []
}