{ "format": "tensorviz.graph", "version": 3, "exportedAt": "2026-09-17T06:47:42.907335+00:00", "status": "static", "warningCount": 0, "document": { "schemaVersion": 2, "semantic": { "graph_id": "snapshot", "name": "TinyDDPM", "revision": 0, "nodes": [ { "node_id": "input", "kind": "input", "op": "Input", "params": { "shape": "1,1,8,8" }, "in_ports": [], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "input", "name": "Input", "displayName": "Input", "description": "Model input node", "icon": "ArrowRight", "category": "io", "parameters": [ { "name": "shape", "type": "string", "defaultValue": "1,3,224,224", "required": true, "description": "Input tensor shape (comma-separated, e.g., 1,3,224,224)" }, { "name": "name", "type": "string", "defaultValue": "input", "required": false, "description": "Name for this input" } ] } } }, { "node_id": "corrupt", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "moduleInstanceId": "corrupt", "moduleTemplateId": "model::ForwardNoise", "isGroup": true, "moduleMembers": [ "python@corrupt#0", "mul@corrupt#0", "mul@corrupt#1", "add@corrupt#0" ], "moduleLayerCount": 4 } }, { "node_id": "denoiser", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "moduleInstanceId": "denoiser", "moduleTemplateId": "model::NoisePredictor", "isGroup": true, "moduleMembers": [ "denoiser.time_embedding", "denoiser.stem", "add@denoiser#0", "silu@denoiser#0", "denoiser.down", "silu@denoiser#1", "add@denoiser#1", "denoiser.bottleneck", "silu@denoiser#2", "denoiser.up", "cat@denoiser#0", "denoiser.merge", "silu@denoiser#3", "denoiser.noise" ], "moduleLayerCount": 14 } }, { "node_id": "output", "kind": "output", "op": "Output", "params": {}, "in_ports": [ "x" ], "out_ports": [], "metadata": { "layerType": { "id": "output", "name": "Output", "displayName": "Output", "description": "Model output node", "icon": "ArrowLeft", "category": "io", "parameters": [ { "name": "name", "type": "string", "defaultValue": "output", "required": false, "description": "Name for this output" } ] } } }, { "node_id": "python@corrupt#0", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "corrupt" } }, { "node_id": "mul@corrupt#0", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "corrupt" } }, { "node_id": "mul@corrupt#1", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "corrupt" } }, { "node_id": "add@corrupt#0", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "corrupt" } }, { "node_id": "denoiser.time_embedding", "kind": "layer", "op": "Embedding", "params": { "num_embeddings": 13, "embedding_dim": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "embedding", "name": "Embedding", "displayName": "Embedding", "description": "Lookup table for dense vector embeddings", "icon": "Hash", "category": "embedding", "parameters": [ { "name": "num_embeddings", "type": "number", "defaultValue": 10000, "required": true, "description": "Size of the dictionary of embeddings", "min": 1 }, { "name": "embedding_dim", "type": "number", "defaultValue": 128, "required": true, "description": "Size of each embedding vector", "min": 1 }, { "name": "padding_idx", "type": "number", "defaultValue": 0, "required": false, "description": "Pad output with zeros at this index", "min": 0 }, { "name": "max_norm", "type": "number", "defaultValue": 0, "required": false, "description": "If > 0, re-normalizes embeddings with norm > max_norm", "min": 0 }, { "name": "sparse", "type": "boolean", "defaultValue": false, "required": false, "description": "If True, gradient w.r.t. weight is sparse" } ] }, "parentId": "denoiser" } }, { "node_id": "denoiser.stem", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 1, "out_channels": 8, "kernel_size": 3, "padding": 1 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "conv2d", "name": "Conv2d", "displayName": "Conv2D", "description": "2D Convolutional Layer", "icon": "Grid3X3", "category": "conv", "parameters": [ { "name": "in_channels", "type": "number", "defaultValue": 3, "required": true, "description": "Number of input channels", "min": 1 }, { "name": "out_channels", "type": "number", "defaultValue": 64, "required": true, "description": "Number of output channels", "min": 1 }, { "name": "kernel_size", "type": "number", "defaultValue": 3, "required": true, "description": "Size of the convolving kernel", "min": 1 }, { "name": "stride", "type": "number", "defaultValue": 1, "required": false, "description": "Stride of the convolution", "min": 1 }, { "name": "padding", "type": "number", "defaultValue": 0, "required": false, "description": "Zero-padding added to both sides", "min": 0 }, { "name": "dilation", "type": "number", "defaultValue": 1, "required": false, "description": "Spacing between kernel elements", "min": 1 }, { "name": "groups", "type": "number", "defaultValue": 1, "required": false, "description": "Number of blocked connections from input to output channels", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, adds a learnable bias" } ] }, "parentId": "denoiser" } }, { "node_id": "add@denoiser#0", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "denoiser" } }, { "node_id": "silu@denoiser#0", "kind": "layer", "op": "SiLU", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "silu", "name": "SiLU", "displayName": "SiLU", "description": "Sigmoid Linear Unit (Swish)", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "inplace", "type": "boolean", "defaultValue": false, "required": false, "description": "Can optionally do the operation in-place" } ] }, "parentId": "denoiser" } }, { "node_id": "denoiser.down", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 8, "out_channels": 8, "kernel_size": 3, "stride": 2, "padding": 1 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "conv2d", "name": "Conv2d", "displayName": "Conv2D", "description": "2D Convolutional Layer", "icon": "Grid3X3", "category": "conv", "parameters": [ { "name": "in_channels", "type": "number", "defaultValue": 3, "required": true, "description": "Number of input channels", "min": 1 }, { "name": "out_channels", "type": "number", "defaultValue": 64, "required": true, "description": "Number of output channels", "min": 1 }, { "name": "kernel_size", "type": "number", "defaultValue": 3, "required": true, "description": "Size of the convolving kernel", "min": 1 }, { "name": "stride", "type": "number", "defaultValue": 1, "required": false, "description": "Stride of the convolution", "min": 1 }, { "name": "padding", "type": "number", "defaultValue": 0, "required": false, "description": "Zero-padding added to both sides", "min": 0 }, { "name": "dilation", "type": "number", "defaultValue": 1, "required": false, "description": "Spacing between kernel elements", "min": 1 }, { "name": "groups", "type": "number", "defaultValue": 1, "required": false, "description": "Number of blocked connections from input to output channels", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, adds a learnable bias" } ] }, "parentId": "denoiser" } }, { "node_id": "silu@denoiser#1", "kind": "layer", "op": "SiLU", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "silu", "name": "SiLU", "displayName": "SiLU", "description": "Sigmoid Linear Unit (Swish)", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "inplace", "type": "boolean", "defaultValue": false, "required": false, "description": "Can optionally do the operation in-place" } ] }, "parentId": "denoiser" } }, { "node_id": "add@denoiser#1", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "denoiser" } }, { "node_id": "denoiser.bottleneck", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 8, "out_channels": 8, "kernel_size": 3, "padding": 1 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "conv2d", "name": "Conv2d", "displayName": "Conv2D", "description": "2D Convolutional Layer", "icon": "Grid3X3", "category": "conv", "parameters": [ { "name": "in_channels", "type": "number", "defaultValue": 3, "required": true, "description": "Number of input channels", "min": 1 }, { "name": "out_channels", "type": "number", "defaultValue": 64, "required": true, "description": "Number of output channels", "min": 1 }, { "name": "kernel_size", "type": "number", "defaultValue": 3, "required": true, "description": "Size of the convolving kernel", "min": 1 }, { "name": "stride", "type": "number", "defaultValue": 1, "required": false, "description": "Stride of the convolution", "min": 1 }, { "name": "padding", "type": "number", "defaultValue": 0, "required": false, "description": "Zero-padding added to both sides", "min": 0 }, { "name": "dilation", "type": "number", "defaultValue": 1, "required": false, "description": "Spacing between kernel elements", "min": 1 }, { "name": "groups", "type": "number", "defaultValue": 1, "required": false, "description": "Number of blocked connections from input to output channels", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, adds a learnable bias" } ] }, "parentId": "denoiser" } }, { "node_id": "silu@denoiser#2", "kind": "layer", "op": "SiLU", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "silu", "name": "SiLU", "displayName": "SiLU", "description": "Sigmoid Linear Unit (Swish)", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "inplace", "type": "boolean", "defaultValue": false, "required": false, "description": "Can optionally do the operation in-place" } ] }, "parentId": "denoiser" } }, { "node_id": "denoiser.up", "kind": "layer", "op": "ConvTranspose2d", "params": { "in_channels": 8, "out_channels": 8, "kernel_size": 2, "stride": 2 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "convtranspose2d", "name": "ConvTranspose2d", "displayName": "ConvTranspose2D", "description": "2D Transposed Convolutional Layer", "icon": "ChevronsRight", "category": "conv", "parameters": [ { "name": "in_channels", "type": "number", "defaultValue": 64, "required": true, "description": "Number of input channels", "min": 1 }, { "name": "out_channels", "type": "number", "defaultValue": 32, "required": true, "description": "Number of output channels", "min": 1 }, { "name": "kernel_size", "type": "number", "defaultValue": 3, "required": true, "description": "Size of the convolving kernel", "min": 1 }, { "name": "stride", "type": "number", "defaultValue": 1, "required": false, "description": "Stride of the convolution", "min": 1 }, { "name": "padding", "type": "number", "defaultValue": 0, "required": false, "description": "Zero-padding added to input", "min": 0 }, { "name": "output_padding", "type": "number", "defaultValue": 0, "required": false, "description": "Additional size added to one side of the output shape", "min": 0 }, { "name": "groups", "type": "number", "defaultValue": 1, "required": false, "description": "Number of blocked connections", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, adds a learnable bias" } ] }, "parentId": "denoiser" } }, { "node_id": "cat@denoiser#0", "kind": "layer", "op": "torch.cat", "params": { "dim": 1 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "cat", "name": "torch.cat", "displayName": "torch.cat", "description": "PyTorch cat operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "denoiser" } }, { "node_id": "denoiser.merge", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 16, "out_channels": 8, "kernel_size": 3, "padding": 1 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "conv2d", "name": "Conv2d", "displayName": "Conv2D", "description": "2D Convolutional Layer", "icon": "Grid3X3", "category": "conv", "parameters": [ { "name": "in_channels", "type": "number", "defaultValue": 3, "required": true, "description": "Number of input channels", "min": 1 }, { "name": "out_channels", "type": "number", "defaultValue": 64, "required": true, "description": "Number of output channels", "min": 1 }, { "name": "kernel_size", "type": "number", "defaultValue": 3, "required": true, "description": "Size of the convolving kernel", "min": 1 }, { "name": "stride", "type": "number", "defaultValue": 1, "required": false, "description": "Stride of the convolution", "min": 1 }, { "name": "padding", "type": "number", "defaultValue": 0, "required": false, "description": "Zero-padding added to both sides", "min": 0 }, { "name": "dilation", "type": "number", "defaultValue": 1, "required": false, "description": "Spacing between kernel elements", "min": 1 }, { "name": "groups", "type": "number", "defaultValue": 1, "required": false, "description": "Number of blocked connections from input to output channels", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, adds a learnable bias" } ] }, "parentId": "denoiser" } }, { "node_id": "silu@denoiser#3", "kind": "layer", "op": "SiLU", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "silu", "name": "SiLU", "displayName": "SiLU", "description": "Sigmoid Linear Unit (Swish)", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "inplace", "type": "boolean", "defaultValue": false, "required": false, "description": "Can optionally do the operation in-place" } ] }, "parentId": "denoiser" } }, { "node_id": "denoiser.noise", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 8, "out_channels": 1, "kernel_size": 3, "padding": 1 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "conv2d", "name": "Conv2d", "displayName": "Conv2D", "description": "2D Convolutional Layer", "icon": "Grid3X3", "category": "conv", "parameters": [ { "name": "in_channels", "type": "number", "defaultValue": 3, "required": true, "description": "Number of input channels", "min": 1 }, { "name": "out_channels", "type": "number", "defaultValue": 64, "required": true, "description": "Number of output channels", "min": 1 }, { "name": "kernel_size", "type": "number", "defaultValue": 3, "required": true, "description": "Size of the convolving kernel", "min": 1 }, { "name": "stride", "type": "number", "defaultValue": 1, "required": false, "description": "Stride of the convolution", "min": 1 }, { "name": "padding", "type": "number", "defaultValue": 0, "required": false, "description": "Zero-padding added to both sides", "min": 0 }, { "name": "dilation", "type": "number", "defaultValue": 1, "required": false, "description": "Spacing between kernel elements", "min": 1 }, { "name": "groups", "type": "number", "defaultValue": 1, "required": false, "description": "Number of blocked connections from input to output channels", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, adds a learnable bias" } ] }, "parentId": "denoiser" } } ], "edges": [ { "edge_id": "edge-1", "source_node_id": "python@corrupt#0", "target_node_id": "mul@corrupt#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-2", "source_node_id": "python@corrupt#0", "target_node_id": "mul@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-3", "source_node_id": "mul@corrupt#1", "target_node_id": "add@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-4", "source_node_id": "mul@corrupt#0", "target_node_id": "add@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-5", "source_node_id": "denoiser.stem", "target_node_id": "add@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-6", "source_node_id": "denoiser.time_embedding", "target_node_id": "add@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-7", "source_node_id": "add@denoiser#0", "target_node_id": "silu@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-8", "source_node_id": "silu@denoiser#0", "target_node_id": "denoiser.down", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-9", "source_node_id": "denoiser.down", "target_node_id": "silu@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-10", "source_node_id": "silu@denoiser#1", "target_node_id": "add@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-11", "source_node_id": "denoiser.time_embedding", "target_node_id": "add@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-12", "source_node_id": "add@denoiser#1", "target_node_id": "denoiser.bottleneck", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-13", "source_node_id": "denoiser.bottleneck", "target_node_id": "silu@denoiser#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-14", "source_node_id": "silu@denoiser#2", "target_node_id": "denoiser.up", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-15", "source_node_id": "silu@denoiser#0", "target_node_id": "cat@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-16", "source_node_id": "denoiser.up", "target_node_id": "cat@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-17", "source_node_id": "cat@denoiser#0", "target_node_id": "denoiser.merge", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-18", "source_node_id": "denoiser.merge", "target_node_id": "silu@denoiser#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-19", "source_node_id": "silu@denoiser#3", "target_node_id": "denoiser.noise", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-20", "source_node_id": "input", "target_node_id": "python@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-21", "source_node_id": "input", "target_node_id": "mul@corrupt#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-22", "source_node_id": "input", "target_node_id": "mul@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-23", "source_node_id": "input", "target_node_id": "corrupt", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-24", "source_node_id": "add@corrupt#0", "target_node_id": "denoiser.time_embedding", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-25", "source_node_id": "add@corrupt#0", "target_node_id": "denoiser.stem", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-26", "source_node_id": "corrupt", "target_node_id": "denoiser", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-27", "source_node_id": "input", "target_node_id": "denoiser.time_embedding", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-28", "source_node_id": "input", "target_node_id": "denoiser.stem", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-29", "source_node_id": "input", "target_node_id": "denoiser", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-30", "source_node_id": "denoiser.noise", "target_node_id": "output", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-31", "source_node_id": "denoiser", "target_node_id": "output", "source_port": "y", "target_port": "x", "metadata": {} } ], "inputs": [], "modules": [], "operations": [], "metadata": {} }, "view": { "graph_id": "snapshot", "node_positions": { "input": { "x": 0, "y": 160 }, "corrupt": { "x": 340, "y": 160 }, "denoiser": { "x": 680, "y": 160 }, "output": { "x": 1020, "y": 160 }, "python@corrupt#0": { "x": 24, "y": 214 }, "mul@corrupt#0": { "x": 24, "y": 364 }, "mul@corrupt#1": { "x": 24, "y": 514 }, "add@corrupt#0": { "x": 24, "y": 664 }, "denoiser.time_embedding": { "x": 24, "y": 214 }, "denoiser.stem": { "x": 24, "y": 364 }, "add@denoiser#0": { "x": 24, "y": 514 }, "silu@denoiser#0": { "x": 24, "y": 664 }, "denoiser.down": { "x": 24, "y": 814 }, "silu@denoiser#1": { "x": 24, "y": 964 }, "add@denoiser#1": { "x": 24, "y": 1114 }, "denoiser.bottleneck": { "x": 24, "y": 1264 }, "silu@denoiser#2": { "x": 24, "y": 1414 }, "denoiser.up": { "x": 24, "y": 1564 }, "cat@denoiser#0": { "x": 24, "y": 1714 }, "denoiser.merge": { "x": 24, "y": 1864 }, "silu@denoiser#3": { "x": 24, "y": 2014 }, "denoiser.noise": { "x": 24, "y": 2164 } }, "collapsed_modules": [ "corrupt", "denoiser" ], "selection": [], "layout": {} } }, "snapshotId": "snapshot-350f57c36906474d", "collaboration": { "walkthrough": { "title": "Learn to reverse a noise schedule", "steps": [ { "id": "corrupt", "title": "Mix a clean image with known Gaussian noise", "note": "The fixed forward process samples xₜ = √ᾱₜ x₀ + √(1−ᾱₜ) ε. A training example can jump directly to any timestep; it does not need to run every preceding noising step.", "view": { "camera": { "x": 263.0769230769231, "y": -70.04195804195803, "z": 0.7114427860696517 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "corrupt": { "x": 340, "y": 160 }, "denoiser": { "x": 680, "y": 160 }, "output": { "x": 1020, "y": 160 }, "python@corrupt#0": { "x": 24, "y": 214 }, "mul@corrupt#0": { "x": 24, "y": 364 }, "mul@corrupt#1": { "x": 24, "y": 514 }, "add@corrupt#0": { "x": 24, "y": 664 }, "denoiser.time_embedding": { "x": 24, "y": 214 }, "denoiser.stem": { "x": 24, "y": 364 }, "add@denoiser#0": { "x": 24, "y": 514 }, "silu@denoiser#0": { "x": 24, "y": 664 }, "denoiser.down": { "x": 24, "y": 814 }, "silu@denoiser#1": { "x": 24, "y": 964 }, "add@denoiser#1": { "x": 24, "y": 1114 }, "denoiser.bottleneck": { "x": 24, "y": 1264 }, "silu@denoiser#2": { "x": 24, "y": 1414 }, "denoiser.up": { "x": 24, "y": 1564 }, "cat@denoiser#0": { "x": 24, "y": 1714 }, "denoiser.merge": { "x": 24, "y": 1864 }, "silu@denoiser#3": { "x": 24, "y": 2014 }, "denoiser.noise": { "x": 24, "y": 2164 } }, "expandedModuleIds": [ "corrupt" ], "selectedNodeIds": [ "corrupt" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "corrupt" }, "snapshotId": "snapshot-350f57c36906474d", "exportedAt": "2026-09-17T06:47:42.907335+00:00", "modelName": "TinyDDPM", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.stem\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.time_embedding\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"corrupt\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"corrupt\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#1\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"mul@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"mul@corrupt#1\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"python@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"python@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#1\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[\"{\\\"id\\\":\\\"add@corrupt#0\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"torch.add\\\",\\\"params\\\":{},\\\"parent\\\":\\\"corrupt\\\"}\",\"{\\\"id\\\":\\\"mul@corrupt#0\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"torch.mul\\\",\\\"params\\\":{},\\\"parent\\\":\\\"corrupt\\\"}\",\"{\\\"id\\\":\\\"mul@corrupt#1\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"torch.mul\\\",\\\"params\\\":{},\\\"parent\\\":\\\"corrupt\\\"}\",\"{\\\"id\\\":\\\"python@corrupt#0\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"Python\\\",\\\"params\\\":{},\\\"parent\\\":\\\"corrupt\\\"}\"],\"node\":{\"id\":\"corrupt\",\"kind\":\"module\",\"op\":\"Module\",\"params\":{}}}" } ] }, { "id": "time", "title": "Tell the denoiser which noise level it sees", "note": "The selected timestep becomes eight learned features, added to fine and coarse image features. The same network weights serve every timestep.", "view": { "camera": { "x": 308.5, "y": 6, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "corrupt": { "x": 340, "y": 160 }, "denoiser": { "x": 680, "y": 160 }, "output": { "x": 1020, "y": 160 }, "python@corrupt#0": { "x": 24, "y": 214 }, "mul@corrupt#0": { "x": 24, "y": 364 }, "mul@corrupt#1": { "x": 24, "y": 514 }, "add@corrupt#0": { "x": 24, "y": 664 }, "denoiser.time_embedding": { "x": 24, "y": 214 }, "denoiser.stem": { "x": 24, "y": 364 }, "add@denoiser#0": { "x": 24, "y": 514 }, "silu@denoiser#0": { "x": 24, "y": 664 }, "denoiser.down": { "x": 24, "y": 814 }, "silu@denoiser#1": { "x": 24, "y": 964 }, "add@denoiser#1": { "x": 24, "y": 1114 }, "denoiser.bottleneck": { "x": 24, "y": 1264 }, "silu@denoiser#2": { "x": 24, "y": 1414 }, "denoiser.up": { "x": 24, "y": 1564 }, "cat@denoiser#0": { "x": 24, "y": 1714 }, "denoiser.merge": { "x": 24, "y": 1864 }, "silu@denoiser#3": { "x": 24, "y": 2014 }, "denoiser.noise": { "x": 24, "y": 2164 } }, "expandedModuleIds": [ "denoiser" ], "selectedNodeIds": [ "denoiser.time_embedding" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "denoiser.time_embedding" }, "snapshotId": "snapshot-350f57c36906474d", "exportedAt": "2026-09-17T06:47:42.907335+00:00", "modelName": "TinyDDPM", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.time_embedding\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"denoiser.time_embedding\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@denoiser#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"denoiser.time_embedding\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@denoiser#1\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.time_embedding\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"denoiser.time_embedding\",\"kind\":\"layer\",\"op\":\"Embedding\",\"params\":{\"embedding_dim\":8,\"num_embeddings\":13},\"parent\":\"denoiser\"}}" } ] }, { "id": "predict", "title": "Combine coarse context with a fine-resolution skip", "note": "The tiny U-Net contracts from 8 × 8 to 4 × 4 and expands again. Concatenation preserves fine features while the bottleneck supplies broader context.", "view": { "camera": { "x": 308.5, "y": -1057.5, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "corrupt": { "x": 340, "y": 160 }, "denoiser": { "x": 680, "y": 160 }, "output": { "x": 1020, "y": 160 }, "python@corrupt#0": { "x": 24, "y": 214 }, "mul@corrupt#0": { "x": 24, "y": 364 }, "mul@corrupt#1": { "x": 24, "y": 514 }, "add@corrupt#0": { "x": 24, "y": 664 }, "denoiser.time_embedding": { "x": 24, "y": 214 }, "denoiser.stem": { "x": 24, "y": 364 }, "add@denoiser#0": { "x": 24, "y": 514 }, "silu@denoiser#0": { "x": 24, "y": 664 }, "denoiser.down": { "x": 24, "y": 814 }, "silu@denoiser#1": { "x": 24, "y": 964 }, "add@denoiser#1": { "x": 24, "y": 1114 }, "denoiser.bottleneck": { "x": 24, "y": 1264 }, "silu@denoiser#2": { "x": 24, "y": 1414 }, "denoiser.up": { "x": 24, "y": 1564 }, "cat@denoiser#0": { "x": 24, "y": 1714 }, "denoiser.merge": { "x": 24, "y": 1864 }, "silu@denoiser#3": { "x": 24, "y": 2014 }, "denoiser.noise": { "x": 24, "y": 2164 } }, "expandedModuleIds": [ "denoiser" ], "selectedNodeIds": [ "denoiser.bottleneck" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "denoiser.bottleneck" }, "snapshotId": "snapshot-350f57c36906474d", "exportedAt": "2026-09-17T06:47:42.907335+00:00", "modelName": "TinyDDPM", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@denoiser#1\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.bottleneck\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"denoiser.bottleneck\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"silu@denoiser#2\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"denoiser.bottleneck\",\"kind\":\"layer\",\"op\":\"Conv2d\",\"params\":{\"in_channels\":8,\"kernel_size\":3,\"out_channels\":8,\"padding\":1},\"parent\":\"denoiser\"}}" } ] }, { "id": "reverse", "title": "Predict epsilon, then reuse the network in reverse", "note": "Training minimizes MSE against the known added noise. Sampling starts from Gaussian noise and converts each prediction into the reverse mean, adding posterior noise except at the final step. Choose a reverse-chain checkpoint in the panel.", "view": { "camera": { "x": 308.5, "y": -1957.5, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "corrupt": { "x": 340, "y": 160 }, "denoiser": { "x": 680, "y": 160 }, "output": { "x": 1020, "y": 160 }, "python@corrupt#0": { "x": 24, "y": 214 }, "mul@corrupt#0": { "x": 24, "y": 364 }, "mul@corrupt#1": { "x": 24, "y": 514 }, "add@corrupt#0": { "x": 24, "y": 664 }, "denoiser.time_embedding": { "x": 24, "y": 214 }, "denoiser.stem": { "x": 24, "y": 364 }, "add@denoiser#0": { "x": 24, "y": 514 }, "silu@denoiser#0": { "x": 24, "y": 664 }, "denoiser.down": { "x": 24, "y": 814 }, "silu@denoiser#1": { "x": 24, "y": 964 }, "add@denoiser#1": { "x": 24, "y": 1114 }, "denoiser.bottleneck": { "x": 24, "y": 1264 }, "silu@denoiser#2": { "x": 24, "y": 1414 }, "denoiser.up": { "x": 24, "y": 1564 }, "cat@denoiser#0": { "x": 24, "y": 1714 }, "denoiser.merge": { "x": 24, "y": 1864 }, "silu@denoiser#3": { "x": 24, "y": 2014 }, "denoiser.noise": { "x": 24, "y": 2164 } }, "expandedModuleIds": [ "denoiser" ], "selectedNodeIds": [ "denoiser.noise" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "denoiser.noise" }, "snapshotId": "snapshot-350f57c36906474d", "exportedAt": "2026-09-17T06:47:42.907335+00:00", "modelName": "TinyDDPM", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"denoiser.noise\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"output\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"silu@denoiser#3\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.noise\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"denoiser.noise\",\"kind\":\"layer\",\"op\":\"Conv2d\",\"params\":{\"in_channels\":8,\"kernel_size\":3,\"out_channels\":1,\"padding\":1},\"parent\":\"denoiser\"}}" } ] } ] }, "threads": [], "savedViews": [ { "id": "corrupt", "title": "Mix a clean image with known Gaussian noise", "note": "The fixed forward process samples xₜ = √ᾱₜ x₀ + √(1−ᾱₜ) ε. A training example can jump directly to any timestep; it does not need to run every preceding noising step.", "view": { "camera": { "x": 263.0769230769231, "y": -70.04195804195803, "z": 0.7114427860696517 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "corrupt": { "x": 340, "y": 160 }, "denoiser": { "x": 680, "y": 160 }, "output": { "x": 1020, "y": 160 }, "python@corrupt#0": { "x": 24, "y": 214 }, "mul@corrupt#0": { "x": 24, "y": 364 }, "mul@corrupt#1": { "x": 24, "y": 514 }, "add@corrupt#0": { "x": 24, "y": 664 }, "denoiser.time_embedding": { "x": 24, "y": 214 }, "denoiser.stem": { "x": 24, "y": 364 }, "add@denoiser#0": { "x": 24, "y": 514 }, "silu@denoiser#0": { "x": 24, "y": 664 }, "denoiser.down": { "x": 24, "y": 814 }, "silu@denoiser#1": { "x": 24, "y": 964 }, "add@denoiser#1": { "x": 24, "y": 1114 }, "denoiser.bottleneck": { "x": 24, "y": 1264 }, "silu@denoiser#2": { "x": 24, "y": 1414 }, "denoiser.up": { "x": 24, "y": 1564 }, "cat@denoiser#0": { "x": 24, "y": 1714 }, "denoiser.merge": { "x": 24, "y": 1864 }, "silu@denoiser#3": { "x": 24, "y": 2014 }, "denoiser.noise": { "x": 24, "y": 2164 } }, "expandedModuleIds": [ "corrupt" ], "selectedNodeIds": [ "corrupt" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "corrupt" }, "snapshotId": "snapshot-350f57c36906474d", "exportedAt": "2026-09-17T06:47:42.907335+00:00", "modelName": "TinyDDPM", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.stem\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.time_embedding\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"corrupt\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"corrupt\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#1\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"mul@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"mul@corrupt#1\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"python@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"python@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#1\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[\"{\\\"id\\\":\\\"add@corrupt#0\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"torch.add\\\",\\\"params\\\":{},\\\"parent\\\":\\\"corrupt\\\"}\",\"{\\\"id\\\":\\\"mul@corrupt#0\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"torch.mul\\\",\\\"params\\\":{},\\\"parent\\\":\\\"corrupt\\\"}\",\"{\\\"id\\\":\\\"mul@corrupt#1\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"torch.mul\\\",\\\"params\\\":{},\\\"parent\\\":\\\"corrupt\\\"}\",\"{\\\"id\\\":\\\"python@corrupt#0\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"Python\\\",\\\"params\\\":{},\\\"parent\\\":\\\"corrupt\\\"}\"],\"node\":{\"id\":\"corrupt\",\"kind\":\"module\",\"op\":\"Module\",\"params\":{}}}" } ] } ], "openingViewId": "corrupt" } }