{ "format": "tensorviz.graph", "version": 3, "exportedAt": "2026-09-17T06:51:48.425092+00:00", "status": "static", "warningCount": 0, "document": { "schemaVersion": 2, "semantic": { "graph_id": "snapshot", "name": "TinyLatentDiffusion", "revision": 0, "nodes": [ { "node_id": "input", "kind": "input", "op": "Input", "params": { "shape": "1,3,8,8" }, "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": "encoder", "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": "encoder", "moduleTemplateId": "model::ImageEncoder", "isGroup": true, "moduleMembers": [ "encoder.down", "silu@encoder#0", "encoder.latent" ], "moduleLayerCount": 3 } }, { "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::LatentNoise", "isGroup": true, "moduleMembers": [ "python@corrupt#0", "mul@corrupt#0", "mul@corrupt#1", "add@corrupt#0" ], "moduleLayerCount": 4 } }, { "node_id": "python@TinyLatentDiffusion#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": [] } } }, { "node_id": "conditioner", "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": "conditioner", "moduleTemplateId": "model::TextConditioner", "isGroup": true, "moduleMembers": [ "conditioner.tokens", "conditioner.positions", "add@conditioner#0" ], "moduleLayerCount": 3 } }, { "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::ConditionalDenoiser", "isGroup": true, "moduleMembers": [ "denoiser.time_embedding", "denoiser.stem", "add@denoiser#0", "silu@denoiser#0", "denoiser.down", "silu@denoiser#1", "flatten@denoiser#0", "python@denoiser#0", "denoiser.cross_attention", "add@denoiser#1", "python@denoiser#1", "denoiser.up", "cat@denoiser#0", "denoiser.merge", "silu@denoiser#2", "denoiser.noise" ], "moduleLayerCount": 21 } }, { "node_id": "estimate", "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": "estimate", "moduleTemplateId": "model::EstimateCleanLatent", "isGroup": true, "moduleMembers": [], "moduleLayerCount": 0 } }, { "node_id": "decoder", "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": "decoder", "moduleTemplateId": "model::ImageDecoder", "isGroup": true, "moduleMembers": [ "decoder.up", "silu@decoder#0", "decoder.rgb" ], "moduleLayerCount": 3 } }, { "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": "encoder.down", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 3, "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": "encoder" } }, { "node_id": "silu@encoder#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": "encoder" } }, { "node_id": "encoder.latent", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 8, "out_channels": 2, "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": "encoder" } }, { "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": "conditioner.tokens", "kind": "layer", "op": "Embedding", "params": { "num_embeddings": 16, "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": "conditioner" } }, { "node_id": "conditioner.positions", "kind": "layer", "op": "Parameter", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "parameter", "description": "", "category": "misc", "icon": "Box", "parameters": [], "name": "Parameter", "displayName": "Parameter" }, "parentId": "conditioner" } }, { "node_id": "add@conditioner#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": "conditioner" } }, { "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": 2, "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": "flatten@denoiser#0", "kind": "layer", "op": "Flatten", "params": { "start_dim": 2 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "flatten", "name": "Flatten", "displayName": "Flatten", "description": "Flattens input tensor contiguously", "icon": "Maximize2", "category": "tensor", "parameters": [ { "name": "start_dim", "type": "number", "defaultValue": 1, "required": false, "description": "First dim to flatten", "min": 0 }, { "name": "end_dim", "type": "number", "defaultValue": -1, "required": false, "description": "Last dim to flatten" } ] }, "parentId": "denoiser" } }, { "node_id": "python@denoiser#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": "denoiser" } }, { "node_id": "denoiser.cross_attention", "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" } ] }, "parentId": "denoiser", "moduleInstanceId": "denoiser.cross_attention", "moduleTemplateId": "model::CrossAttention", "isGroup": true, "moduleMembers": [ "denoiser.cross_attention.query", "denoiser.cross_attention.key", "div@denoiser.cross_attention#0", "denoiser.cross_attention.value", "denoiser.cross_attention.softmax", "denoiser.cross_attention.project" ], "moduleLayerCount": 6 } }, { "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": "python@denoiser#1", "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": "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#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.noise", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 8, "out_channels": 2, "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": "denoiser.cross_attention.query", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "denoiser.cross_attention" } }, { "node_id": "denoiser.cross_attention.key", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "denoiser.cross_attention" } }, { "node_id": "div@denoiser.cross_attention#0", "kind": "layer", "op": "torch.div", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "div", "name": "torch.div", "displayName": "torch.div", "description": "PyTorch div operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "denoiser.cross_attention" } }, { "node_id": "denoiser.cross_attention.value", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "denoiser.cross_attention" } }, { "node_id": "denoiser.cross_attention.softmax", "kind": "layer", "op": "Softmax", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "softmax", "name": "Softmax", "displayName": "Softmax", "description": "Softmax over a dimension", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "dim", "type": "number", "defaultValue": 1, "required": true, "description": "Dimension along which Softmax is computed" } ] }, "parentId": "denoiser.cross_attention" } }, { "node_id": "denoiser.cross_attention.project", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "denoiser.cross_attention" } }, { "node_id": "decoder.up", "kind": "layer", "op": "ConvTranspose2d", "params": { "in_channels": 2, "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": "decoder" } }, { "node_id": "silu@decoder#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": "decoder" } }, { "node_id": "decoder.rgb", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 8, "out_channels": 3, "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": "decoder" } } ], "edges": [ { "edge_id": "edge-1", "source_node_id": "encoder.down", "target_node_id": "silu@encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-2", "source_node_id": "silu@encoder#0", "target_node_id": "encoder.latent", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-3", "source_node_id": "python@corrupt#0", "target_node_id": "mul@corrupt#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-4", "source_node_id": "python@corrupt#0", "target_node_id": "mul@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-5", "source_node_id": "mul@corrupt#1", "target_node_id": "add@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-6", "source_node_id": "mul@corrupt#0", "target_node_id": "add@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-7", "source_node_id": "conditioner.tokens", "target_node_id": "add@conditioner#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-8", "source_node_id": "conditioner.positions", "target_node_id": "add@conditioner#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-9", "source_node_id": "denoiser.cross_attention.query", "target_node_id": "div@denoiser.cross_attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-10", "source_node_id": "denoiser.cross_attention.key", "target_node_id": "div@denoiser.cross_attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-11", "source_node_id": "div@denoiser.cross_attention#0", "target_node_id": "denoiser.cross_attention.softmax", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-12", "source_node_id": "denoiser.cross_attention.softmax", "target_node_id": "denoiser.cross_attention.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-13", "source_node_id": "denoiser.cross_attention.value", "target_node_id": "denoiser.cross_attention.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-14", "source_node_id": "denoiser.stem", "target_node_id": "add@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-15", "source_node_id": "denoiser.time_embedding", "target_node_id": "add@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-16", "source_node_id": "add@denoiser#0", "target_node_id": "silu@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-17", "source_node_id": "silu@denoiser#0", "target_node_id": "denoiser.down", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-18", "source_node_id": "denoiser.down", "target_node_id": "silu@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-19", "source_node_id": "silu@denoiser#1", "target_node_id": "flatten@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-20", "source_node_id": "flatten@denoiser#0", "target_node_id": "python@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-21", "source_node_id": "python@denoiser#0", "target_node_id": "denoiser.cross_attention.query", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-22", "source_node_id": "python@denoiser#0", "target_node_id": "denoiser.cross_attention.key", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-23", "source_node_id": "python@denoiser#0", "target_node_id": "denoiser.cross_attention.value", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-24", "source_node_id": "python@denoiser#0", "target_node_id": "denoiser.cross_attention", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-25", "source_node_id": "python@denoiser#0", "target_node_id": "add@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-26", "source_node_id": "denoiser.cross_attention.project", "target_node_id": "add@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-27", "source_node_id": "denoiser.cross_attention", "target_node_id": "add@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-28", "source_node_id": "add@denoiser#1", "target_node_id": "python@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-29", "source_node_id": "silu@denoiser#1", "target_node_id": "python@denoiser#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-30", "source_node_id": "python@denoiser#1", "target_node_id": "denoiser.up", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-31", "source_node_id": "silu@denoiser#0", "target_node_id": "cat@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-32", "source_node_id": "denoiser.up", "target_node_id": "cat@denoiser#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-33", "source_node_id": "cat@denoiser#0", "target_node_id": "denoiser.merge", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-34", "source_node_id": "denoiser.merge", "target_node_id": "silu@denoiser#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-35", "source_node_id": "silu@denoiser#2", "target_node_id": "denoiser.noise", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-36", "source_node_id": "decoder.up", "target_node_id": "silu@decoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-37", "source_node_id": "silu@decoder#0", "target_node_id": "decoder.rgb", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-38", "source_node_id": "input", "target_node_id": "encoder.down", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-39", "source_node_id": "input", "target_node_id": "encoder", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-40", "source_node_id": "encoder.latent", "target_node_id": "python@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-41", "source_node_id": "encoder.latent", "target_node_id": "mul@corrupt#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-42", "source_node_id": "encoder.latent", "target_node_id": "mul@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-43", "source_node_id": "encoder", "target_node_id": "corrupt", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-44", "source_node_id": "input", "target_node_id": "python@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-45", "source_node_id": "input", "target_node_id": "mul@corrupt#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-46", "source_node_id": "input", "target_node_id": "mul@corrupt#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-47", "source_node_id": "input", "target_node_id": "corrupt", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-48", "source_node_id": "input", "target_node_id": "conditioner.tokens", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-49", "source_node_id": "input", "target_node_id": "conditioner", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-50", "source_node_id": "add@corrupt#0", "target_node_id": "denoiser.stem", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-51", "source_node_id": "add@corrupt#0", "target_node_id": "denoiser.time_embedding", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-52", "source_node_id": "add@corrupt#0", "target_node_id": "denoiser.cross_attention.query", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-53", "source_node_id": "add@corrupt#0", "target_node_id": "denoiser.cross_attention.key", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-54", "source_node_id": "add@corrupt#0", "target_node_id": "denoiser.cross_attention.value", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-55", "source_node_id": "corrupt", "target_node_id": "denoiser", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-56", "source_node_id": "input", "target_node_id": "denoiser.stem", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-57", "source_node_id": "input", "target_node_id": "denoiser.time_embedding", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-58", "source_node_id": "input", "target_node_id": "denoiser.cross_attention.query", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-59", "source_node_id": "input", "target_node_id": "denoiser.cross_attention.key", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-60", "source_node_id": "input", "target_node_id": "denoiser.cross_attention.value", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-61", "source_node_id": "input", "target_node_id": "denoiser", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-62", "source_node_id": "add@conditioner#0", "target_node_id": "denoiser.stem", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-63", "source_node_id": "add@conditioner#0", "target_node_id": "denoiser.time_embedding", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-64", "source_node_id": "add@conditioner#0", "target_node_id": "denoiser.cross_attention.query", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-65", "source_node_id": "add@conditioner#0", "target_node_id": "denoiser.cross_attention.key", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-66", "source_node_id": "add@conditioner#0", "target_node_id": "denoiser.cross_attention.value", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-67", "source_node_id": "conditioner", "target_node_id": "denoiser", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-68", "source_node_id": "add@corrupt#0", "target_node_id": "python@TinyLatentDiffusion#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-69", "source_node_id": "corrupt", "target_node_id": "python@TinyLatentDiffusion#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-70", "source_node_id": "input", "target_node_id": "python@TinyLatentDiffusion#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-71", "source_node_id": "add@corrupt#0", "target_node_id": "estimate", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-72", "source_node_id": "corrupt", "target_node_id": "estimate", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-73", "source_node_id": "denoiser.noise", "target_node_id": "estimate", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-74", "source_node_id": "denoiser", "target_node_id": "estimate", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-75", "source_node_id": "python@TinyLatentDiffusion#0", "target_node_id": "estimate", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-76", "source_node_id": "estimate", "target_node_id": "decoder.up", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-77", "source_node_id": "estimate", "target_node_id": "decoder", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-78", "source_node_id": "decoder.rgb", "target_node_id": "output", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-79", "source_node_id": "decoder", "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 }, "encoder": { "x": 340, "y": 160 }, "corrupt": { "x": 680, "y": 160 }, "python@TinyLatentDiffusion#0": { "x": 1020, "y": 160 }, "conditioner": { "x": 1360, "y": 160 }, "denoiser": { "x": 1700, "y": 160 }, "estimate": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "encoder.down": { "x": 24, "y": 214 }, "silu@encoder#0": { "x": 24, "y": 364 }, "encoder.latent": { "x": 24, "y": 514 }, "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 }, "conditioner.tokens": { "x": 24, "y": 214 }, "conditioner.positions": { "x": 24, "y": 364 }, "add@conditioner#0": { "x": 24, "y": 514 }, "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 }, "flatten@denoiser#0": { "x": 24, "y": 1114 }, "python@denoiser#0": { "x": 24, "y": 1264 }, "denoiser.cross_attention": { "x": 24, "y": 1414 }, "add@denoiser#1": { "x": 24, "y": 1564 }, "python@denoiser#1": { "x": 24, "y": 1714 }, "denoiser.up": { "x": 24, "y": 1864 }, "cat@denoiser#0": { "x": 24, "y": 2014 }, "denoiser.merge": { "x": 24, "y": 2164 }, "silu@denoiser#2": { "x": 24, "y": 2314 }, "denoiser.noise": { "x": 24, "y": 2464 }, "denoiser.cross_attention.query": { "x": 24, "y": 214 }, "denoiser.cross_attention.key": { "x": 24, "y": 364 }, "div@denoiser.cross_attention#0": { "x": 24, "y": 514 }, "denoiser.cross_attention.value": { "x": 24, "y": 664 }, "denoiser.cross_attention.softmax": { "x": 24, "y": 814 }, "denoiser.cross_attention.project": { "x": 24, "y": 964 }, "decoder.up": { "x": 24, "y": 214 }, "silu@decoder#0": { "x": 24, "y": 364 }, "decoder.rgb": { "x": 24, "y": 514 } }, "collapsed_modules": [ "encoder", "corrupt", "conditioner", "denoiser", "estimate", "decoder", "denoiser.cross_attention" ], "selection": [], "layout": {} } }, "snapshotId": "snapshot-48f0d017e40d4345", "collaboration": { "walkthrough": { "title": "Move denoising into a compressed space", "steps": [ { "id": "compress", "title": "Train diffusion on an encoded image", "note": "The fixed first stage reduces 192 pixel-space values to 32 latent values. This teaching encoder is random, so its compression is a shape fact rather than evidence of retained perceptual quality.", "view": { "camera": { "x": 308.5, "y": -307.5, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "encoder": { "x": 340, "y": 160 }, "corrupt": { "x": 680, "y": 160 }, "python@TinyLatentDiffusion#0": { "x": 1020, "y": 160 }, "conditioner": { "x": 1360, "y": 160 }, "denoiser": { "x": 1700, "y": 160 }, "estimate": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "encoder.down": { "x": 24, "y": 214 }, "silu@encoder#0": { "x": 24, "y": 364 }, "encoder.latent": { "x": 24, "y": 514 }, "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 }, "conditioner.tokens": { "x": 24, "y": 214 }, "conditioner.positions": { "x": 24, "y": 364 }, "add@conditioner#0": { "x": 24, "y": 514 }, "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 }, "flatten@denoiser#0": { "x": 24, "y": 1114 }, "python@denoiser#0": { "x": 24, "y": 1264 }, "denoiser.cross_attention": { "x": 24, "y": 1414 }, "add@denoiser#1": { "x": 24, "y": 1564 }, "python@denoiser#1": { "x": 24, "y": 1714 }, "denoiser.up": { "x": 24, "y": 1864 }, "cat@denoiser#0": { "x": 24, "y": 2014 }, "denoiser.merge": { "x": 24, "y": 2164 }, "silu@denoiser#2": { "x": 24, "y": 2314 }, "denoiser.noise": { "x": 24, "y": 2464 }, "denoiser.cross_attention.query": { "x": 24, "y": 214 }, "denoiser.cross_attention.key": { "x": 24, "y": 364 }, "div@denoiser.cross_attention#0": { "x": 24, "y": 514 }, "denoiser.cross_attention.value": { "x": 24, "y": 664 }, "denoiser.cross_attention.softmax": { "x": 24, "y": 814 }, "denoiser.cross_attention.project": { "x": 24, "y": 964 }, "decoder.up": { "x": 24, "y": 214 }, "silu@decoder#0": { "x": 24, "y": 364 }, "decoder.rgb": { "x": 24, "y": 514 } }, "expandedModuleIds": [ "encoder" ], "selectedNodeIds": [ "encoder.latent" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "encoder.latent" }, "snapshotId": "snapshot-48f0d017e40d4345", "exportedAt": "2026-09-17T06:51:48.425092+00:00", "modelName": "TinyLatentDiffusion", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"encoder.latent\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"encoder.latent\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#1\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"encoder.latent\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"silu@encoder#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"encoder.latent\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"encoder.latent\",\"kind\":\"layer\",\"op\":\"Conv2d\",\"params\":{\"in_channels\":8,\"kernel_size\":3,\"out_channels\":2,\"padding\":1},\"parent\":\"encoder\"}}" } ] }, { "id": "noise", "title": "Add noise in latent coordinates", "note": "During diffusion training the image encoder supplies the clean latent z₀. The fixed schedule samples zₜ directly at a chosen t. The denoiser predicts the added latent noise; the frozen encoder and decoder are not updated by this loss.", "view": { "camera": { "x": -76.9230769230769, "y": -70.04195804195803, "z": 0.7114427860696517 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "encoder": { "x": 340, "y": 160 }, "corrupt": { "x": 680, "y": 160 }, "python@TinyLatentDiffusion#0": { "x": 1020, "y": 160 }, "conditioner": { "x": 1360, "y": 160 }, "denoiser": { "x": 1700, "y": 160 }, "estimate": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "encoder.down": { "x": 24, "y": 214 }, "silu@encoder#0": { "x": 24, "y": 364 }, "encoder.latent": { "x": 24, "y": 514 }, "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 }, "conditioner.tokens": { "x": 24, "y": 214 }, "conditioner.positions": { "x": 24, "y": 364 }, "add@conditioner#0": { "x": 24, "y": 514 }, "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 }, "flatten@denoiser#0": { "x": 24, "y": 1114 }, "python@denoiser#0": { "x": 24, "y": 1264 }, "denoiser.cross_attention": { "x": 24, "y": 1414 }, "add@denoiser#1": { "x": 24, "y": 1564 }, "python@denoiser#1": { "x": 24, "y": 1714 }, "denoiser.up": { "x": 24, "y": 1864 }, "cat@denoiser#0": { "x": 24, "y": 2014 }, "denoiser.merge": { "x": 24, "y": 2164 }, "silu@denoiser#2": { "x": 24, "y": 2314 }, "denoiser.noise": { "x": 24, "y": 2464 }, "denoiser.cross_attention.query": { "x": 24, "y": 214 }, "denoiser.cross_attention.key": { "x": 24, "y": 364 }, "div@denoiser.cross_attention#0": { "x": 24, "y": 514 }, "denoiser.cross_attention.value": { "x": 24, "y": 664 }, "denoiser.cross_attention.softmax": { "x": 24, "y": 814 }, "denoiser.cross_attention.project": { "x": 24, "y": 964 }, "decoder.up": { "x": 24, "y": 214 }, "silu@decoder#0": { "x": 24, "y": 364 }, "decoder.rgb": { "x": 24, "y": 514 } }, "expandedModuleIds": [ "corrupt" ], "selectedNodeIds": [ "corrupt" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "corrupt" }, "snapshotId": "snapshot-48f0d017e40d4345", "exportedAt": "2026-09-17T06:51:48.425092+00:00", "modelName": "TinyLatentDiffusion", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.cross_attention.key\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.cross_attention.query\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.cross_attention.value\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"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\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"estimate\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@corrupt#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@TinyLatentDiffusion#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"corrupt\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"corrupt\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"estimate\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"corrupt\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@TinyLatentDiffusion#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"encoder\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"corrupt\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"encoder.latent\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"encoder.latent\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@corrupt#1\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"encoder.latent\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@corrupt#0\\\",\\\"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": "condition", "title": "Let spatial queries consult text keys and values", "note": "Queries come from the U-Net’s 2 × 2 bottleneck. Keys and values come from the separate text conditioner. Softmax normalizes over text tokens for every spatial query. Compare two texts with the same image, noise and timestep.", "view": { "camera": { "x": 308.5, "y": -594, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "encoder": { "x": 340, "y": 160 }, "corrupt": { "x": 680, "y": 160 }, "python@TinyLatentDiffusion#0": { "x": 1020, "y": 160 }, "conditioner": { "x": 1360, "y": 160 }, "denoiser": { "x": 1700, "y": 160 }, "estimate": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "encoder.down": { "x": 24, "y": 214 }, "silu@encoder#0": { "x": 24, "y": 364 }, "encoder.latent": { "x": 24, "y": 514 }, "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 }, "conditioner.tokens": { "x": 24, "y": 214 }, "conditioner.positions": { "x": 24, "y": 364 }, "add@conditioner#0": { "x": 24, "y": 514 }, "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 }, "flatten@denoiser#0": { "x": 24, "y": 1114 }, "python@denoiser#0": { "x": 24, "y": 1264 }, "denoiser.cross_attention": { "x": 24, "y": 1414 }, "add@denoiser#1": { "x": 24, "y": 1564 }, "python@denoiser#1": { "x": 24, "y": 1714 }, "denoiser.up": { "x": 24, "y": 1864 }, "cat@denoiser#0": { "x": 24, "y": 2014 }, "denoiser.merge": { "x": 24, "y": 2164 }, "silu@denoiser#2": { "x": 24, "y": 2314 }, "denoiser.noise": { "x": 24, "y": 2464 }, "denoiser.cross_attention.query": { "x": 24, "y": 214 }, "denoiser.cross_attention.key": { "x": 24, "y": 364 }, "div@denoiser.cross_attention#0": { "x": 24, "y": 514 }, "denoiser.cross_attention.value": { "x": 24, "y": 664 }, "denoiser.cross_attention.softmax": { "x": 24, "y": 814 }, "denoiser.cross_attention.project": { "x": 24, "y": 964 }, "decoder.up": { "x": 24, "y": 214 }, "silu@decoder#0": { "x": 24, "y": 364 }, "decoder.rgb": { "x": 24, "y": 514 } }, "expandedModuleIds": [ "denoiser", "denoiser.cross_attention" ], "selectedNodeIds": [ "denoiser.cross_attention.softmax" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "denoiser.cross_attention.softmax" }, "snapshotId": "snapshot-48f0d017e40d4345", "exportedAt": "2026-09-17T06:51:48.425092+00:00", "modelName": "TinyLatentDiffusion", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"denoiser.cross_attention.softmax\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.cross_attention.project\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"div@denoiser.cross_attention#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.cross_attention.softmax\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"denoiser.cross_attention.softmax\",\"kind\":\"layer\",\"op\":\"Softmax\",\"params\":{},\"parent\":\"denoiser.cross_attention\"}}" } ] }, { "id": "decode", "title": "Decode after the latent sampling loop", "note": "Generation begins from Gaussian latent noise, without an image-encoder call. It repeats the denoiser with the same conditioning, then decodes the final latent once. The displayed forward graph instead decodes a single clean estimate so all subsystems remain inspectable.", "view": { "camera": { "x": 308.5, "y": -307.5, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "encoder": { "x": 340, "y": 160 }, "corrupt": { "x": 680, "y": 160 }, "python@TinyLatentDiffusion#0": { "x": 1020, "y": 160 }, "conditioner": { "x": 1360, "y": 160 }, "denoiser": { "x": 1700, "y": 160 }, "estimate": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "encoder.down": { "x": 24, "y": 214 }, "silu@encoder#0": { "x": 24, "y": 364 }, "encoder.latent": { "x": 24, "y": 514 }, "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 }, "conditioner.tokens": { "x": 24, "y": 214 }, "conditioner.positions": { "x": 24, "y": 364 }, "add@conditioner#0": { "x": 24, "y": 514 }, "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 }, "flatten@denoiser#0": { "x": 24, "y": 1114 }, "python@denoiser#0": { "x": 24, "y": 1264 }, "denoiser.cross_attention": { "x": 24, "y": 1414 }, "add@denoiser#1": { "x": 24, "y": 1564 }, "python@denoiser#1": { "x": 24, "y": 1714 }, "denoiser.up": { "x": 24, "y": 1864 }, "cat@denoiser#0": { "x": 24, "y": 2014 }, "denoiser.merge": { "x": 24, "y": 2164 }, "silu@denoiser#2": { "x": 24, "y": 2314 }, "denoiser.noise": { "x": 24, "y": 2464 }, "denoiser.cross_attention.query": { "x": 24, "y": 214 }, "denoiser.cross_attention.key": { "x": 24, "y": 364 }, "div@denoiser.cross_attention#0": { "x": 24, "y": 514 }, "denoiser.cross_attention.value": { "x": 24, "y": 664 }, "denoiser.cross_attention.softmax": { "x": 24, "y": 814 }, "denoiser.cross_attention.project": { "x": 24, "y": 964 }, "decoder.up": { "x": 24, "y": 214 }, "silu@decoder#0": { "x": 24, "y": 364 }, "decoder.rgb": { "x": 24, "y": 514 } }, "expandedModuleIds": [ "decoder" ], "selectedNodeIds": [ "decoder.rgb" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "decoder.rgb" }, "snapshotId": "snapshot-48f0d017e40d4345", "exportedAt": "2026-09-17T06:51:48.425092+00:00", "modelName": "TinyLatentDiffusion", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"decoder.rgb\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"output\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"silu@decoder#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"decoder.rgb\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"decoder.rgb\",\"kind\":\"layer\",\"op\":\"Conv2d\",\"params\":{\"in_channels\":8,\"kernel_size\":3,\"out_channels\":3,\"padding\":1},\"parent\":\"decoder\"}}" } ] } ] }, "threads": [], "savedViews": [ { "id": "condition", "title": "Let spatial queries consult text keys and values", "note": "Queries come from the U-Net’s 2 × 2 bottleneck. Keys and values come from the separate text conditioner. Softmax normalizes over text tokens for every spatial query. Compare two texts with the same image, noise and timestep.", "view": { "camera": { "x": 308.5, "y": -594, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "encoder": { "x": 340, "y": 160 }, "corrupt": { "x": 680, "y": 160 }, "python@TinyLatentDiffusion#0": { "x": 1020, "y": 160 }, "conditioner": { "x": 1360, "y": 160 }, "denoiser": { "x": 1700, "y": 160 }, "estimate": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "encoder.down": { "x": 24, "y": 214 }, "silu@encoder#0": { "x": 24, "y": 364 }, "encoder.latent": { "x": 24, "y": 514 }, "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 }, "conditioner.tokens": { "x": 24, "y": 214 }, "conditioner.positions": { "x": 24, "y": 364 }, "add@conditioner#0": { "x": 24, "y": 514 }, "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 }, "flatten@denoiser#0": { "x": 24, "y": 1114 }, "python@denoiser#0": { "x": 24, "y": 1264 }, "denoiser.cross_attention": { "x": 24, "y": 1414 }, "add@denoiser#1": { "x": 24, "y": 1564 }, "python@denoiser#1": { "x": 24, "y": 1714 }, "denoiser.up": { "x": 24, "y": 1864 }, "cat@denoiser#0": { "x": 24, "y": 2014 }, "denoiser.merge": { "x": 24, "y": 2164 }, "silu@denoiser#2": { "x": 24, "y": 2314 }, "denoiser.noise": { "x": 24, "y": 2464 }, "denoiser.cross_attention.query": { "x": 24, "y": 214 }, "denoiser.cross_attention.key": { "x": 24, "y": 364 }, "div@denoiser.cross_attention#0": { "x": 24, "y": 514 }, "denoiser.cross_attention.value": { "x": 24, "y": 664 }, "denoiser.cross_attention.softmax": { "x": 24, "y": 814 }, "denoiser.cross_attention.project": { "x": 24, "y": 964 }, "decoder.up": { "x": 24, "y": 214 }, "silu@decoder#0": { "x": 24, "y": 364 }, "decoder.rgb": { "x": 24, "y": 514 } }, "expandedModuleIds": [ "denoiser", "denoiser.cross_attention" ], "selectedNodeIds": [ "denoiser.cross_attention.softmax" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "denoiser.cross_attention.softmax" }, "snapshotId": "snapshot-48f0d017e40d4345", "exportedAt": "2026-09-17T06:51:48.425092+00:00", "modelName": "TinyLatentDiffusion", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"denoiser.cross_attention.softmax\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.cross_attention.project\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"div@denoiser.cross_attention#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"denoiser.cross_attention.softmax\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"denoiser.cross_attention.softmax\",\"kind\":\"layer\",\"op\":\"Softmax\",\"params\":{},\"parent\":\"denoiser.cross_attention\"}}" } ] } ], "openingViewId": "condition" } }