Skip to main content

Node Reference

CodefyUI ships 94 built-in nodes across 15 categories. Installed plugin packs and your own custom nodes add more.

tip

This list is the source of truth at the time of writing, but the backend is authoritative: the live palette and GET /api/nodes always reflect exactly what your install has. Use the in-app search (double-click the canvas) to find a node fast.

CategoryNodesCount
CNNConv2d, Conv1d, ConvTranspose2d, MaxPool2d, AvgPool2d, AdaptiveAvgPool2d, BatchNorm2d, Dropout, Activation9
RNNLSTM, GRU, RNNCell3
TransformerMultiHeadAttention, TransformerEncoder, TransformerDecoder, MoELayer4
RLDQN, PPO, EnvWrapper, RewardModel, KLDivergence5
DataDataset, DataLoader, Transform, HuggingFaceDataset, KaggleDataset, TensorInput, TextInput, CSVReader, ColumnSelector, Normalize, SyntheticDataset, TrainTestSplit12
Data FlowMap, Reduce, Switch3
TrainingOptimizer, Loss, TrainingLoop, LRScheduler, SequentialModel, BackwardOnce6
IOImageReader, ImageWriter, ImageBatchReader, FileReader, CheckpointSaver, CheckpointLoader, ModelLoader, ModelSaver, Inference9
ControlStart1
UtilityPrint, Reshape, Concat, Flatten, Linear, Visualize, Embedding7
NormalizationBatchNorm1d, LayerNorm, GroupNorm, InstanceNorm2d4
Tensor OperationsAdd, MatMul, Mean, Multiply, Permute, Softmax, Split, Squeeze, Stack, TensorCreate, Unsqueeze11
LLMTokenizer, WordVector, EmbeddingScatter, CosineSimilarity, AttentionMask, AttentionHeatmap, PositionalEncoding7
ClassicalKNN, LinearRegression, LogisticRegression, DecisionTreeClassifier, SVMClassifier, MLPClassifier, Accuracy7
DiffusionUpsample, TimestepEmbedding, Lerp, GaussianNoise, DDPMSampler, DiffusionUNet6

Notable nodes

  • Start (Control) — the execution entry point. Every runnable graph needs one; see Your First Graph.
  • TensorInput (Data) — an inline grid editor to hand-feed explicit tensors into a pipeline; the backbone of Teaching Inspector demos.
  • TrainingLoop (Training) — drives training and emits the live loss chart in the Results panel.
  • EmbeddingScatter (LLM) — projects embeddings to 2D (PCA / t-SNE) for a zoomable scatter plot.
  • AttentionHeatmap (LLM) — renders attention matrices as images.
  • Switch (Data Flow) — conditional routing so only one branch executes.

Port data types

Edges are typed. The built-in data types include: Tensor, Model, Dataset, DataLoader, Optimizer, Loss, Scalar, String, Image, List, Any, Trigger. The Trigger type is what Start nodes emit to drive execution order.