Canvas Basics
Overview of the CodefyUI sidebar, canvas, type-safe connections, notes, node cards, toolbar, configuration panel, results panel, and notifications.
Your First Graph
Build and run a minimal pipeline, and learn why every graph needs a Start node to drive execution.
Running Graphs
How execution works — WebSocket streaming, the results panel, live loss charts, and partial re-execution.
Run Queue
Queue graphs per device and walk away — lanes, FIFO order, concurrency limits, cancelling a waiting run, and the cdui run CLI.
Studies
Follow durable studies in the Studies panel, read rankings with ties, and set the server-enforced budgets that halt a study before it runs too long.
Multi-GPU Execution
Use a machine with several GPUs from CodefyUI — when to queue a run instead of clicking Run, starting a worker agent on the same machine, choosing a device or a device set, training one model on several cards, slots and their limits, the Workers panel, what the canvas does, peak memory, and the current limits.
Remote Workers
Run graphs on another machine — enroll a GPU server as a worker agent from the Workers panel, reach the control plane directly, through a TLS proxy or over SSH, share or transfer files, keep versions and nodes in sync, trust, and a systemd unit for the agent.
Saving and Loading Models
What ModelSaver writes and what ModelLoader will read back — state_dict versus full_model, and why loading a model file is a trust decision.
Data & Augmentation
Build a preprocessing pipeline out of nodes, augment your training data, load your own images, and send the metrics to TensorBoard.
Reproducing Baselines
A research-grade walkthrough — reproduce the standard ResNet-18 / CIFAR-10 result end to end from GUI nodes, with a fixed seed, on the run queue.
Build Your Own Eval Pipeline
Evaluate a Hugging Face model on your own data with the eval pack — an eval set with ids and hashes, a pinned model, a prediction table, metrics and an item-aligned A/B comparison.
Teaching Inspector
Record per-node outputs, inspect input→output tensor diffs, compare a subgraph segment, capture gradients, and view step traces.
Tabs & Persistence
Multi-tab workspaces, automatic in-browser saving, importing/exporting graphs as JSON, and moving every open tab to another browser.
Key Bindings
Keyboard and mouse shortcuts for the CodefyUI editor.
CLI Graph Runner
Execute a saved graph.json directly from the command line with run_graph.py — no server required.
Graph as a Function
Call any saved graph headlessly over HTTP as a named function — declared inputs in, declared outputs out.
Publish (Graphs as Apps)
Freeze a saved graph as a versioned app behind a stable, API-key-protected invoke endpoint, with every run recorded to SQLite.
Deployment Behind a Reverse Proxy
Give CodefyUI its own hostname, bind it to loopback, and put nginx in front for TLS and single sign-on -- with a tested systemd unit and nginx config.
Shared Instances
What one CodefyUI instance shares between everyone who can reach it -- ambient credentials, who gets billed, and what is stored per graph.
Version Control Your Graphs
Keep your graph JSON in a git service repo, validate every graph in CI with run_graph.py, and publish from a versioned source.
Source Control
Commit, branch, stash, push and review a CodefyUI project from the editor's Source Control tab, with a diff view that says what changed in the graph.
Project Directories
Turn a service into a self-contained git repo (cdui project) with a logic/layout split, per-project assets and secrets, CI validation, and publish provenance.
Example Projects
Official ready-to-clone example services -- each one a standalone CodefyUI project repository you can run, publish, and fork.
Node Reference
Every built-in node — 158 nodes across 16 categories, from CNN and Transformer layers to RL, LLM, Diffusion, and classical ML.
Optional Packs
Install the optional packages and model files required by selected LLM nodes and GPU backends.
Plugin Center
Install teaching node packs and GitHub plugins from inside the editor, and read what each one asks for before it goes in.
Examples Gallery
Pre-built example workflows — model architectures, end-to-end training, and LLM demos you can load and run.