# ComfyUI — full corpus # LLM Wiki An open-source template for building LLM-powered knowledge bases, following [Andrej Karpathy's "LLM Wiki" pattern](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f). You provide raw sources. The LLM reads them, writes structured wiki pages, cross-links everything, and maintains it over time. You never edit the wiki directly — you curate sources and ask questions. ## How It Works The system has three layers: ``` raw/ Sources you collect (articles, transcripts, notes, PDFs) wiki/ LLM-written & maintained pages (summaries, concepts, entities, syntheses) CLAUDE.md Schema that tells the LLM how to structure everything ``` Three operations drive the workflow: | Operation | Trigger | What happens | |-----------|---------|--------------| | **Ingest** | "ingest raw/my-source.txt" | LLM reads the source, creates a summary page, creates/updates concept and entity pages, adds cross-links, updates the index and log | | **Query** | Ask any question | LLM searches the wiki, synthesizes an answer with citations, optionally creates a synthesis page for novel insights | | **Lint** | "lint" or "health check" | LLM audits all pages for orphans, contradictions, missing links, incomplete sections, and low-confidence claims — fixes what it can, reports the rest | ## Quick Start 1. **Clone this repo** ```bash git clone https://github.com/YOUR_USERNAME/llm-wiki.git my-knowledge-base cd my-knowledge-base ``` 2. **Customize CLAUDE.md** for your domain - Update the Purpose section with your topic - Replace the placeholder tagging taxonomy with your own categories - Adjust confidence level descriptions if needed - Everything else (workflows, page formats, linking rules) works as-is 3. **Drop sources into `raw/`** - Text files, transcripts, articles, notes — any plain text - These are immutable once added; the LLM never modifies them 4. **Tell the LLM to ingest** ``` ingest raw/my-first-source.txt ``` The LLM will create summary pages, concept pages, entity pages, cross-links, and update the index. 5. **Ask questions** ``` What are the key differences between X and Y? ``` The LLM answers from the wiki, citing specific pages. 6. **Run health checks** ``` lint ``` The LLM audits the wiki and fixes issues. ## Directory Structure ``` . ├── CLAUDE.md # Schema — the LLM's instructions ├── raw/ # Your source documents (immutable) └── wiki/ ├── index.md # Master catalog of all pages ├── log.md # Append-only activity log ├── dashboard.md # Dataview dashboard (Obsidian) ├── analytics.md # Charts View analytics (Obsidian) ├── flashcards.md # Spaced repetition cards ├── summaries/ # One page per source document ├── concepts/ # Concept and framework pages ├── entities/ # People, tools, organizations, etc. ├── syntheses/ # Cross-cutting analyses and comparisons ├── journal/ # Research/session journal entries │ └── template.md # Journal entry template └── presentations/ # Marp slide decks ``` ## Enhancements This template includes several extras beyond the core wiki pattern: ### Dataview Dashboard (`wiki/dashboard.md`) Live queries that surface low-confidence pages, recent updates, concepts by tag, and pages with the most sources. Requires the [Dataview](https://github.com/blacksmithgu/obsidian-dataview) Obsidian plugin. ### Charts View Analytics (`wiki/analytics.md`) Visual analytics with pie charts, bar charts, and word clouds. Requires the [Charts View](https://github.com/caronchen/obsidian-chartsview-plugin) Obsidian plugin. ### Mermaid Diagrams Use Mermaid code blocks in any wiki page to create flowcharts, sequence diagrams, or concept maps. Native support in Obsidian and GitHub. ### Marp Slides (`wiki/presentations/`) Create slide decks from markdown using [Marp](https://marp.app/). Drop presentation files in this directory. ### Research Journal (`wiki/journal/`) Track your research sessions, experiments, or applied work with the included template. The LLM can reference journal entries when answering queries. ### Spaced Repetition (`wiki/flashcards.md`) Flashcards in the format used by the [Spaced Repetition](https://github.com/st3v3nmw/obsidian-spaced-repetition) Obsidian plugin. Ask the LLM to generate flashcards from any wiki page. ### MCP Server This repo works with Claude Code's MCP server capabilities. Point an MCP-compatible client at this repo and the LLM can read/write the wiki programmatically. ## Customizing for Your Domain The schema in `CLAUDE.md` is domain-agnostic. To adapt it: 1. **Purpose** — Describe your knowledge domain in one paragraph 2. **Tagging taxonomy** — Replace placeholder categories with your own (e.g., for a cooking KB: `cuisine`, `technique`, `ingredient`, `equipment`) 3. **Confidence levels** — Adjust the descriptions to match your domain's evidence standards 4. **Entity types** — Update the entity page description to match what entities mean in your domain (people, tools, companies, etc.) 5. **Journal template** — Customize `wiki/journal/template.md` for your workflow Everything else — page format, linking conventions, workflows, rules — is universal and works across domains. ## Example Domains This template works for any knowledge-intensive topic: - **Research notes** — papers, experiments, methodologies - **Book analysis** — themes, characters, author techniques - **Competitive analysis** — companies, products, market trends - **Course notes** — lectures, readings, key concepts - **Personal development** — frameworks, habits, book summaries - **Technical documentation** — APIs, architectures, design patterns - **Hobby deep-dives** — any subject you want to master ## License MIT --- title: "ComfyUI KB — Master Index" type: index updated: 2026-06-09 comfyui_version: "v0.24.0" --- # ComfyUI KB — Master Index **Domain:** ComfyUI — node-based engine for generative media workflows (images, video, audio) running locally or on Comfy Cloud. Repo: Comfy-Org/ComfyUI; docs: docs.comfy.org. **Corpus:** 177 provenance-stamped sources in `raw/` (121 official guide pages, 45 solved tracker issues, 10 release notes, README). **Pages:** 26 — core guides plus the operator/developer ring. ## Deep-dive layer - [[concepts/core-concepts]] — nodes/models/dependency mental model - [[concepts/node-dev-backend-deep-dive]] — lifecycle, datatypes, hidden inputs, tensors, lazy eval - [[concepts/node-dev-js-extensions]] — LiteGraph, hooks, all extension API surfaces - [[concepts/interface-settings-reference]] — settings categories + subgraphs, partial execution, node docs - [[concepts/desktop-usage-guide]] — instances, snapshots, desktop settings, migration - [[concepts/manager-deep-dive]] — new/legacy UI, config paths (V3.38 change), git-path fix, bisection method - [[concepts/server-internals-and-headless]] — startup flags, comms routes/messages, execution-model inversion (PR #2666) - [[concepts/cloud-workflows]] — sharing links, Civitai/HF imports - [[concepts/cloud-account-and-billing]] — subscription/Stripe, data retention, API-key partner nodes (PR #8041) - [[summaries/release-detail]] — patch-train detail + recurring themes - [[syntheses/update-breakage-casebook]] — diagnostic questions + the broke-my-node signature ## Concepts - [[concepts/install-windows-portable]] — portable download per GPU (CUDA 13.0 / cu126), requirements (Python 3.13, Chrome 143+), update scripts + the dangerous one - [[concepts/install-desktop]] — Comfy Desktop multi-instance manager (Win/macOS; Linux unsupported), Manager pre-enabled - [[concepts/install-manual-and-cli]] — venv+git+pip path; comfy-cli (launch, snapshot, bisect, `comfy generate` partner nodes Beta) - [[concepts/workflows-and-first-generation]] — node-graph model, templates with auto model download, workflows embedded in PNGs - [[concepts/models-and-files]] — `models/checkpoints/` + refresh, external model dirs, Cloud imports - [[concepts/interface-and-settings]] — keybindings, Desktop-only Server Config, Mask Editor - [[concepts/custom-nodes-and-manager]] — security warning, 3 install methods, Manager enablement per install type - [[concepts/node-development]] — official walkthrough, backend (Python), JS extensions (WEB_DIRECTORY + registerExtension) - [[concepts/server-and-api]] — 127.0.0.1:8188, API-format-vs-save-format gotcha, /prompt route, 3 integration patterns ## Entities - [[entities/comfy-cloud]] — official hosted ComfyUI (RTX 6000 Pro, 400 free credits/mo, Civitai/HF imports on Creator tier) ## Summaries - [[summaries/release-digest]] — v0.19.0 → v0.24.0 (10 releases in <2 months), highlights per release ## Syntheses - [[syntheses/install-methods-compared]] — portable vs desktop vs manual vs CLI vs cloud, decision rules - [[syntheses/model-hardware-compat-notes]] — Mac/MPS patterns, new-model support cycle, GPU edge cases (reported, low-confidence by design) - [[syntheses/troubleshooting-playbook]] — CUDA-not-compiled, black images, VAE OOM, IMPORT FAILED, model-not-found… symptom → cause → fix ## Statistics - **Total pages**: 14 · **Sources ingested**: 177 (immutable raw/) - **Confidence**: 9 high · 4 medium · 1 low ## Coverage notes Strong: install (all methods), workflows basics, custom nodes + Manager, server/API, troubleshooting, releases. Thinner (raw/ has sources; pages pending): built-in node reference (714 pages deliberately excluded from gather), interface deep-dive (settings subpages), node-dev backend details (13 docs), JS hook reference (15 docs), Comfy Cloud workflows. Recency: sources fetched 2026-06-09; ComfyUI ships ~weekly — expect drift. --- title: "Cloud Account, Billing & API-Key Partner Nodes" type: concept tags: [cloud, billing, subscription, api-key] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-support-subscription-subscribing-mdx.md, raw/github_doc-support-subscription-canceling-mdx.md, raw/github_doc-support-payment-accepted-payment-methods-mdx.md, raw/github_doc-support-data-retention-mdx.md, raw/github_doc-support-contact-support-mdx.md, raw/github_doc-development-comfyui-server-api-key-integration-mdx.md] --- # Cloud Account, Billing & API-Key Partner Nodes ## Subscription model You can **browse Comfy Cloud without a subscription, but running workflows requires one**. Subscribe/cancel at `cloud.comfy.org` (model imports need Creator tier — [[concepts/cloud-workflows]]). ## Payments Processor is **Stripe**: major cards (Visa/Mastercard/Amex) plus regional wallets (Google Pay, Link). What you see at checkout is what's available in your region. Further docs cover payment history, invoices, currency, and failed payments (`raw/github_doc-support-payment-*`). ## Data retention & support A retention table per data type is published (with ToS/Privacy links); official help = `support.comfy.org` + the per-repo issue trackers. ## API-key partner nodes (the developer-relevant bit) Since **PR #8041**, **built-in paid Partner Nodes work via your ComfyUI Account API key directly** — no frontend required at all. That enables hybrid workflows mixing **local OSS models + hosted partner models** in one graph, even on headless servers ([[concepts/server-internals-and-headless]]). ## Related [[entities/comfy-cloud]] · [[summaries/release-detail]] (partner-node release stream) --- title: "Comfy Cloud — Sharing & Model Imports" type: concept tags: [cloud, sharing, models] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-cloud-share-workflow-mdx.md, raw/github_doc-cloud-import-models-mdx.md, raw/github_doc-get-started-cloud-mdx.md] --- # Comfy Cloud — Sharing & Model Imports Operational detail behind [[entities/comfy-cloud]]. ## Workflow sharing Generate a **shareable link**; recipients see the workflow **with its inputs, outputs, and any referenced assets** — distribution without JSON-file passing (the local equivalent remains workflow-JSON / PNG-embedded workflows, [[concepts/workflows-and-first-generation]]). ## Model imports Import models **directly from Civitai and Hugging Face** into your cloud workspace. Requires **Creator tier or above**. (Local equivalent: the models directory rules in [[concepts/models-and-files]].) ## Positioning Same ComfyUI, pre-installed, on RTX 6000 Pro GPUs; 400 free credits/month with Google sign-in — the no-hardware on-ramp in [[syntheses/install-methods-compared]]. --- title: "Core Concepts — Nodes, Models, Workflow Dependencies" type: concept tags: [concepts, nodes, models, dependencies] updated: 2026-06-09 confidence: high sources: [raw/github_doc-development-core-concepts-nodes-mdx.md, raw/github_doc-development-core-concepts-models-mdx.md, raw/github_doc-development-core-concepts-dependencies-mdx.md, raw/github_doc-development-core-concepts-workflow-mdx.md] --- # Core Concepts — Nodes, Models, Workflow Dependencies The mental model behind everything else. ## Nodes Nodes are ComfyUI's building blocks — each an independent module (**Comfy Core** node or **Custom Node**) with inputs, outputs, and parameters, connected by links "like LEGO blocks." Behind each node is Python logic (e.g. the K-Sampler node: multiple inputs/outputs + the parameters that drive sampling) — you compose capability without writing code. ## Models, in plain terms Models = the **weight files** workflows actually run: **checkpoints, VAEs, LoRAs, ControlNets, upscalers**. The app install is small and ships **no models** — you download them, place them under `ComfyUI/models/`, and select them in the matching **Load-** node ([[concepts/models-and-files]]). ## Why a downloaded workflow won't run (dependencies) A workflow file depends on more than itself: **media assets, models, custom nodes, and Python dependencies**. Community workflows fail to run out-of-the-box when any of those are missing — Manager's missing-node detection covers one class ([[concepts/custom-nodes-and-manager]]); models and assets are on you. ## Related [[concepts/workflows-and-first-generation]] · [[concepts/node-development]] --- title: "Custom Nodes & ComfyUI-Manager" type: concept tags: [custom-nodes, manager, extensions] updated: 2026-06-09 confidence: high sources: [raw/github_doc-installation-install-custom-node-mdx.md, raw/github_doc-manager-overview-mdx.md, raw/github_doc-manager-install-mdx.md] --- # Custom Nodes & ComfyUI-Manager **Custom nodes** are community extensions that add functionality. Every install boils down to: (1) node code into `ComfyUI/custom_nodes/`, (2) install its Python dependencies. ## Security first (official warning) Malicious plugins can exploit custom nodes: install only from **trusted authors** commonly used by the community; understand functionality before installing; avoid obscure/suspicious plugins — unverified ones can compromise the system. ## Three install methods (official comparison) | Method | Pros | Cons | |---|---|---| | **ComfyUI-Manager** (recommended) | automated install, dependency handling, GUI | can't find nodes not in the registry | | **Git clone** | any repo, any revision | manual deps | | **ZIP** | no git needed | manual deps, no updates | ## ComfyUI-Manager Built into current ComfyUI core. Features: **install/update/remove/disable custom nodes, model downloads, snapshot save/restore, missing-node detection** (auto-detects nodes a loaded workflow needs). **Enabling, per install type:** - **Desktop:** included and enabled by default — nothing to do. - **Portable:** `.\python_embeded\python.exe -m pip install -r ComfyUI\manager_requirements.txt`, then launch with `--enable-manager` (alongside `--windows-standalone-build`). - **Manual:** same idea inside your venv — install `manager_requirements.txt`, launch with `--enable-manager`. ## Related [[concepts/node-development]] (writing your own) · [[syntheses/troubleshooting-playbook]] (IMPORT FAILED, missing-manager cases) --- title: "Comfy Desktop — Instances, Snapshots, Settings, Migration" type: concept tags: [desktop, instances, snapshots, settings] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-installation-desktop-usage-instance-management-mdx.md, raw/github_doc-installation-desktop-usage-snapshots-mdx.md, raw/github_doc-installation-desktop-usage-migrate-mdx.md, raw/github_doc-installation-desktop-usage-settings-mdx.md, raw/github_doc-installation-desktop-faq-mdx.md, raw/github_doc-installation-desktop-usage-overview-mdx.md] --- # Comfy Desktop — Day-to-Day Usage The operational layer of [[concepts/install-desktop]] (Comfy Desktop = multi-instance launcher). ## Instance management Create, edit, rename, and delete **separate ComfyUI instances** from the launcher — each its own environment (the isolation that prevents one setup's custom-node deps breaking another's). ## Snapshots **Back up, restore, and share instance configurations.** This is Desktop's answer to update-breakage: snapshot before risky changes, restore when an update or node install goes wrong (the CLI's equivalent is `comfy-cli` snapshot/bisect — [[concepts/install-manual-and-cli]]). ## Settings (Desktop-global) Global settings cover **theme, proxy/mirror configuration, and data locations** — proxy/mirror matters in restricted networks where model downloads fail ([[syntheses/troubleshooting-playbook]] § Hub errors). ## Migration A dedicated guide covers migrating a **Legacy Desktop** installation to current Comfy Desktop. ## FAQ A desktop-specific FAQ doc handles its common questions/troubleshooting — distinct from the general tracker ([[syntheses/troubleshooting-playbook]]). ## Related [[concepts/install-desktop]] · [[concepts/interface-settings-reference]] --- title: "Install — Comfy Desktop" type: concept tags: [install, desktop, windows, macos] updated: 2026-06-09 confidence: high sources: [raw/github_doc-installation-desktop-windows-mdx.md, raw/github_doc-installation-desktop-macos-mdx.md, raw/github_doc-installation-desktop-linux-mdx.md, raw/github_doc-manager-install-mdx.md] --- # Install — Comfy Desktop **Comfy Desktop** is a multi-installation manager: run multiple ComfyUI instances from a single launcher. Download from `comfy.org/download` (auto-detects platform). ## Requirements | OS | Requirements | |---|---| | **Windows** | Windows 10+, x64 or ARM64; dedicated GPU (NVIDIA/AMD) recommended but not required; ≥4.85 GB disk **per installation** | | **macOS** | macOS 13 (Ventura)+, Apple Silicon (M1+); ≥4.85 GB per installation | | **Linux** | **Not officially supported** — build from source (`Comfy-Org/Comfy-Desktop` repo) or use portable/manual/CLI instead | ## What you get - Multi-instance management (create/manage/migrate instances, snapshots — see the desktop usage docs in `raw/`) - **ComfyUI-Manager included and enabled by default** (portable/manual users must enable it — [[concepts/custom-nodes-and-manager]]) ## Related [[syntheses/install-methods-compared]] · [[concepts/install-windows-portable]] · [[concepts/install-manual-and-cli]] --- title: "Install — Manual & comfy-cli" type: concept tags: [install, manual, comfy-cli, linux, macos] updated: 2026-06-09 confidence: high sources: [raw/github_doc-installation-manual-install-mdx.md, raw/github_doc-comfy-cli-getting-started-mdx.md, raw/github_doc-installation-system-requirements-mdx.md] --- # Install — Manual & comfy-cli ## Manual install (Windows / macOS / Linux) Four steps, per the official guide: 1. **Create a virtual environment** (conda or venv — required: ComfyUI's deps conflict with system packages) 2. **Clone the repo** (`git` required; on Windows also Microsoft Visual C++ Redistributable) 3. **Install dependencies** (`pip install -r requirements.txt` + the GPU-correct torch — see [[concepts/install-windows-portable]] § Requirements) 4. **Start ComfyUI** (`python main.py`) ## comfy-cli (the command-line manager) `comfy-cli` (github.com/Comfy-Org/comfy-cli) does two jobs: 1. **Manage a local install** — install, launch, update, **snapshot, and bisect** ComfyUI and its custom nodes: ```bash comfy launch ``` 2. **Call hosted partner nodes directly** (`comfy generate`, Beta) — generate images/video from Seedance, Nano Banana (Gemini), Grok, Flux, Ideogram, DALL·E, Recraft, Stability, Kling, Luma, Runway, Pika, Vidu, Hailuo, Moonvalley and more, **with no local ComfyUI or workflow JSON**. Beta caveat (official): flag names, model aliases, and output formats may change. The CLI reference covers `generate`, `nodes`, and `models` command groups. ## Related [[syntheses/install-methods-compared]] · [[concepts/custom-nodes-and-manager]] --- title: "Install — Windows Portable" type: concept tags: [install, windows, portable, update] updated: 2026-06-09 confidence: high sources: [raw/github_doc-installation-comfyui-portable-windows-mdx.md, raw/github_doc-installation-system-requirements-mdx.md, raw/github_doc-installation-update-comfyui-mdx.md] --- # Install — Windows Portable **ComfyUI Portable** is a self-contained Windows build with its own embedded Python (`python_embeded`) — extract with 7-Zip and run; no environment setup. ## Download — match your GPU | Package | For | |---|---| | `ComfyUI_windows_portable_nvidia.7z` (CUDA 13.0, Python 3.13) | modern NVIDIA (RTX series) | | `ComfyUI_windows_portable_nvidia_cu126.7z` (CUDA 12.6, Python 3.12) | NVIDIA 10-series and older | From `github.com/Comfy-Org/ComfyUI/releases/latest`. Extracted layout: `ComfyUI_windows_portable/` containing the `ComfyUI` main program (+ `python_embeded`, `update/`). ## Requirements (all install methods) - Windows / Linux / macOS (Apple Silicon M-series) - **Python 3.13 recommended** (3.14 works, some custom nodes break; 3.12 is the fallback when custom-node deps fail on 3.13) - **Chrome 143+** for the UI (≤142 has known visual/performance issues) - NVIDIA: stable PyTorch with CUDA 13.0 — `pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130`; AMD on Linux: ROCm 7.2 ## Updating (portable) Batch scripts live in the `update/` folder: ``` update_comfyui.bat → latest development version update_comfyui_stable.bat → latest stable version update_comfyui_and_python_dependencies.bat → ⚠️ DANGER: reinstalls deps — may cause conflicts ``` Use the dependencies script only as a last resort — it's the documented cause of dependency breakage (see [[syntheses/troubleshooting-playbook]] for the "torch not compiled with CUDA" failure it can trigger). ## Related [[syntheses/install-methods-compared]] · [[concepts/install-desktop]] · [[concepts/install-manual-and-cli]] --- title: "Interface, Settings & Mask Editor" type: concept tags: [interface, settings, shortcuts, maskeditor] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-interface-shortcuts-mdx.md, raw/github_doc-interface-maskeditor-mdx.md, raw/github_doc-interface-settings-server-config-mdx.md, raw/github_doc-interface-overview-mdx.md] --- # Interface, Settings & Mask Editor ## Frontend & localization The ComfyUI frontend is a **separate project** (`Comfy-Org/ComfyUI_frontend`), released as an independent pip package. The UI supports English, Chinese, Russian, French, Japanese, Korean — switch via `Settings → Comfy → Locale`. ## Settings & keyboard shortcuts ComfyUI supports **custom keybindings**: `Settings (gear icon) → Keybinding` lists every command's current shortcut; click the edit icon on a command to rebind it. ## Server Config settings (Desktop only) The `Server Config` settings menu **exists only in Comfy Desktop** (not portable/manual). It exposes server options like the **Host** (IP to listen on) and related network settings — on other installs these are launch flags instead ([[concepts/server-and-api]]). ## Mask Editor Built-in mask creation/editing on your images — no external editor. Drawing tools, **layer management**, undo/redo, canvas transforms, GPU-accelerated brush rendering. Open it via (1) the **Selection Toolbox** mask button on a selected Load Image node, or the other documented entry points (right-click menu). ## Related [[concepts/workflows-and-first-generation]] · [[concepts/custom-nodes-and-manager]] --- title: "Interface Settings Reference & Power Features" type: concept tags: [interface, settings, subgraph, templates] updated: 2026-06-09 confidence: high sources: [raw/github_doc-interface-settings-overview-mdx.md, raw/github_doc-interface-settings-comfy-mdx.md, raw/github_doc-interface-settings-lite-graph-mdx.md, raw/github_doc-interface-settings-extension-mdx.md, raw/github_doc-interface-settings-3d-mdx.md, raw/github_doc-interface-settings-mask-editor-mdx.md, raw/github_doc-interface-features-template-mdx.md, raw/github_doc-interface-features-subgraph-mdx.md, raw/github_doc-interface-features-partial-execution-mdx.md, raw/github_doc-interface-features-node-docs-mdx.md] --- # Interface Settings Reference & Power Features ## Settings system - Open with **`Ctrl + ,`**. All user settings persist to **`ComfyUI/user/default/comfy.settings.json`**. - Custom nodes can register their own settings categories — official docs cover only native ones; some options are **Desktop-only** (marked per page). | Settings category | Controls | |---|---| | Comfy / Comfy Desktop | general app + desktop-specific options | | **LiteGraph** | the canvas engine: canvas, node, and link behavior/appearance | | **Extension** | enable/disable **frontend extensions** registered by custom nodes — *not* the custom nodes themselves ([[concepts/custom-nodes-and-manager]] for those) | | 3D | initialization defaults for 3D components (camera, lighting, scene) | | Mask Editor | brush and editor behavior ([[concepts/interface-and-settings]]) | | Keybinding / Appearance / Server Config | shortcuts, themes, Desktop server options | ## Power features - **Workflow Templates** — the browser for natively-supported model workflows *plus example workflows shipped by custom nodes* ([[concepts/workflows-and-first-generation]]). - **Subgraphs** — collapse node groups into reusable units; requires **frontend ≥1.24.3** (if missing, update — converting subgraphs back to nodes is still future work per the doc). - **Partial Execution** — run only part of a workflow: select an **output node** and click the green-triangle icon in the selection toolbox. - **Node Docs** — in-app documentation for any node (functionality, parameters, examples). ## Related [[concepts/interface-and-settings]] · [[concepts/node-dev-js-extensions]] (registering settings/commands) --- title: "ComfyUI-Manager — Pack Management, Configuration, Legacy UI" type: concept tags: [manager, custom-nodes, configuration] updated: 2026-06-09 confidence: high sources: [raw/github_doc-manager-pack-management-mdx.md, raw/github_doc-manager-configuration-mdx.md, raw/github_doc-manager-legacy-ui-mdx.md, raw/github_doc-manager-troubleshooting-mdx.md] --- # ComfyUI-Manager — Deep Dive Operational detail behind [[concepts/custom-nodes-and-manager]]. ## New UI vs legacy UI Manager has two interfaces; switching between them is covered in the install doc. The **new UI** ("pack management") is the current install/update/manage surface; the **legacy UI** flow (Manager button → install nodes) remains documented for older setups. ## Configuration paths (security-relevant change) Since **Manager V3.38**, configuration lives in a **protected system path**: | ComfyUI version | Manager path | |---|---| | v0.3.76+ (System User API) | `/__manager/` | | older | `/default/ComfyUI-Manager/` | ## Manager troubleshooting (from the official doc) - **Custom git path:** if `git.exe` isn't on the default path — run ComfyUI once with Manager installed, then set the path in `/default/ComfyUI-Manager/config.ini`. - Prerequisite for everything git-based (Manager installs, comfy-cli): **git must be installed**. ## The custom-node bisection method (official flowchart) When something breaks: **disable all custom nodes** → issue gone? it's a custom node → re-enable in halves to find the culprit. Issue persists with all disabled? It's core — go to [[syntheses/troubleshooting-playbook]]. ## Related [[concepts/custom-nodes-and-manager]] · [[concepts/install-manual-and-cli]] --- title: "Models — Where Files Go" type: concept tags: [models, checkpoints, files] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-get-started-first-generation-mdx.md, raw/github_issue-issues-compyui-can-t-find-the-ckpt-name.md, raw/github_doc-installation-comfyui-portable-windows-mdx.md, raw/github_doc-cloud-import-models-mdx.md] --- # Models — Where Files Go ## The rule that solves most "model not found" problems Checkpoints go in **`ComfyUI/models/checkpoints/`** — and after adding files, **refresh the browser page** so loader nodes see them. ("Can't find the ckpt_name" in a loader node = the folder is empty or unrefreshed; straight from the maintainers' answer to that exact issue.) Other model types follow the same pattern under `ComfyUI/models//` (loras, vae, etc. — the subfolders ship with the install). ## Getting models - **Workflow templates auto-prompt** to download the models they need ([[concepts/workflows-and-first-generation]]) — the easiest path. - **Manual:** download from model hubs and place per the rule above. The portable build's docs include adding external model directories (`extra_model_paths` configuration — see the portable install doc's AddExternalModels section). - **Comfy Cloud:** import directly from **Civitai and Hugging Face** (requires the Creator subscription tier or higher) — [[entities/comfy-cloud]]. ## Related [[concepts/install-windows-portable]] · [[syntheses/troubleshooting-playbook]] (model-load failures) --- title: "Node Development — Backend Deep Dive" type: concept tags: [development, backend, python, datatypes, tensors] updated: 2026-06-09 confidence: high sources: [raw/github_doc-custom-nodes-backend-datatypes-mdx.md, raw/github_doc-custom-nodes-backend-more-on-inputs-mdx.md, raw/github_doc-custom-nodes-backend-lifecycle-mdx.md, raw/github_doc-custom-nodes-backend-images-and-masks-mdx.md, raw/github_doc-custom-nodes-backend-tensors-mdx.md, raw/github_doc-custom-nodes-backend-lazy-evaluation-mdx.md, raw/github_doc-custom-nodes-backend-server-overview-mdx.md] --- # Node Development — Backend Deep Dive The detail layer behind [[concepts/node-development]] — for actually writing Python nodes. ## Lifecycle — how Comfy loads your node On startup, Comfy scans `custom_nodes/` for **Python modules** (a directory with `__init__.py`; exports controlled by `__all__`). A module that exports **`NODE_CLASS_MAPPINGS`** is treated as a custom node. ## Datatypes Comfy datatypes are client-side "strong typing": the JS client refuses to connect an output to an input of a different datatype (with a few documented exceptions). Built-ins cover the core objects (IMAGE, LATENT, MASK, MODEL, …) and you can define **custom datatypes** for your own node families. ## Inputs — beyond required/optional `INPUT_TYPES` may also return a **`hidden`** dict requesting server-side info per execution: - `PROMPT` — the executing prompt/workflow - `EXTRA_PNGINFO` — metadata destined for the saved PNG - `UNIQUE_ID` — the node's id in the graph ## Images, latents, masks = torch.Tensor All number-crunching is PyTorch. Key working facts: - Images/latents/masks are `torch.Tensor`s with Comfy-specific shape conventions (the tensors doc is the intro). - **Single-tensor output gotcha (official warning):** return `(image,)` — *with the comma* — not `(image)`. - The snippets doc collects copy-paste examples for these patterns. ## Lazy evaluation By default every `required`/`optional` input is evaluated before the node runs. Mark inputs lazy when they may go unused — e.g. `ModelMergeSimple` with ratio `0.0` needn't load the first model at all. ## Server-side pattern A node class declares `INPUT_TYPES` (classmethod) → typed inputs; execution returns the output tuple ([[concepts/node-development]] has the minimal example; the server-overview doc the full property set). ## Related [[concepts/node-dev-js-extensions]] · [[concepts/node-development]] --- title: "Node Development — JS Extension APIs" type: concept tags: [development, javascript, frontend, extensions] updated: 2026-06-09 confidence: high sources: [raw/github_doc-custom-nodes-js-javascript-overview-mdx.md, raw/github_doc-custom-nodes-js-javascript-hooks-mdx.md, raw/github_doc-custom-nodes-js-javascript-objects-and-hijacking-mdx.md, raw/github_doc-custom-nodes-js-javascript-settings-mdx.md, raw/github_doc-custom-nodes-js-javascript-dialog-mdx.md, raw/github_doc-custom-nodes-js-javascript-commands-keybindings-mdx.md, raw/github_doc-custom-nodes-js-javascript-sidebar-tabs-mdx.md, raw/github_doc-custom-nodes-js-javascript-selection-toolbox-mdx.md, raw/github_doc-custom-nodes-js-javascript-examples-mdx.md] --- # Node Development — JS Extension APIs The client-side API surface for extensions (setup: `WEB_DIRECTORY` + `app.registerExtension` — [[concepts/node-development]]). ## The architecture facts that matter first - **The Comfy UI is built on LiteGraph** (`jagenjo/litegraph.js`) — complex client work means learning LiteGraph's objects; much Comfy behavior *is* LiteGraph behavior ("objects and hijacking" doc). - **Hooks:** Comfy calls `#invokeExtensionsAsync` / `#invokeExtensions` with a hook name at defined points — your extension implements those hook names to get called. ## The extension API surfaces (one line each) | API | Lets an extension… | |---|---| | **Settings** | add entries to the ComfyUI settings panel | | **Dialog** | show standardized prompt/confirm dialogs, consistent across desktop+web | | **Commands & Keybindings** | register commands and bind them to keyboard shortcuts | | **Sidebar Tabs** | add persistent custom tabs to the sidebar | | **Selection Toolbox** | add context-sensitive buttons that appear over selected nodes/groups | | **Toast / About badges / Bottom panel / Topbar menu** | notifications, about-page badges, panel tabs, menu items (one doc each in `raw/`) | | **Annotated examples** | copy-paste fragments incl. right-click menu extensions | ## Related [[concepts/node-dev-backend-deep-dive]] · [[concepts/interface-and-settings]] --- title: "Custom Node Development (Backend + JS)" type: concept tags: [development, custom-nodes, python, javascript] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-custom-nodes-walkthrough-mdx.md, raw/github_doc-custom-nodes-js-javascript-overview-mdx.md, raw/github_doc-custom-nodes-overview-mdx.md, raw/github_doc-custom-nodes-tips-mdx.md, raw/github_doc-custom-nodes-backend-server-overview-mdx.md] --- # Custom Node Development ## The walkthrough (start here) The official getting-started walkthrough builds a complete custom node step-by-step — a node that takes a batch of images and returns one (the lightest on average), then extends it with selection criteria and client-side code. It assumes "very little knowledge of Python or Javascript" and hands off to the backend and frontend deep-dives. ## Backend (Python) A node is a Python class registered with ComfyUI's server — the `custom-nodes/backend/` docs (13 pages in `raw/`) cover the server overview, datatypes, lifecycle, and more. After the walkthrough, `backend/server_overview` is the entry point — its core pattern: a node class declares `INPUT_TYPES` as a classmethod returning `required` inputs with types like `("IMAGE", {})`. ## Frontend (JavaScript extensions) Extending the Comfy client takes three steps (official): 1. Export `WEB_DIRECTORY` from your Python module 2. Place `.js` files in that directory 3. Call `app.registerExtension` to register Then use the documented **hooks** (`custom-nodes/js/javascript_hooks`) to get your code called — the `custom-nodes/js/` docs (15 pages in `raw/`) cover objects, settings, dialogs, annotated examples. ## Distribution Publish to the registry (`registry.comfy.org`) so [[concepts/custom-nodes-and-manager]] can find it; workflow templates and help pages can ship with the node (see `custom-nodes/workflow_templates`, `help_page` docs). ## Related [[concepts/server-and-api]] · [[concepts/custom-nodes-and-manager]] --- title: "The Registry — Versioned Custom-Node Publishing" type: concept tags: [registry, publishing, versioning, custom-nodes] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-registry-overview-mdx.md, raw/github_doc-registry-publishing-mdx.md, raw/github_doc-community-contributing-mdx.md] --- # The Registry — Versioned Custom-Node Publishing The **Comfy Registry** (registry.comfy.org) is the public custom-node collection that **powers ComfyUI-Manager**: developers publish/version/deprecate nodes and track metrics; users discover, install, and rate through Manager. ## The headline feature: semantic versioning Registry nodes are **semver'd**, and **workflow JSON stores the node version used** — so workflows reproduce reliably, and users can upgrade/lock/deprecate knowing the impact in advance. This is the structural fix for the "update broke my node" class ([[syntheses/update-breakage-casebook]]). ## Publishing (for node authors) Set up a registry account → create a **Publisher** → publish (video walkthrough + step docs; ~29 more registry docs in `raw/github_doc-registry-*` cover the API surface — tokens, versions, reviews, deprecation). ## Related [[concepts/custom-nodes-and-manager]] · [[concepts/node-development]] --- title: "Server & API (Programmatic ComfyUI)" type: concept tags: [api, server, automation, websocket] updated: 2026-06-09 confidence: high sources: [raw/github_doc-development-comfyui-server-api-examples-mdx.md, raw/github_doc-development-comfyui-server-comms-routes-mdx.md, raw/github_doc-development-api-development-workflow-api-format-mdx.md, raw/github_issue-feature-request-programmatic-use-and-api.md] --- # Server & API (Programmatic ComfyUI) ComfyUI's web client talks to a local server — default `127.0.0.1:8188` — and your code can too. (Programmatic pipelines-as-JSON was one of the community's most-requested features; it's now first-class.) ## Workflow API format — the gotcha that bites first Workflows have **two JSON formats**: the browser **Save format** and the **API format**. Programmatic submission requires the **API format** — export it from the UI before calling the server (see the workflow-api-format doc for the differences). ## Core routes Defined in `server.py` (search `@routes`). Submitting a workflow POSTs to **`/prompt`**, which validates and queues it, returning `prompt_id` + `number` (queue position) — or `error` + `node_errors` on validation failure. The prompt queue and `PromptExecutor` live in `execution.py`. ## Three official integration patterns (with example scripts) 1. **Submit and forget** — HTTP-only POST to `/prompt` (`basic_api_example.py`); check outputs later. 2. **Submit + wait via WebSocket** — track execution progress in real time (`websocket-client` package). 3. **WebSocket with real-time image output** — receive images as they render. Examples live in the repo's `script_examples/` (Python stdlib + `websocket-client`); the protocol is language-agnostic. ## Related [[concepts/workflows-and-first-generation]] · [[concepts/interface-and-settings]] (Desktop server-config UI) · [[entities/comfy-cloud]] (hosted API) --- title: "Server Internals — Headless, Startup Flags, Execution Model" type: concept tags: [server, headless, api, execution] updated: 2026-06-09 confidence: high sources: [raw/github_doc-development-comfyui-server-comms-overview-mdx.md, raw/github_doc-development-comfyui-server-comms-routes-mdx.md, raw/github_doc-development-comfyui-server-startup-flags-mdx.md, raw/github_doc-development-comfyui-server-execution-model-inversion-guide-m.md, raw/github_doc-development-comfyui-server-comms-messages-mdx.md] --- # Server Internals — Headless, Startup Flags, Execution Model Deeper layer under [[concepts/server-and-api]]. ## The server, headless ComfyUI **is** an HTTP server (`http://127.0.0.1:8188` on launch) — submit workflows, upload files, download outputs, and monitor progress programmatically, browser never opened. Startup flags (dedicated doc) control listen address, ports, and behavior for service deployments. ## Comms model - **Routes** (`@routes` in `server.py`): `/prompt` validates + queues (returns `prompt_id` + queue `number`, or `error`/`node_errors`); queue + `PromptExecutor` live in `execution.py` ([[concepts/server-and-api]]). - **Messages**: a WebSocket message stream reports execution progress/events (the comms-messages doc catalogs them) — what the three integration patterns consume. ## The execution-model inversion (matters for node authors) **PR #2666 inverted execution from back-to-front recursion to front-to-back topological sort.** Most nodes "just work," but (per the official guide): - **Monkey-patches of the execution model break** — and the new engine outperforms the popular patches, making many unnecessary. - Optional-input validation semantics changed (the guide details the cases). If an old custom node misbehaves around execution order/caching, this transition is the first suspect ([[syntheses/update-breakage-casebook]]). ## Related [[concepts/server-and-api]] · [[concepts/node-dev-backend-deep-dive]] --- title: "Generation Basics — Text-to-Image, LoRA, ControlNet" type: concept tags: [tutorials, text-to-image, lora, controlnet] updated: 2026-06-09 confidence: high sources: [raw/github_doc-tutorials-basic-text-to-image-mdx.md, raw/github_doc-tutorials-basic-lora-mdx.md, raw/github_doc-tutorials-controlnet-controlnet-mdx.md] --- # Generation Basics — Text-to-Image, LoRA, ControlNet The official tutorial layer for the three foundational techniques. ## Text-to-image The canonical first workflow ([[concepts/workflows-and-first-generation]]): the tutorial walks the complete graph while teaching **diffusion-model principles** and what each node contributes (checkpoint loader → CLIP text encode (positive/negative) → sampler → VAE decode → save). ## LoRA **LoRA (Low-Rank Adaptation)** fine-tunes a large model by training **low-rank matrices over a frozen base** — only a fraction of parameters — so adapters are *small, cheap to train, and stack on a base model* (vs full checkpoints like SD1.5). In workflows: a LoRA-loader node between checkpoint and sampler; the multiple-LoRAs tutorial covers stacking (`raw/github_doc-tutorials-basic-multiple-loras-mdx.md`). ## ControlNet **ControlNet** (Zhang & Agrawala et al., 2023 — "Adding Conditional Control to Text-to-Image Diffusion Models") adds **multimodal conditioning** (pose, depth, edges…) to diffusion generation — the answer to "one prompt can't control composition." Dedicated tutorials cover depth ControlNet, depth T2I-adapter, pose two-pass, and **mixing multiple ControlNets** in one graph. ## Related [[concepts/models-and-files]] (where these model files go) · [[summaries/tutorials-catalog]] --- title: "Workflows & Your First Generation" type: concept tags: [workflows, basics, text-to-image, templates] updated: 2026-06-09 confidence: high sources: [raw/github_doc-get-started-first-generation-mdx.md, raw/github_doc-development-core-concepts-workflow-mdx.md, raw/github_doc-installation-system-requirements-mdx.md] --- # Workflows & Your First Generation ## The core idea A **workflow** is a network ("graph") of connected **nodes** — ComfyUI is a visual programming environment, the same node-graph paradigm as Nuke, Maya, Blender, or Unreal. A workflow can generate any media type: image, video, audio, even AI-model/agent pipelines. ## First text-to-image, step by step (official guide) 1. **Launch ComfyUI** and reach the interface (or skip local entirely with Comfy Cloud — 400 free credits/month with Google sign-in; [[entities/comfy-cloud]]). 2. **Load an example workflow**: - `Workflow → Browse Workflow Templates` — built-in templates use only Core nodes and **auto-prompt you to download required models**, or - load from an **image with embedded workflow metadata** (ComfyUI saves workflows into generated PNGs). 3. **Install models** — automatically via the template prompt, or manually ([[concepts/models-and-files]]). 4. **Generate** — positive/negative prompts in, image out. ## Sharing & the ecosystem Workflows are JSON; generated images embed their workflow, so any ComfyUI image can be dragged back in to reproduce its graph. The community extends the node set via the registry (`registry.comfy.org`) — [[concepts/custom-nodes-and-manager]]. ## Related [[concepts/interface-and-settings]] · [[concepts/server-and-api]] (programmatic execution) --- title: "Comfy Cloud" type: entity tags: [cloud, hosted] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-get-started-cloud-mdx.md, raw/github_doc-cloud-import-models-mdx.md, raw/github_doc-cloud-share-workflow-mdx.md, raw/github_doc-get-started-first-generation-mdx.md] --- # Comfy Cloud The **official hosted ComfyUI** (comfy.org/cloud): same features as local, pre-installed and ready, running on **NVIDIA RTX 6000 Pro GPUs**. Zero local setup — sign in with Google for **400 free credits/month**. ## Notables - **Model import from Civitai and Hugging Face** — requires **Creator subscription tier or higher**. - Workflow sharing (cloud/share-workflow doc). - The first-generation tutorial offers it as the no-install path. ## When it fits No GPU / no install (try-before-install), or burst capacity beyond local hardware — vs. local installs in [[syntheses/install-methods-compared]]. --- title: "Activity Log" type: log --- # Activity Log Append-only record of all wiki changes. ## Format Each entry follows this format: ``` ### YYYY-MM-DD HH:MM — [Action Type] - **Source/Trigger**: what initiated the action - **Pages created**: list of new pages - **Pages updated**: list of updated pages - **Notes**: any contradictions flagged, decisions made ``` --- ### 2026-04-08 00:00 — Setup - **Source/Trigger**: Repository initialized - **Pages created**: index.md, log.md, dashboard.md, analytics.md, flashcards.md - **Pages updated**: none - **Notes**: Empty knowledge base ready for first source ingestion --- ## 2026-06-10 — removed Obsidian scaffolding from the served wiki Deleted `analytics.md`, `dashboard.md`, `flashcards.md` (Obsidian plugin pages — Dataview/Charts View/Spaced Repetition markup, unusable when served as plain Markdown to agents) and the `journal/` scaffold (template only). `CLAUDE.md` directory layout updated: production/planning material lives at repo root, never under `wiki/` (everything under `wiki/` is served publicly). --- title: "Release Detail — Patch Trains & Themes (v0.19–v0.24)" type: summary tags: [releases, changelog, detail] updated: 2026-06-09 confidence: high sources: [raw/github_release-v0-19-1.md, raw/github_release-v0-19-2.md, raw/github_release-v0-19-3.md, raw/github_release-v0-20-1.md, raw/github_release-v0-21-1.md, raw/github_release-v0-24-0.md] --- # Release Detail — Patch Trains & Themes The lower-altitude companion to [[summaries/release-digest]]: what *patch* releases actually carry. ## Per-patch notables - **v0.19.1** — Ernie image fix; Ministral text-gen working; Sonilo Audio partner nodes; templates v0.9.50 - **v0.19.2** — **OOM regression fix in `_apply()` for quantized models**; option to disable the default template in the text-gen node; `ErnieTEModel_` naming fix - **v0.19.3** — `use_default_template` for LTX text-gen; StabilityAI price-badge fix; new arrow-1.1 SVG partner models - **v0.20.1** — **SUPIR model support**; Ernie inference speedups; internal files-endpoint annotation fix; Intel portable link - **v0.21.1** — partner-node wave (Flux2Image, GrokImageEditV2, ByteDance Seedream V2, OpenAI Image w/ DynamicCombo); **fp8 safetensors saving fix** - **v0.24.0** — birefnet/dino3 dtype fixes; triposplat preview fix; Math Expression overflow fix ## Recurring themes a maintainer-watcher should know 1. **Partner/API nodes ship continuously** (Sonilo, Seedream, Flux2, Grok, OpenAI Image…) — hosted-model nodes are a steady release stream alongside local features. 2. **Workflow-template version bumps ride almost every release** (v0.9.50 → .57 → .77…). 3. **Quantization/dtype fixes are a constant patch category** (OOM in quantized `_apply()`, fp8 saving, cast fixes) — if you run quantized models, patch releases matter to you. ## Related [[summaries/release-digest]] · [[syntheses/update-breakage-casebook]] --- title: "Release Digest — v0.19.0 → v0.24.0" type: summary tags: [releases, versions, changelog] updated: 2026-06-09 confidence: high sources: [raw/github_release-v0-19-0.md, raw/github_release-v0-19-1.md, raw/github_release-v0-19-2.md, raw/github_release-v0-19-3.md, raw/github_release-v0-20-1.md, raw/github_release-v0-21-0.md, raw/github_release-v0-21-1.md, raw/github_release-v0-22-0.md, raw/github_release-v0-23-0.md, raw/github_release-v0-24-0.md] --- # Release Digest — v0.19.0 → v0.24.0 (Apr–Jun 2026) ComfyUI ships **fast** — ten releases in under two months across this window (full PR-level changelogs in each `raw/github_release-*` file). Current as of fetch: **v0.24.0 (2026-06-03)**. ## Cadence & highlights | Release | Date | Notables (from the changelog) | |---|---|---| | v0.19.0 | 2026-04-13 | LongCat-Image edit; dynamic-VRAM-disable warning; Manager bumped to 4.1b8 | | v0.19.1–.3, v0.20.1 | Apr–May | patch trains following each minor | | v0.21.0 | 2026-05-11 | video loader reads audio+video together; meta-tensor CPU TE crash fix; CI dispatch to Comfy Cloud | | v0.22.0 | 2026-05-20 | model unique-identifier tests; audio-node robustness; workflow templates v0.9.77 | | v0.23.0 | 2026-06-01 | MediaPipe face detection; **multi-threaded model load from disk (big load-time speedups + offload-to-disk)**; repo security work | | v0.24.0 | 2026-06-03 | cast/dtype fixes (birefnet, dino3); triposplat preview fix; Math Expression overflow fix | ## Why this page matters Version churn is the top trigger for "it worked yesterday" breakage — updates change PyTorch/deps and node behavior ([[syntheses/troubleshooting-playbook]]). Before/after updating, skim the release notes here or in the repo's releases tab. --- title: "Tutorials Catalog — What Official Walkthroughs Exist (128)" type: summary tags: [tutorials, catalog, models] updated: 2026-06-09 confidence: medium sources: [raw/github_doc-tutorials-flux-flux-2-dev-mdx.md, raw/github_doc-tutorials-video-wan-wan2-2-animate-mdx.md, raw/github_doc-tutorials-llm-gemma4-gemma4-mdx.md, raw/github_doc-tutorials-utility-face-detection-mediapipe-mdx.md, raw/github_doc-tutorials-basic-text-to-image-mdx.md] --- # Tutorials Catalog docs.comfy.org ships **128 step-by-step tutorials** (all in `raw/github_doc-tutorials-*`). The category map: | Category | Count | Flavor (sampled) | |---|---|---| | **partner** | 49 | hosted partner-node workflows (the API-node ecosystem — [[summaries/release-detail]]) | | **video** | 22 | e.g. **Wan-Animate**: animate/replace characters from a performer video, preserving expression+lighting | | **image** | 19 | model-specific image workflows (Qwen-Image, etc.) | | **utility** | 9 | e.g. **MediaPipe face detection** (BlazeFace + FaceMesh v2 + ARKit-52 blendshapes, native since PR #14009) | | **flux** | 8 | e.g. **FLUX.2** (BFL): 4MP photorealism, up to 10 reference images, pro text rendering | | **basic** | 7 | text-to-image, img2img, inpaint, outpaint, LoRA, multi-LoRA, upscale ([[concepts/tutorials-basics]]) | | **controlnet** | 5 | depth/pose/T2I-adapter/mixing | | **audio** | 4 | ACE-Step, Stable Audio | | **llm** | 3 | e.g. **Gemma 4** — ComfyUI's default Text Generation model (text/image/video/audio understanding + tool use) | | **3d** | 2 | Hunyuan3D, TripoSplat | ## How to use this catalog When a user asks "how do I do X in ComfyUI" — check this map first: if a model/technique has an official tutorial, the answer is "load that tutorial's template + its model files" rather than hand-building. Each tutorial doc embeds the workflow, model links, and parameters. --- title: "Install Methods Compared — Portable vs Desktop vs Manual vs CLI vs Cloud" type: synthesis tags: [install, decision] updated: 2026-06-09 confidence: high sources: [raw/github_doc-installation-comfyui-portable-windows-mdx.md, raw/github_doc-installation-desktop-windows-mdx.md, raw/github_doc-installation-desktop-linux-mdx.md, raw/github_doc-installation-manual-install-mdx.md, raw/github_doc-comfy-cli-getting-started-mdx.md, raw/github_doc-get-started-cloud-mdx.md, raw/github_doc-manager-install-mdx.md] --- # Install Methods Compared | | Portable (Win) | Comfy Desktop | Manual | comfy-cli | Comfy Cloud | |---|---|---|---|---|---| | Platforms | Windows x64 | Win 10+ (x64/ARM64), macOS 13+ (Apple Silicon); **Linux unsupported** | Win/macOS/Linux | Win/macOS/Linux | browser | | Setup effort | extract & run | installer | venv + git + pip | one CLI | none | | Python env | embedded (`python_embeded`) | managed per instance | yours (venv/conda) | managed | n/a | | Manager | enable manually (`--enable-manager`) | **included, enabled** | enable manually | — | pre-installed platform | | Multi-instance | one per extract | **yes — built for it** | manual | snapshots/bisect | n/a | | Updates | `update/` batch scripts | launcher-managed | `git pull` | `comfy` update | always current | ## Decision rules - **Windows beginner with a GPU** → **Portable**: zero environment work; mind the GPU-matched download (CUDA 13.0 vs cu126 for 10-series) and avoid the dependencies update script ([[concepts/install-windows-portable]]). - **Want multiple ComfyUI instances / cleanest GUI experience** → **Desktop** (Manager already on). Not on Linux. - **Linux, or you want control** → **Manual** (venv discipline is mandatory) or **comfy-cli** — the CLI adds snapshot/bisect, which doubles as the update-breakage debugging tool ([[syntheses/troubleshooting-playbook]]). - **No hardware / try first** → **[[entities/comfy-cloud]]** (400 free credits/month). - **Automation/scripting against hosted models without local install** → `comfy generate` partner nodes (Beta — [[concepts/install-manual-and-cli]]). ## The cross-cutting trap Whatever the method, **each install is its own Python environment** — breakage usually means *that environment's* torch/deps changed, not your system Python ([[syntheses/troubleshooting-playbook]] § CUDA). --- title: "Model & Hardware Compatibility — Known Reported Patterns" type: synthesis tags: [compatibility, mac, amd, models, hardware] updated: 2026-06-09 confidence: low sources: [raw/github_issue-flux-color-noise-instead-of-picture.md, raw/github_issue-trying-to-convert-float8-e4m3fn-to-the-mps-backend-but-it-do.md, raw/github_issue-after-the-last-update-black-images-after-few-steps-on-mac-mi.md, raw/github_issue-support-flux1-bnb-nf4.md, raw/github_issue-stable-cascade.md, raw/github_issue-new-model-hidream-i1-full.md, raw/github_issue-hidream-e1-isn-t-working-at-all-comfyui-portable.md, raw/github_issue-zluda-support.md, raw/github_issue-pinned-memory-causes-error-with-gguf-model.md, raw/github_issue-slow-generation-times-in-flux-using-some-loras.md, raw/github_issue-headdim-should-be-in-64-96-128.md, raw/github_issue-xpu-out-of-memory-error-after-loadimage-and-ksampler.md, raw/github_issue-undefined-symbol-ijit-notifyevent.md] --- # Model & Hardware Compatibility — Known Reported Patterns **Read this page's confidence honestly: these are *reported patterns* from the issue tracker, not verified fixes.** Value: if you hit one of these, you're not alone, and the issue thread is the place to check for the current state. ## Apple Silicon / MPS - **Flux on Mac producing color noise or barely-visible images** — multiple reports (M3 Max/Sonoma: `dev` = RGB noise, `schnell`/fp8 = heavy noise). Related: post-update **black images** on M2 had maintainer guidance (PyTorch nightly, `--force-upcast-attention` — [[syntheses/troubleshooting-playbook]]). - **`float8_e4m3fn` is not supported on the MPS backend** — fp8-quantized models can fail on Mac; pick non-fp8 variants. ## New-model support follows a request → native-support cycle Stable Cascade, FLUX nf4, and HiDream I1/E1 all appear as "please add support" issues before (or without) native support landing. **Rule:** if a brand-new model fails to load, check the tracker for its support status before debugging your setup — and check [[summaries/release-digest]] for when support shipped. ## GPU edge cases (reported) - **ZLUDA (CUDA-on-AMD)** — unsupported community territory; threads show hand-patched `model_management.py` experiments, not official support. - **Pinned memory + GGUF** — load errors reported (also on non-GGUF workflows, RTX 4060). - **Intel XPU** — out-of-memory reported after LoadImage + KSampler. - **`undefined symbol: iJIT_NotifyEvent`** — Intel oneAPI/MKL symbol error reported at startup. - **`headdim should be in [64, 96, 128]`** — attention-kernel constraint surfacing with certain model/attention-backend combos; uninstalling xformers/SageAttention didn't resolve in the report. - **Flux + certain LoRAs slow** — reported as not LoRA-size-related; cause under discussion. ## Related [[syntheses/troubleshooting-playbook]] (solved problems) · [[concepts/models-and-files]] --- title: "Troubleshooting Playbook — Solved Problems from the Issue Tracker" type: synthesis tags: [troubleshooting, errors, cuda, vram, updates] updated: 2026-06-09 confidence: medium sources: [raw/github_issue-assertionerror-torch-not-compiled-with-cuda-enabled.md, raw/github_issue-torch-not-being-compiled-with-cuda-enabled.md, raw/github_issue-issues-compyui-can-t-find-the-ckpt-name.md, raw/github_issue-i-can-t-see-the-comfyui-manager.md, raw/github_issue-sos-after-the-update-the-picture-that-comes-out-is-all-black.md, raw/github_issue-after-the-last-update-black-images-after-few-steps-on-mac-mi.md, raw/github_issue-vaedecode-decode-using-an-additional-10gb-of-vram-when-execu.md, raw/github_issue-update-broke-pulid.md, raw/github_issue-import-failed-comfyui-pulid-flux-enhanced.md, raw/github_issue-importerror-cannot-import-name-get-full-repo-name-from-huggi.md, raw/github_issue-error-no-naistyles-csv-found-when-connect-comfyui-web.md, raw/github_issue-hip-error-invalid-device-function-when-running-comfyui.md, raw/github_issue-error-first-try-sd3-directml-rx580.md, raw/github_issue-an-error-happened-while-trying-to-locate-the-file-on-the-hub.md, raw/github_issue-announcement-the-frontend-will-no-longer-be-shipped-in-the-m.md, raw/github_doc-manager-troubleshooting-mdx.md, raw/github_doc-troubleshooting-custom-node-issues-mdx.md, raw/github_doc-comfy-cli-troubleshooting-mdx.md] --- # Troubleshooting Playbook Symptom → cause → fix, from *solved* tracker issues (45 in `raw/github_issue-*`) + the official troubleshooting docs. The meta-pattern: **most breakage follows an update** — check [[summaries/release-digest]] first. ## "AssertionError: Torch not compiled with CUDA enabled" **Cause:** an update pulled the default (CPU-only) PyTorch — typically after running the portable's `update_comfyui_and_python_dependencies.bat` or a stray `pip install torch`. **Fix (maintainer):** reinstall the CUDA build in the install's Python — the CUDA-indexed torch install command in [[concepts/install-windows-portable]] § Requirements. (Two separate well-discussed issues, same root cause.) ## Loader shows "can't find the ckpt_name" / model dropdown empty **Fix:** put the checkpoint in `ComfyUI/models/checkpoints/` and **refresh the page** — [[concepts/models-and-files]]. ## Black images after a few steps (esp. Mac, post-update) **Maintainer guidance:** update PyTorch to the latest nightly; if it persists, launch with `--force-upcast-attention`. Seen on Apple Silicon (M2, Sonoma) after updates. ## Huge VRAM spike at VAE decode (OOM at the end of generation) **Community-verified mitigation:** use **`VAEDecodeTiled`** instead of `VAEDecode` (measured 23 GB → 18 GB in the issue); tiled decode trades speed for memory. ## "(IMPORT FAILED) " at startup / update broke a custom node **Cause:** the node's Python deps don't match the updated environment (e.g. `ImportError: cannot import name 'get_full_repo_name' from 'huggingface_hub'` — a version mismatch). **Fix:** reinstall/update the node via the Manager, or pin/install its requirements into ComfyUI's own Python env; check the node's repo for the post-update fix. Official docs: `troubleshooting/custom-node-issues` + `manager/troubleshooting` (in `raw/`). ## Manager button missing **Seen when:** the frontend/Manager versions drift (the frontend ships as a separate pip package since Aug 2024 — announcement in `raw/`). Ensure Manager is **enabled** for your install type ([[concepts/custom-nodes-and-manager]]) and the frontend package is current. ## Hub/download errors ("cannot find the requested files in the local cache") **Cause:** a node tried to fetch from Hugging Face without connectivity/cache. **Fix:** pre-download the model and place it locally ([[concepts/models-and-files]]); check proxies/firewall. ## AMD: "HIP error: invalid device function" (Linux) / DirectML errors (RX580-class) **Pattern:** ROCm/arch mismatches — verify the ROCm version matches the official requirement (ROCm 7.2 per [[concepts/install-windows-portable]] § Requirements); older cards via DirectML have limited support (per the SD3/RX580 issue). ## Stray "Error. No naistyles.csv found" **Cause:** a custom node from an imported workflow expects its data files. **Fix:** identify the node that owns the file and reinstall it fully — or remove it from the workflow. ## General diagnostics - Update-related? → [[summaries/release-digest]], then the tracker for your exact version. - Custom-node-related? → start ComfyUI with all custom nodes disabled, re-enable in halves (official custom-node-issues doc's bisection approach; `comfy-cli` has `bisect` — [[concepts/install-manual-and-cli]]). --- title: "Update & UI Breakage Casebook" type: synthesis tags: [troubleshooting, updates, frontend, diagnosis] updated: 2026-06-09 confidence: medium sources: [raw/github_issue-copy-paste-of-nodes-stopped-working.md, raw/github_issue-comfyui.md, raw/github_issue-unet-loader-node-broken.md, raw/github_issue-controlnet-object-has-no-attribute-device.md, raw/github_issue-exception-message-typeerror-cannot-delete-property-value-of-.md, raw/github_doc-troubleshooting-custom-node-issues-mdx.md, raw/github_issue-update-broke-pulid.md, raw/github_issue-face-detailer-differentialdiffusion-object-has-no-attribute-.md, raw/github_issue-torchcompilemodelwanvideo-failed-to-compile-model.md] --- # Update & UI Breakage Casebook The "it worked before the update" class, beyond [[syntheses/troubleshooting-playbook]]. The maintainers' diagnostic moves, observed across these threads: ## The three standard diagnostic questions (use them first) 1. **`--disable-all-custom-nodes`** — does the issue persist? (Asked verbatim on the copy/paste-broken thread.) Custom nodes cause most UI breakage. 2. **Which exact ComfyUI + ComfyUI_frontend versions?** — Settings → **About** panel (asked on the post-update workflow-broken thread). Frontend and core version drift is its own failure class since the frontend became a separate pip package. 3. **What's inside `custom_nodes/`?** — stray non-node directories break loading (the unet-loader thread was fixed by *removing* foreign dirs like `dlib`/`huggingface` that had been dumped into `custom_nodes/`). ## Case patterns (reported) - **Copy/paste of nodes stopped working** → custom-node interference suspected; bisect. - **Post-update: workflow won't load, connections/groups gone** → core/frontend version mismatch; check About, update both. - **Widget TypeErrors (`Cannot delete property 'value'`)** → frontend/custom-node JS conflicts after updates. - **`'ControlNet' object has no attribute 'device'`** → version-specific regression class; check tracker for your version. ## The "core update broke my custom node" signature A recognizable class: core changes a function signature, custom nodes calling it crash with **positional-argument errors** — e.g. PuLID's `forward_orig() takes from 7 to 9 positional arguments but 10 were given`, and FaceDetailer's `'DifferentialDiffusion' object has no attribute 'apply'` right after a ComfyUI update. **Fix path:** update the custom node (its repo usually patches within days), or pin/roll back core until it does. For torch.compile failures (e.g. WanVideo), the maintainers' first ask is your **PyTorch version** — try the nightly. ## The rule Update breakage = bisect custom nodes first, then version-match core+frontend, then check the tracker for your exact version ([[summaries/release-digest]] tells you what changed when).