Workflows

Workflows

Saved workflow requirements after checking. Missing models and stopped Comfy keep Run disabled. Captured September 25, 2026.

Workflows is the ComfyUI graph operator: choose a JSON graph, verify its models and custom nodes, then run it. Requirement checks do not load weights. Running does.

Open ComfyUI → Workflows. Keep Engine & queue nearby; Run needs a ready Comfy instance, but you can import and check files while the engine is stopped.

The in-app GPU & API guide (/static/comfy-api-guide.html) is the same contract as docs/comfy-api.md. This page is the GUI how-to.

Three-step strip

1. Choose — import or pick a saved graph
2. Verify — models, nodes, API format, GPU plan
3. Run — queue on a ready instance

If ComfyUI is not running, a warning banner offers Start ComfyUI. Model checks still work. Node checks and Run wait for an instance.

Import a workflow

Do this:

1. Click Import workflow (or use the file field in Add a workflow).
2. Select a .json file exported from ComfyUI.
3. Check Replace a workflow with the same name only when you intend to overwrite.
4. Click Import and check.

Then: the graph appears under Your workflows and Omni inspects requirements without loading models.

API format versus UI format

Comfy's editor JSON is not directly queueable. Export API format from ComfyUI (File → Export (API) in upstream UI). If analysis reports API export needed, open the graph in the native Comfy window from Engine & queue, export API format, and import that file.

Do not paste a screenshot. Do not import a .png workflow image here; Media library can show embedded metadata, but this importer wants JSON.

Your workflows

Each card shows the name, node count, size, and a readiness badge:

Badge	Meaning	What you do
Ready	Models and nodes verified	Plan GPUs, then Run
N models missing	Files not on disk	Download verified when sources are known, or find a source
N nodes missing	Custom node not in the live instance	Install the node on Engine & queue, restart if required, Check again
API export needed	UI-format graph	Re-export API JSON
Review required	Unknown custom model input	Search all folders / find source
Checking	Scan in flight	Wait
Models checked	Files OK, Comfy not up for node verify	Start Comfy, Check again

Run is disabled until ready_to_run is true on a live instance. The tooltip says to resolve readiness first. That is the GUI form of analyze-before-run.

Other card actions:

• Plan GPUs — runs analysis with the current placement policy, still without loading weights
• Check again — refresh requirements
• Download verified (N) — install missing models that have a known source
• Open JSON / Delete — under Requirements and options

Delete is immediate. There is no recycle bin.

Run on

When at least one Comfy instance is ready, Run on picks the instance. That instance also enables live custom-node verification. Auxiliary GPUs are only usable if they are in that instance's gpu_pool. If a plan says restart-required, stop and start Comfy with the missing GPU checked. You cannot hot-add a GPU to a running instance.

GPU placement on this page

Open GPU placement. This is the same single / auto / manual policy object the API stores as workflow metadata (not inside portable graph JSON).

• Mode
  • Single GPU — every component on the preferred GPU
  • Automatic pool — place components across eligible cards when useful
  • Manual / hybrid — lock selected components; the rest are still planned automatically
• Preferred GPU — single mode uses only this card; pool modes prefer it for the main diffusion model. Options use stable GPU UUIDs.
• VRAM reserve (MB) — headroom kept free on every eligible GPU (default 1024).
• Checkboxes — which UUIDs are eligible. Unchecked cards are invisible to the planner even if they exist in the host.
• Advanced: require every selected GPU to receive a component — require_all. Leave off unless you truly need a component on every card. "Use all GPUs" normally means they are eligible, not that the planner must invent dummy assignments.

The hint on the panel is the product rule: Plan checks do not load models. GPU UUIDs are saved in policies so selections survive CUDA index changes.

After Plan GPUs, the card shows a GPU plan with status, component rows, blockers, and warnings. Ordinary routing is component placement (MODEL / CLIP / VAE), not layer sharding of one UNet.

Analyze-before-run and valid: false

Omni always re-analyzes at queue time. You should still press Plan GPUs (or Check draft) and read the plan.

If the plan's valid field is false, or the badge is blocked:

• That is a hard stop. Do not Run.
• Read blockers (missing VRAM, missing pool GPU, host RAM, cgroup budget).
• Shrink the graph, pick a lighter checkpoint, enable a real low/none VRAM instance, or unload other workers.
• Never "just queue it anyway."

valid: false and ready_to_run: false are the same operator meaning: the app is protecting the GPU. Bypass is not offered in the GUI because bypass is how you OOM the WSL VM.

Warnings are not blockers. Foreign compute PIDs, overlap of ordinary loaders, and staged-vs-ordinary notices are warnings. Blockers are blockers.

A Comfy prompt_id on a successful Run means the engine accepted the prompt. It does not mean the PNG exists. Watch Engine & queue, then confirm the file in Media library.

Templates and JSON editor

Open Templates and JSON editor.

• Search installed Comfy templates by purpose, node, or text, then Search templates.
• Filename + textarea: paste API JSON.
• Load safe sample — a tiny known-good graph for wiring tests.
• Check draft — analyze-before-run on the textarea.
• Save JSON — persist to the workflow store.
• Run draft — disabled without a ready instance; still respects a blocked plan.

This editor is for API JSON, not Python.

Missing models and nodes

Inside Requirements and options:

• Missing models with verified source can be downloaded from the card.
• Find source / Search all folders for unclassified inputs. Do not invent a Comfy folder from a display name. Category + relative filename must be exact.
• Missing node types need a custom node install on Engine & queue, then a restart of that instance if the node only loads at process start.

Core Comfy update and Manager update_all are different operations. Do not assume one did the other.

Staged graphs

Staged H3/LTX nodes release their owned models between phases. Generic staged VAE decode releases its own VAE only, leaving ordinary upstream loaders resident. The plan's summary.staged flag tells you whether every heavy component actually has that contract. Ordinary OmniRoute* placement can keep weights resident. If the combined graph cannot fit, use staged nodes or a smaller stack — do not hope auto placement will layer-shard a diffusion model.

Related pages

• Engine & queue
• GPU placement
• Media library — where PNGs land
• Model library — ComfyUI models view

JSON editor example

The example above uses EmptyImage and PreviewImage to explain the API graph format. It is an unsaved draft; this screenshot does not show a queued or completed workflow. Check draft validates readiness before Run draft.