A ComfyUI image file can carry more than visible pixels, and a saved workflow may not be enough to reproduce a result on another computer. ComfyUI has an editable graph format, an execution-oriented API format, and optional workflow information embedded in generated files. Those are different deliverables. This guide uses a fictional postcard illustration to build a transfer checklist, explain what is saved and what is not, and prevent accidental disclosure of client prompts or local file names. It complements a beginner’s generation tutorial by focusing on handoff, not on building the first node graph.
Understand the three artifacts before exporting
The official Workflow API Format documentation distinguishes an editable save-format JSON from an API-format JSON. A regular workflow preserves the graphical editor state—nodes, connections, groups and layout—so a designer can inspect it. API export preserves executable node types, inputs, and connections while omitting positions and visual layout. A generated PNG can optionally embed workflow and execution metadata. These are not interchangeable guarantees: a file can be a useful image without a workflow, and a valid JSON graph can be impossible to run if the receiver lacks a required model. Label the deliverable “editable graph,” “API payload,” or “review image” instead of using the vague word export.
Save a human-editable graph for collaborators
For a design handoff, use the regular save operation to preserve the layout and explanatory groups. Keep the graph small and readable: Load Image, model setup, prompt and conditioning, sampler, and Save Image are conceptual stages that can vary with the actual model. Rename graph groups by purpose rather than hiding required decisions in cryptic node labels. Save a clean baseline before experimental changes. When reviewing a graph, look for dangling links, hidden parameters, and unused nodes that may make the next person think they matter. Never claim an exact output can be regenerated merely because node positions and widget values were saved; the checkpoints, custom nodes, versions, input images, and runtime environment still matter.
Export API format for scripts and backend integrations
The ComfyUI developer reference documents File → Export Workflow (API), which produces a JSON structure keyed by node IDs for server submission. This is not the best artifact for teaching a visual artist because layout information is missing. An application integrator should validate required input nodes, permitted model identifiers, output locations, and limits before accepting an API workflow. A JSON file is not automatically safe just because it has no executable Python code in its text. Extensions, custom nodes, and external endpoints can perform actions. Ask a technical owner to review the dependency list and access controls before connecting unfamiliar graphs to a server that has production data or credentials.
Inspect image metadata before sharing a preview
ComfyUI’s Workflow Metadata documentation explains that built-in PNG saves can contain `workflow` and `prompt` JSON fields. Those fields may reveal source text, local paths, model names, seeds, and project-specific details. A customer who only needs to approve the postcard artwork should normally receive a deliberately cleaned image rather than a graph-bearing production asset. Open a copy in a metadata inspector you control and confirm what is present before delivery; do not assume renaming an image removes embedded data. The documentation describes a server setting to disable saving prompt metadata, but that also removes convenient drag-to-recover behavior. Choose this trade-off deliberately.
Write a dependency manifest outside the picture
A workflow file does not bundle model checkpoints, model licenses, input images, or custom node packages. Create a separate handoff inventory with model name and version, checksum where available, license, custom node source and revision, expected folder paths, input asset permissions, required hardware, and a known good sample. For the postcard example, record whether the input is a licensed shoreline photograph or original sketch. Do not include credentials, customer names, private endpoint URLs, or unauthorized checkpoints. Before installing missing community nodes, review their repository, maintenance status, and permissions. An unknown node should never be automatically installed merely because a shared JSON expects it.
Test the recipient’s round trip before declaring success
An effective handoff test is: open the editable graph, inspect the expected node classes, confirm files resolve, export or run on a safe test input, and compare the result against a defined visual acceptance checklist. A legitimate changed model version may alter the output even with identical settings. For the fictional postcard, check shoreline placement, palette, headline space and aspect ratio, not pixel-perfect identity unless the production process truly guarantees it. Log the failure type—missing model, invalid link, changed sampler, rights issue, or unexpected generated content—then resolve that dependency. Do not report “reproduced” when only the graph was opened successfully.
Pick the right handoff package for the receiving team
A designer needs an editable graph plus an approved flat image and a source/license record. An engineer needs the API-format graph, a schema or example payload, dependency versions, and safe error handling. A client reviewer usually needs only the approved image with private workflow metadata removed, plus a brief disclosure when appropriate. Never send a company’s complete internal workflow or raw input images just to share a preview. Keep distinct backups of the working graph and approved export because ordinary image recompression can strip metadata. The simplest production package is small, documented and explicit about what it cannot reproduce across environments.
How We Research
We checked the Comfy-Org official Save Format vs API Format specification and workflow metadata explanation on October 28, 2026. They distinguish graph formats, describe embedded PNG fields, and note that workflows omit checkpoints, input assets, and custom packages. The postcard handoff and dependency manifest are editorial examples; we did not install community nodes or run a transfer benchmark. Exact menu locations and metadata behavior can vary by version or save node, so inspect the current software and target file before delivery.
Frequently Asked Questions
No. UI workflow JSON preserves visual graph layout; API JSON describes execution inputs and connections.
Yes. Built-in saved PNG images can carry prompt and workflow metadata unless disabled or removed.
No. Model files, input assets and custom node packages are not bundled with the graph.
Not by itself. Versions, models, inputs, seeds and environment can change behavior.
An approved image with reviewed metadata, not necessarily your editable production graph or private assets.