ComfyUI can save node graphs as JSON and retain workflow metadata in generated files, making handoff look like sending one file. But a graph describes relationships, not the entire runtime. Model weights, custom-node versions, input assets and hardware differences can still change the output—or prevent the workflow from running.
JSON preserves the graph; metadata provides a way back
ComfyUI documents workflows as node graphs that can be saved to JSON and recovered from generated files containing workflow metadata. Built-in templates also provide starting points for common routes.
These capabilities support version review, troubleshooting and handoff because a recipient can inspect how nodes connect instead of receiving an unexplained final image.
The environment—not the nodes—is easiest to lose
A shared node name does not guarantee two machines load the same implementation. Custom nodes can change interfaces, model files may be identified only by nicknames, paths may point to a personal directory, and library versions beyond the seed can change results.
Each approved workflow should become a release package: lock the ComfyUI version, custom-node commits, model checksums, input hashes and required environment values, then attach expected outputs or tolerance notes.
The first acceptance pass checks whether JSON loads, nodes resolve and paths are valid. The second runs fixed inputs and compares output specification, key visuals, elapsed time and errors. Only the third tests whether another team member can replace project assets without altering the core route.
Passing only the first stage produces a browsable diagram, not a production workflow.
What remains unknown: Even a locked environment can produce pixel differences because of GPUs, low-level libraries and nondeterministic operations. Production reproducibility should define narrative and delivery tolerances, not demand pixel-identical output.