Guides
How to Make a ComfyUI Workflow Reproduce on Every Artist's Machine
ยท RenderBob team
"It works on my machine" is a punchline in software and a genuine crisis in a ComfyUI studio.
"It works on my machine" is a punchline in software and a genuine crisis in a ComfyUI studio. A workflow that renders perfectly on one artist's 4090 and fails on a colleague's identical-looking setup is version drift, and it is fixable.
Pin everything that can change
ComfyUI output depends on a long chain: the ComfyUI version, custom node versions, model and LoRA files, the sampler, the seed, and the exact settings. Change any one and the result can shift or the job can crash. Reproducibility means pinning all of them, not just the seed.
Use a shared model registry, not local folders
When each artist keeps their own /models folder, you get duplication, subtle version mismatches, and outputs nobody can recreate. A single source of truth for checkpoints, LoRAs and custom nodes, versioned and referenced by every workstation, removes the most common cause of "works here, not there."
Capture full run state, automatically
The parameters that made a good frame are gone the moment the artist moves on, unless someone manually saved and named the workflow JSON. Capture every run's complete state (nodes, settings, models, inputs, outputs) so any teammate can open it, change one value, and rerun. That turns "give me more like last week's" into a solvable request instead of an archaeology project.
Match the environment across local and cloud
If you burst to cloud nodes, they must mirror the pinned local environment exactly. A cloud node running a newer custom node version will produce a different result, or a crash, and you will lose an afternoon finding out why.
Separate precision from taste
NVFP4/FP8 can change output subtly. Decide precision per workflow, pin it, and don't let it vary artist to artist. Otherwise "reproducible" quietly stops being true across your own hardware.
You cannot nag artists into reproducibility. Pin it in the pipeline. When the model registry, the versions, the run history and the environment are shared by the pipeline itself, the same workflow produces the same frame on an owned node, a colleague's workstation, or a cloud node during a burst.
More from the blog
- A Risk Ladder for AI in Documentary
Screenweaver's 8 October guide ranks five documentary uses of AI by risk, from archive restoration to a synthetic face, and pairs them with EU disclosure rules now in force.
- The B-Roll Gap: Generated, Selected, or Shot
Every cut eventually needs a shot that does not exist. AI can generate it or search a library for it, and stock libraries are answering with different rules. Here is how to choose.