Case study
Case Study: When One 4090 Workflow Broke on Everyone Else's Machine
ยท RenderBob team
The workflow was beautiful. A senior artist had spent two days tuning a ComfyUI video pipeline: a specific model version, a couple of custom nodes, a LoRA stack, precise sampler settings. Then they handed it to the rest of the team.

Illustrative composite of a widely reported pattern.
The workflow was beautiful. A senior artist had spent two days tuning a ComfyUI video pipeline: a specific model version, a couple of custom nodes, a LoRA stack, precise sampler settings. The output was exactly what the client had approved. Then they handed it to the rest of the team, and it fell apart.
On one colleague's machine it threw a custom-node error. On another it ran but produced subtly different motion. On a third it OOMed on a card that should have had the headroom. Three artists, three different failures, one workflow that worked perfectly on the machine it was born on. The team lost most of a day to it during a delivery week.
Nothing here was mysterious in hindsight. The senior artist's machine had a particular ComfyUI version, particular custom node versions, and models pulled at a particular time. The colleagues' machines had drifted: a newer node here, a slightly different checkpoint there, a different precision setting. In ComfyUI, output depends on the entire chain, and every link had quietly diverged across the room. The "it works on my machine" problem, arriving in a creative studio.
They did not lecture artists about hygiene. They removed the drift. A shared, version-pinned model registry became the single source of truth: every workstation referenced the same checkpoints, LoRAs and custom nodes at the same pinned versions. Full run state was captured automatically, so any artist could open a teammate's run, change one value, and rerun with confidence. Precision was fixed per workflow rather than left to each machine.
After that, the beautiful workflow was beautiful everywhere. The same graph produced the same frame on every owned node, and on the cloud nodes the studio burst to during peaks, because those mirrored the pinned environment too. Reproducibility stopped being a property of one artist's lucky setup and became a property of the pipeline.
As a studio grows past one or two artists, ComfyUI's greatest strength, infinite flexibility, becomes its greatest operational risk. Flexibility without pinning is drift. And drift, discovered mid-deadline, is one of the most expensive failures a generative studio can suffer. The insurance is boring: a shared registry, pinned versions, captured state.
More from the blog
- Critterz and the Retired Model: A Feature-Length Lesson in Tool Dependency
Critterz was pitched as an AI-assisted animated feature in about nine months for around $30 million. Its Cannes story is what happens when the video model in the pipeline is shut down.
- Wan 2.2 on a 6GB Card: A Real Low-VRAM Case Study
A community Wan 2.2 14B workflow generates a second of video on an RTX 3050 6GB in under five minutes. Here is the stacked technique set that makes that possible.