Guides
How to Add Cloud Burst to a ComfyUI Pipeline Without Rewriting Workflows
· RenderBob team
The promise of a local+cloud expansion pipeline is simple: baseline load on owned hardware, peaks on rented cloud nodes, and artists who never have to think about which is which.
A local+cloud expansion pipeline means baseline load on owned hardware, peaks on rented cloud nodes, and artists who never have to think about which is which. The failure mode is equally simple: two disconnected setups, two ways to submit, and a team that treats "the cloud one" as a separate, scary machine. Add burst capacity so it feels like one farm.
Keep one submission layer
Artists should submit the same way regardless of where a job runs. If bursting means a different app, a different login, or a manual file upload, adoption dies. The submission point stays put; only the destination of the job changes.
Make deployment a configuration, not a rebuild
The same ComfyUI workflow (same nodes, same model versions, same settings) must run on an owned node and a cloud node without edits. That requires the cloud node to mirror the owned environment: identical model registry, pinned versions, matching custom nodes. Environment drift between local and cloud is the number-one cause of "it worked here but not there."
Solve model availability before the first burst
Cloud nodes need the same checkpoints, LoRAs and custom nodes your workflow expects, staged and version-matched. Pre-sync them; do not discover a missing 12GB checkpoint mid-deadline.
Put the routing decision in one place
Define, in writing, what sends a job to cloud: a VRAM ceiling the local node can't meet, a full local queue during a delivery window, or a resolution/length threshold. That is your overflow rule (see post #20). It should be a policy, not an artist's judgement call at 9pm.
Cap the spend
Cloud burst without a ceiling produces a surprise invoice. Set a hard fleet-size cap, per-job node limits, budget alarms with named recipients, and automatic node shutdown on idle and on job completion. The studio should own its cloud account and see the bill directly: no hidden compute margin, no incentive to run inefficiently.
Report on both halves
One monthly view of utilisation and cost-per-frame across owned and cloud nodes keeps the pipeline honest and shows exactly when bursting paid for itself.
The heavy jobs quietly find capacity, the baseline jobs stay home, and nobody re-learns their tools. That is the whole point of adding burst.
More from the blog
- The Launch-Flag Field Guide: Community Command-Line Hacks for Low VRAM
Beyond block swap and GGUF, a second layer of VRAM control lives on ComfyUI's command line. Here is what the widely shared flags actually do, including where current defaults have moved.
- How to Run a 744B-Parameter Model on Hardware You Already Own: A Colibrì Walkthrough
Colibrì runs GLM-5.2, a 744-billion-parameter Mixture-of-Experts model, on a 12-core laptop with 25GB of RAM and no GPU by streaming experts from disk.