Field reports

Send someone a ComfyUI workflow and they get the graph. The models are the hard part. We handed one to another account and ran it on their machine.

Face swap is one of the setups people most often lose when a rented machine goes away. Thirteen model files you never placed, a GPU build of onnxruntime, and a licence question.

A video setup is big. We restored one on four cards: four to ten minutes, and the longest wait wasn't the card or the download.

One setup, built once on an A40, restored on an A5000, a 3090, a 4090 and a 5090. The card that renders fastest isn't the one that costs you the most time, and the pictures aren't all equally close.

Five hosts with the same GPU model regenerated one image. The results fell into exactly two byte-identical files — and the CPU vendor, not the driver, sorted them.

Same GPU model: the regenerated video is literally the same file. Different model — even same architecture: a uniform ~0.94 plateau. And the 0.9814 mystery is solved: it was the CPU vendor.

A workflow with 3 custom nodes and a LoRA is captured and revived intact — including a node that was already broken — while new data forces honest revisions to Vol. 1.

July 2026: one ComfyUI + SDXL environment packed, the machine destroyed, and the nest restored 26 times across 2 clouds, 7 datacenters and 6 GPU types, every file verified. The pass line used here has since been replaced by a stricter one; the current record is on the Proof page.