The problem
You rent the machine by the hour — then spend the first hours making it feel like yours again. Three ways the same evening goes wrong.
The first hours
Fresh machine, empty disk. pip resolves a different set of versions than it did last week, the custom nodes have moved on, and every model downloads again.
The meter runs long before the first render does.
The next card
A different card, a newer GPU generation — and the setup that ran last night stops with an error nobody can explain, on the morning it was due.
You can’t promise a delivery when you can’t promise your own environment.
The monthly bill
It lives in one region, on one provider. It can’t follow you to a cheaper card somewhere else, and it can’t be handed to anyone.
It bills every night, including the nights nothing runs.
The method
A nest isn’t a disk image. It’s the recipe of a run that already worked — the files, the locked dependencies and the workflow — captured once, kept in your drive, and rebuilt on whatever you rent next.
What we promise is the files and dependencies coming back, each one checked. Whether it then runs is something we test on every restore and report back — not something we assume.
Step 01 · Capture
On the machine where a ComfyUI run — or a kohya_ss / LLaMA-Factory fine-tune — already works, one command (or one button in the ComfyUI panel) records the models, the custom nodes, the locked dependencies and the workflow. Every file is recorded by its own content on the way in, so “roughly the same” can’t sneak aboard.
Repacking a ComfyUI + SDXL run whose model file an earlier nest had already stored. Upload segment only, one run. field report
Step 02 · Store
The nest lands in your drive, filed by content instead of by filename. A file that is already stored is never stored twice, and only files the drive hasn’t seen travel. Nothing runs, so nothing bills by the hour.
A 23 GB video nest: the files a failed earlier upload had already stored were not sent again. One run. field report
Step 03 · Check
Before anything downloads, the restore checks the machine you just rented against what the nest was packed on. A pairing that cannot work — say, a GPU generation the packed code was never built for — is refused up front, not twenty minutes into a dependency install.
free diskGPU driverGPU generationsystem libraries the nest needs
Step 04 · Restore
One command on the new machine. It walks these stages in order and tells you which one stopped, if one does:
Every file is checked against its fingerprint as it lands. Then the app starts and your workflow runs once — and at the end it prints where the output is and how to open the app yourself. Lost the machine halfway? The same command works on another one.
Door to door: boot, transfer, dependency install, verification and a first render. The 6.5 GB nest held one model and no custom nodes; the 23 GB video nest was one run on the same GPU model. field report · field report
All of it lands in an open format. Every nest ships a plain restore.sh that brings back its files and dependencies with zero Renest code — so your work never depends on us still being here. the escape hatch
Why it holds
Every figure below comes from a published report, with the hardware and the sample size next to it. The full record of every restore we have run — failures included — is on the Proof page.
Determinism
Every file is checked byte for byte on the way in — that part we guarantee. Images depend on the hardware: on the same GPU model and CPU brand, rebuilt renders have come back as the same file; with a different CPU brand they differ by a hair, the same way every time.
rendered images compared by fingerprint across hosts field report
A different card
Dependencies are installed again for the card you landed on, so a nest isn’t tied to the card it was packed on. Where the model of GPU changes, expect slight, stable differences in the output — and the pre-check refuses the pairings that cannot work at all.
a 23 GB video nest rebuilt on an A10 and on an Ada card guide
Resilience
Uploads and downloads resume instead of starting over, and a file that fails its check is fetched again on its own.
A 23 GB video nest: the files a failed earlier upload had already stored were not sent again. One run. field report
Across machines and regions
Each route below comes from a published report and says whether it timed the whole rebuild door to door or only the transfer.
Field reportsDoor-to-door times include boot, dependency install and verification; the Mbps figure is the transfer segment alone. Single runs — each report states its sample size.
Canada → Sweden13.0 min door to door · 23 GB
US-West drive → Japan867 Mbps transfer
RunPod → Vast7.3 min door to door · 6.5 GB
Questions
Not when the files are back — when the app works. A restore checks every file against its checksum, then starts ComfyUI, runs your workflow once and reports what passed, naming any custom node that failed to load because a system library is missing. Our own results, failures included, are on the Proof page.
Claim a starter nest — a run we packed and checked — restore it on a machine you rent, and read the result yourself. Free, no card.