One measured case · 9 October 2026

Packed on RunPod. Restored on a clean Vast.ai instance.

This page is the full record behind the case on the home page. It is one run that we carried out and measured ourselves: the same account packed the nest and restored it, so nothing was handed to anyone. It is not a rate, and it says nothing about machines we did not use.

The two machines and the tool on each

Packed onRestored on
CloudRunPodVast.ai
GPURTX A5000RTX 3090 — the same GPU generation
WhereCA-MTL-1, CanadaUkraine
Date7 October 20269 October 2026
renest tool0.1.20 (pre-release build)0.1.21 (as published on PyPI)

The result holds for this pairing of tool versions only. This page reports that pairing and no other.

What moved

Anima Standard, a community ComfyUI image workflow published on Civitai. The nest held 14 files, 5.67 GB in all: 3 model files, the ComfyUI code, 5 custom-node packs the workflow uses (ComfyUI-Easy-Use, ComfyUI-Image-Saver, ComfyUI-Impact-Pack, ComfyUI-KJNodes, ComfyUI-Lora-Manager), the dependency lock, and the workflow itself. The 164 Python packages in the lock were reinstalled at the packed versions, not copied.

Every stage, as the restore timed it

StageTimeWhat the restore said
S0 checking this machine27 smachine checked; 2 notices, nothing was held back
S1 downloading76 s13 downloaded, 1 already here
S2 putting files in place52 sAll 5 files check out, 6 code folders in place, lockfile back where it belongs
S3 installing dependencies293 sPython environment ready (python 3.11.16)
S4 starting up25 sThe app started and is answering
S5 test render48 sRe-ran the packed recipe and it produced 1 image in 48s
Door to door536 sThe restore reported it as 8 min 56 s; exit code 0.

The first stage printed machine check: warn. Behind it: 14 checks passed, 1 check could not be answered, and there were 2 notices — a GPU driver version different from the packing machine's; a CPU with fewer instruction sets than the packing machine's. The restore went ahead: nothing was held back.

Door to door is the restore command, from its start to its last line; renting and booting the machine and installing the tool are not in it. Starting the rented machine took another 352 s, before the restore began. The stages do not add up to the door-to-door figure exactly: the rest is the time between them.

renest restore · S5 test render Read-only
✓ Re-ran the packed recipe and it produced 1 image in 48s

What was checked afterwards

Files

A separate check, not the product, re-read 5 nest files + 1 dependency lock on the restored machine and compared each one byte for byte with the nest: 6 files compared, 0 differed. It could not re-compare 7 items — the code archives are unpacked and removed during a restore, and the workflow recipe is stored without a fixed path — so for those the only check is the product's own, on the way in.

Output

The restore's last stage ran the packed workflow and produced 1 image. The separate check then started the app itself and ran the workflow again on the same machine: the second image was the same file as the first. The image was not compared with the one made on the packing machine; across different machines we compare files, not images.

How it was run, and what a person would do differently

  • ·The restore code was issued through the service by our test tooling, not clicked in the web console.
  • ·Our tooling chose the Vast.ai machine by price, card, reliability and measured download speed.
  • ·A scripted terminal pasted the console's command block and pressed Enter at the one question.
  • ·The restore asked 1 question: Install them now? [Y/n] — ffmpeg was not on the new machine. The answer was Enter (the default, yes), and the restore reported Installed: ffmpeg. No other command was typed after the paste.

Run the same path yourself.

Each environment we offer is a run we packed after it produced output. Take one and restore it on a machine you rent; the restore prints the same stage lines you see above.