Where it stands.
What we promise, and what our own restores on real rented GPUs show.
Every restore we have judged came back and ran again.
In every restore we have judged, the files and dependencies came back, each one checked.
As of 2026-10-11, on real rented GPUs.
Counted here: restores with released versions that reached a verdict under our current standard. Failures of earlier versions that have since been fixed are not listed.
Starter nests, across GPU models
20 restores of our own starter nests on rented GPUs. Every one ended with the workflow producing its image or video again. Times run from the start of the restore command to its last line.
| Environment | Restored on | Restores | Fastest | Version |
|---|---|---|---|---|
| FLUX.2 Klein 4B | RTX 3090, RTX 4090, RTX 5090, RTX A5000 | 10 | 1 min 37 s | 0.1.17 – 0.1.20 |
| Z-Image-Turbo | RTX 3090, RTX 4090, RTX 5090 | 4 | 2 min 12 s | 0.1.18 |
| Wan 2.2 TI2V 5B | RTX 3090, RTX 4090, RTX 5090, RTX A5000 | 5 | 3 min 01 s | 0.1.17 – 0.1.18 |
| Wan Animate 2 | A40 | 1 | 12 min 33 s | 0.1.18 |
Community workflows and hard cases
Setups picked because they are easy to get wrong — community templates, and the ways an environment hides what it depends on. 18 restores; every one came back and produced output again.
| Setup, and what makes it hard | Restored on | Fastest | Version |
|---|---|---|---|
| Anima Workflows Standard A community ComfyUI template with its own node packs and models. | RunPod, Vast.ai RTX 3090, RTX A5000, RTX PRO 4500 Blackwell | 2 min 24 s | 0.1.20 – 0.1.21 |
| ComfyUI Image Workflows Standard A community ComfyUI template with its own node packs and models. | RunPod RTX A5000 | 3 min 17 s | 0.1.21 |
| LTX 2.3 all-in-one video workflow A video workflow with more than a dozen node packs. | RunPod RTX 4090 | 14 min 09 s | 0.1.20 |
| IPAdapter Plus IP-Adapter and CLIP vision models next to the checkpoint. | RunPod RTX A5000 | 1 min 57 s | 0.1.21 |
| ReActor face swap GPU build of onnxruntime, and models the node pack fetches on first run. | RunPod, Vast.ai RTX 3060, RTX 3090, RTX A5000 | 2 min 28 s | 0.1.21 – 0.1.22 |
| A node pack that downloads its own models The models were never placed by the user; the pack fetched them at run time. | RunPod RTX A5000, RTX PRO 4500 Blackwell | 1 min 43 s | 0.1.20 – 0.1.21 |
| A package built from git source No wheel on PyPI: the built package itself has to travel. | RunPod A40 | 5 min 09 s | 0.1.20 |
| Two OpenCV variants installed at once Variants of one package that overwrite each other's files. | RunPod RTX A6000 | 6 min 40 s | 0.1.20 |
| CLIPSeg, model cached in the home directory The model lives outside the ComfyUI folder, in the Hugging Face cache. | RunPod RTX 4090, RTX A5000 | 1 min 23 s | 0.1.22 |
| IP-Adapter + CLIPSeg + depth showcase Three node packs, models in the app folder and in the home-directory cache. | RunPod RTX A5000 | 2 min 45 s | 0.1.22 |
| Handed off to another account One file's licence does not let it travel: the recipient's restore named it, fetched it from its original source and checked it. | RunPod RTX A6000 | 4 min 11 s | 0.1.23 |
On record since we started counting
| Restores that came back and ran again | 101 | on rented GPUs, each judged by the standard in force on the day it ran |
| Fine-tuning environments (kohya_ss) | 16 | restored on a new machine, then trained a non-empty adapter |
| Countries the rented machines were in | 24 | United States, South Korea, Canada, Poland, Romania, Norway, Czechia, Sweden, and others |
| GPU families | 6 | A40, GTX, H100, RTX, Tesla, Titan |
| Unfit machines stopped by the pre-restore check | 42 | wrong GPU, driver or CPU for the packed build — refused before anything downloaded |
On a different GPU model
When a restore lands on a different GPU model from the one the environment was packed on, we measure how close the new image is to the original and list the figure here. It is not part of the verdict. On the same GPU model it is.
- 2026-10-03Packed on A40, restored on RTX A5000. The restore came back and rendered again; the new image's similarity to the original (SSIM) was 0.973. On the same GPU model, the line a restore has to reach is 0.98.
These docs describe renest 0.1.23, the latest release.