Short answer: yes, once. The workflow came back on a Vast.ai RTX 3090 in Ukraine and made its picture. But to get that one run we rented fifteen Vast.ai machines over two rounds, and the first round, thirteen machines, gave us no answer in either direction.
The workflow
Legendaer's Anima workflow is one of the most downloaded on Civitai: 34,416 downloads when we took it on 7 October. The "Standard" version uses a new 2B image model, Anima, with a small Qwen3 text encoder and seven node packs: rgthree, KJNodes, Impact Pack and Subpack, Image Saver, Lora Manager and Easy-Use. The models come to about 5.6 GB. The licence on the model is non-commercial.
We rebuilt it as published: 416 seconds to set up on a RunPod RTX A5000 in Montreal, 58 seconds to download the models, 56 seconds to render the first picture, and 89 seconds to pack the whole thing.
Round one: thirteen machines, no answer
We wanted to see it restored on Vast.ai, a cheaper marketplace with a much more varied pool of machines. We tried four times, on thirteen machines:
- Several machines wouldn't start the container at all.
- Some started but wouldn't accept a connection.
- Some had a download link too slow to finish in any reasonable time.
- One offer vanished between our search and our order.
- On one, the base image was still being fetched after thirty minutes.
- On one, an RTX 3090 in Czechia, everything went right until the end: the files came back and checked out, the dependencies installed, and ComfyUI would not start.
CUDA unknown error: the host's GPU wouldn't initialise. We confirmed it with our own check on the same machine. The machine was broken; the setup wasn't.
That is not "it fails on Vast", and it is not "it works on Vast". We wrote it down as not measured.
Round two: the machine that worked
Two machines. The first, in California, didn't start: the platform reported a networking error on the container. The second was an RTX 3090 in Ukraine at about 18 cents an hour.
We pasted the restore command from our drive and it ran to the end: 27 seconds checking the machine, 76 seconds downloading, 52 putting files in place, 293 installing packages, 25 starting the app, and 48 for the test render, which produced one picture. From pasting to finished picture: 8 minutes 56 seconds. Our own pass afterwards re-read the six files the setup records, found none different, started the app itself, ran the recipe itself and got another picture. The two pictures are byte-identical to each other.
We didn't compare either with the original from the A5000: it's a different GPU model, and on a different model the similarity is a number we record, not a pass or fail.
The tool on RunPod was 0.1.20 and on Vast.ai 0.1.21. Since 0.1.21 the machine check also asks the GPU driver directly whether it can start CUDA, and stops early if it can't. That is aimed at hosts like the Czech one. We haven't seen the stop fire on a real broken host yet.
What we'd tell someone moving between clouds
- Expect to rent more than one Vast machine. We needed fifteen to get one run. Most of it was the machine, not the setup.
- Watch the first minutes. A host with a dead GPU can cost you the whole download and install before it tells you anything, so don't walk away until the app has started.
- Look at the licence before you move the workflow. The model here is non-commercial, which moving it between clouds does nothing about.
What this doesn't say
One workflow, one machine that worked, one pair of tool versions. It isn't a rate. The second round also cost us only $0.06, but fifteen machines is the real price of the pool at that moment. Our step-by-step guide to Vast.ai covers how to restore there.