My favourite webaite for working with Stable Diffusion My favourite service for Stable Diffusion. Really love this website. Good prices for renting GTX4090. All informations are clear. I'm using vast.ai servers for about 4 months already and I didn't encounter any disconnects yet. All negative comments here seems to be written by people with low IQ who can't read FAQ or are too lazy to read things and then cry. Anyway, website is great. All you need to do is to get a server with good learning rate and low ports amount. Then it should work great.

FloatScale.com launched its AI Data… FloatScale.com launched its AI Data Center and soon needed to fill capacity quickly. Vast.ai has the best expertise in this. As early adopters we don't seek perfection just mutual understanding. Vast.ai support helped us out within minutes. You can't get that anywhere else. The world is changing for the better and companies like ours are set to accelerate things together. Thank you!

The best prices and options for… The best prices and options for consumer GPUs and easy to start and scale up. Some machines don't work for various reasons though so you end up having to test and eval them yourself. Support is active and helpful during US office hours, not as much off hours. Still the best place for cheap GPU compute.

Cheap and usable Cheap and usable, can't ask for more. Only downside is expensive data transfer charges. Otherwise fine.

Good intentions and pricing, poor execution Good intentions and pricing, but so far so-so execution. My custom Docker image did not let me use GPU in TF or PyTorch, and yet the site info said it should be compatible (not with newer CUDA drivers, no - it cannot be, as these drivers are not backward compatible). My custom Jupyter token did not work, and they did not know why. They auto-generated Jupyter token even though I specified my own with a properly set env. variable. The SSH tunnel option did not work... the machines refused SSH connections. I got disconnected after a mere hour of working because the PC (which it was, with a GPU that had an actual fan - meaning: retail hardware, not a proper datacenter GPU, double-check please if you are not violating NVIDIA's hardware licensing agreements by hosting retail hardware like that, in a mock-datacenter fashion). Anyway, I pressed "report this machine" button and it went up again soon, so I did not suffer any data loss (maybe the PC's Jupyter notebook server or other system server simply went down, which the website reported in a very hysterical manner that the machine was "no longer part of their network"). A notch better SLA than an anonymous proxy, because you have actual tech support. Extra star given for the CUDA driver compatibility tests that I managed to conduct (the info provided on each server is very detailed, e.g. on max CUDA version). I hope these are just budding problems, and they will get better soon! Update: I've tested also the Interruptible instances. Waited until I got outbid (for cheapest instances it happens after an 1 hour or so). Then tried to transfer the data out from the stopped instance (you pay for storage of non-running instances precisely for that reason), using the Copy functionality (having spinned up another instance as target). It failed with "Invalid input format" error - I had a text file (JSON) and a binary file (a JPEG). So don't choose these instances if you value your data until such bugs are fixed - if you get outbid, you won't be able to access your data anymore (unless you wait for it to be returned by the person who outbid you, which I did). Update: I was unable to make my containers work with their service, because of non-standard execution mechanisms (these containers work perfectly under both docker and k8s). They still kept charging me even though I could not log in. I subtracted 1 star for being unable to withdraw my money (you have to pre-pay at least 5 USD - I paid 10 - and there is no self-service way to get the unused balance back). I subracted one more start for the age of this service. I've found out that it is now ... 5 years old and it's still full of bugs ("Piotr Inny 15 października, 2018 | 8:26 am Ta usługa to faktycznie beta, niestety pełna bugów."