image-upscaling.net

still training... and still months until i expect it to be complete....

still training and it will probably take months until this is ready because it just runs on a single 5090 currently

going into day 4 of training, loss & validation looking good so far.

starting day 3 of training with a new all time low on the loss curve

Training a new upscaling model, sharing results after 1 day of training. 4x rtx5090 cloud instance running at 50$/day

fully local image converter, supports .heic files image-upscaling.net/image-converte…

pypi project for upscaling api pypi.org/project/image-…

Now I'm sidetracked again. Guess what I'm cooking up here...

if you are banned and it is truly an accident, get a new ip and wait 24 hours before you submit a new request. Do not send a pic to your friend and have him upscale it for you. This is considered cheating on the daily free quota

I offer a generous daily free quota. As of now, a single ip can get up to 2 minutes compute time on rtx5090 for free daily. It is important to protect the site from users cheating around this limit. Any ip cheating will be banned. False Positives are < 0.1%. No more exceptions!

sample image generated after 1.9M patches seen. i think the hands look better compared to what the diffuser generates, mostly because the number of fingers is correct

as training goes on, here are evaluation samples taken after around 500k, 700k, and 1M patches seen

first 21k steps (700k patches seen)

still training...

Both images are from small test models after 60k iterations. One is my old model. The other is my new model. both models are similar in size. both are trained using esrgan.

Since the test runs provide very good results, I will start training a 10GB model starting next week on my new rtx5090. I will probably take weeks to months until it is ready

as training goes on, the learned residual maps get more detailed