nerdyrodent / VQGAN-CLIP

Just playing with getting VQGAN+CLIP running locally, rather than having to use colab.
Other
2.59k stars 427 forks source link

Blocky output with unknown cause #102

Open gertdeschepper opened 2 years ago

gertdeschepper commented 2 years ago

Hi. Let me start by saying this project is great. Using it to generate larger images using cpu after experimenting to find interesting texts with gpu (limited by RAM). So far so good. But for some reason the files I get are always very "blocky" (the slightly burned woodgrain-ish texture permeating the image) image even after 500 or 1000 iterations. I seem to be unable to get "clear" images. The above was generated with python generate.py -cd cpu -s 600 294 -opt Adamax -p "sea Tri-X 400 TX" -o sea_trix_Adamax.png but neither the "Adamax" nor the "Tri-X 400 TX" nor the use of cpu is the culprit (I am using lots of permutations to find the cause) Anyone out there with an idea ? Regards, Gert

altbelly commented 2 years ago

Also unable to get clear images, all results have a characteristic blocky texture. I'm using the stock model checkpoint from the readme setup instructions. Here's an example, prompt was something like: "silhouette of a man watching sunset over mountain city." Running on 32 GB Nvidia GPU setup per instructions. IMG_20211226_225051

Muwusiz commented 2 years ago

i third this. i was on a RTX A6000 was using the following command: python generate.py -p "sunset" -s 1250 1250 -o output/a_beautiful_sunset.png. i found the max resolution you can use without getting the blocky effect is 850x850

a_beautiful_sunset