Low quality results

[DrakeFruit]

my test prompt was car -W 256 -H 256 -n 2 -s 50 -S 10 but the results I get are horrible quality, whats wrong here?

[ghost]

my test prompt was car -W 256 -H 256 -n 2 -s 50 -S 10 but the results I get are horrible quality, whats wrong here?

you need one dimension (W or H) to be at least 512 or higher

[DrakeFruit]

I immediately run out of vram on a 2060 super if I try that

[DrakeFruit]

got it to run but the results still suck

looks closer to craiyon than stable diffusion, did I download the wrong weights? lol

[1blackbar]

test same prompt on discord and post if you want... we cant really tell, without side by side comparison

Ok i tested it with same prompts and settings - yes it is true,the model they released is worse than discord one and i have a feeling they wont release discord one but slightly older one... knowing how business works.

[lstein]

Which weights file are you using, LAION400m or stable-diffusion-v1? Low quality is expected with the former.

[1blackbar]

v1.3 that are about 8gb , but the thing with SD weights is they released model that looks worse than even first version of the bot from discord.
Now they alsoswapped to worse version for 2 days and a lot of people noticed, they might release new version that might not look as good as their own used in dreamstudio.We will see. For now my results are fine with 1.3 but not as good as the ones from bot.
Also i recommended your repo fon SD discord channel to people who want discord bot experience back, they shut down all the bots today.

By the way can you crate identical python script like dream but for img2img ? so you can run model once and rerun it multiple times on image, you just input path to image and image filename and all the rest , pretty much like dream one with image path and prompt.

[DrakeFruit]

Which weights file are you using, LAION400m or stable-diffusion-v1? Low quality is expected with the former.

I'm using v1, the 7gb file

[lstein]

At the dream> command you can include the option -I/path/to/file.png (that's a "dash-eye") and it will modify the image like img2img. The --strength (or -f) option lets you balance how aggressively to modify the image - 0.0 keeps the original; 1.0 ignores the original completely.

I am thinking that the difference in quality is due to the latent diffusion source code using plms and ddim sampling, and not k_lms as the Discord bot used. I do have examples of k_lms implementations and am planning to fold it in.

Did you receive the weights as an early access developer or researcher? There seems to be a leaked version of the weights going around that came from an earlier checkpoint

[DrakeFruit]

I guess I got the leaked one, I thought it was a pre-release or something haha