Home Artists Posts Import Register

Content

Hi all. First of all. After messing with my computer in the weekend it seen to be working fine, in fact I spent most of the day working on SD with no problems.

There was already a some fixes, like memory usage for Upscale and some GUI smaller bugs. 

There been some reports that Upscale is slower than 0.41
That is correct. I spent some time optimizing the upscaler for 0.41, but then it was time to put the Face Fix to work, and it was a mess with my optimizations. I could not make it work. So I removed the optimizations and made the face fix a option. I may try to optimize again in the future.


There been problems to make full precision to work as well. I was gonna test it out, but then it hit me. Why not change the fp16 to work as full precision? As far I can tell the only difference is to make the model work on graphic cards that can't handle half P. The results barely change. So I think just gonna make a option to force to run at Full P.


In the next update will also add the option to select an scheduler and Graphic card (if you have more than one)


Plus, there been some changes to img2img and inpaiting. img2img make more sense to use now. It will always run the correct number of steps instead of changing it.

inpaiting now actually look good. I was not pleased with the results so made some changes, now is looking better. There is probably others modifications that can be done to it.

Comments

Anonymous

Which sampler is this running? k_lms?

Anonymous

hi the img 2 img seems much better to me (thumbs up: )

Anonymous

is textual inversion a possibility?

Anonymous

question I've never asked until now - Is the "Grid" check mark working for some people and not others I've always noticed the option is there, but assumed it's something you're adding later. Should it work already? Not really a huge concern, I'd choose to save them individual anyway, but just asking.

rune hansen

will this update come before the next billing cycle?

Anonymous

From my tests I think it is plms, I get same results when I use plms in other software

Anonymous

So I found a major oddity with this GUI, perhaps a bug or maybe just a feature? So I was making prompts and I had a rather long one but it was generating good results. So I decided to copy and paste it into notepad for when I need it again. I went to get groceries so turned off my computer I come back and copy pasted the prompt from notepad into the GUI field and I noticed everything being generated was nowhere even near the quality it was the first time I used it. Curious I tried some things. I thought for whatever reason, maybe I need to manually type this prompt and not copy paste? So I did, and wouldn’t you know it, it generated correctly again. Weird. Also I was using () for words I wanted to see more of and [] for words I didn’t want in the prompt. It works. Though not when copy pasted…. I have no real clue why

Anonymous

cant wait for the steps to go above 500 steps.

Anonymous

Thank you for everything. I have found something that may be for a future GRisk AI SD. https://huggingface.co/Gustavosta/MagicPrompt-Stable-Diffusion It seems that if you enter a little prompt, it will complete the rest of the prompt. GPT-2 model intended to generate prompt text for AI image processing. I know it is not a priority, but I have reported it in case you are interested.

Galewolf

I noticed people are asking on the itch page about updates and stuff - maybe you could post a little explainer on there for them?

DAINAPP

Will see if I can focus a little there, its just to much places to answer users. Usually I start with Patreon and if have the time check the other places.

DAINAPP

That is a pretty clever modification on GPT-2. I will save it for now to check it out later. Thanks for the link

DAINAPP

That is pretty strange, perhaps there is some invisible characters on your paste? like /r or /r. It should not change the result if the text and seed are the same.

DAINAPP

It should be working, but I don't test it very much, so it could be broken. lol Will check it out and fix it today if is not working.

Anonymous

Question about the img2img functionality.. before v0.5 the image set at 70% seemed to get a 70% weight on the full img + prompt session, i.e. 100 steps with a 70% image strength ended up with running 30 steps due to having the 50% weight divided.. But now at v0.5 it looks like the ratio of weights has reversed.. putting in 30% img strength and 100 steps runs 30 steps while 70% img strength runs 70 steps, Not sure, but other implementations behaves as 80% strength in 100 steps executes 20 steps (20%) for the prompt... what is the expected behaviour? Perhaps im doing something wrong :)

Anonymous

He announced that for 0.5 the weights were reversed. It should be fixed in 0.51

DAINAPP

Like Eduardo said, it was reversed on 0.5 Now its fixed and will always run the number of steps you requested, at 10% or 90%

Anonymous

Ah yes, great! Didnt find the announcement for that. Do you plan to use a bug tracker in the future for the patreons? :)