Mountain river valley, by Ivan Kramskoi and Makoto Shinkai Steps: 2500, Sampler: Heun, CFG scale: 7, Seed: 232122911, Size: 2048x512
Time taken: 4752.52s
are steps not linear? I was kindof under the impression they were, but it seems more steps stays blurry for the longest time, and only gets fine detail at the very end. This one didnt start to really clear up until 2400 steps.
Okay, so here is a problem I have. I generated pic related. Originally it had a fucked up face. So I cropped out a square around the face, dirty upscaled it in paint.net to 512x512 and popped it into img2img, it came out with the face you see. I then placed that back on the image roughly where I wanted it, and at the scale I wanted it. I took the whole thing as you see it, into inpainting. I masked off the face and like, her necklace basically because I didn't want it to touch that. I planned to use a low denoise level, and a similar prompt to what I generated the image with, but then it goes totally nuts and doesn't seem to follow the image at all.
Any advice on img2img based on real people? I have a willing model and a few photos but it's super hit or miss. I'm just masking the face and some hair and letting it fill everything else based on original, if I clear the background it just poops out but 99% of the time it just ends up with a face floating randomly in the image.
Noah Miller
on voldy's colab, it links to a very nice webui that i open in a new tab problem is that i tend to time out when using it because i'm not interacting with the colab notebook any workarounds for this? or do i just have to remember to switch tabs every once in a while
Kayden Cox
this might be the best one it has made for me in a minute
I have 8GB VRAM, but I would like to use one of the optimization settings from voldy's repository just to normally use the rest of my programs. Using medvram and opt-split-attention both drop VRAM usage to ~45%, which is way too low and slow, I am ready to give it some more to get more performance. Which one should I leave enabled?
I find to get anything reasonably interesting, I am pretty often going between img2img and some basic editing in gimp, or photoshop if you have it, whatever. Treating it like a one-stop black box solution is kind of fun for exploring or playing around, but it's better to give it a little guidance with outside tools. Even basic bitch copypasting and smudging in gimp works a lot of the time. Photoshop would make this part exceptionally good. There's also the krita plugin which people seem to love, haven't tried yet. But basically, get involved in the process. Erase those extra faces. Also by playing around, you'll get better at knowing what CFG/denoise/steps should be, when to inpaint, blah blah. Make a batch of 4, take the best. Or a batch of 10, take the best parts from each, and repeat.
Kevin Brown
heun isnt ancestral? I'm testing shit man.
thanks
Jackson Hernandez
I know, but I did a git pull since I made that Brit-waifu and it changed the random seeds, so now I can't make corrective variants of her ;_;
Evan Phillips
Huh. Anons this is a wild one. I trained a textual inversion model on a woman's face to see if it was a viable means of referencing specific people. I ran my next inversion model in the same Colab window and apparently it didn't overwrite the old output file, so I ended up with a duplicate of that first model, which I then downloaded and named vintage-scifi.bin, thinking it was my new model. What's wild is, I never ran that model on it's own, so I didn't realize the mistake. I swear to you both though, I have been running that inversion model as part of my prompts which is rather hilarious. I'm running a batch right now, but as soon as it's done, I'm going to check that same prompt with the faulty model removed and see what kind of difference that makes lmao Interesting, I noticed the opposite when I tested, but I could also be wrong given what I said above regarding the incorrect nature of my one inversion model
I'm not confident in this, but I believe opt-split-attention needs to be run alongside a VRAM specification, so if you're going to run one on it's own, I'd recommend running just --medvram. Conversely, if I'm wrong then --medvram is absolutely what's tanking your usage, and you should run with just opt-split-attention. I'd experiment with both and see which works better.
Cooper Price
Can you sure the actually correct SF model .bin then?
Brandon Robinson
alright. is gimp the free one? I used to have cs3 but that computer caught on fire.
nta. GIMP is free, may be a bit unusual at first, but there is a settings preset for those used to Photoshop. I use it, it's good enough to make asset bash for img2img and make a mask. But CS3 is based, last good version basically.
Ethan Peterson
gimp and krita are free. krita is probably better, I'm just used to gimp and don't do anything very complicated with it. Some people were really raving about the SD plugin for krita, it probably is worth checking out.
Keep in mind, doing an image at 100 steps vs how an image looks at 100 steps when doing an image at 200 steps are not the same thing. It tries to make it converge at the step count you give it.
Also, with an ancestral sampler, 20 steps is often more than enough and over 100 steps is probably meaningless. Ancestral samplers quickly get to the goal and steps after that just change the image. Other samplers move toward a specific goal so more steps means more refined, but ancestral samplers just keep changing randomly through all the steps it goes through. Doing 2500 steps is completely pointless and wasteful. Try the same image at 20 steps or so.
Sebastian Lopez
>I have been running that inversion model as part of my prompts which is rather hilarious
Yeah no this isn't true, because if you had been using it you would have noticed almost instantly. That model inserts her face into at least half the images you generate, even with totally unrelated prompts
I don't know what the point of this ruse is, you wasted like 3 minutes of my time at best and I assure you my time is worthless
Hunter Cruz
Bros how do you cope with the fact that big corporations like google, apple, and ibm have already achieved perfection as far as generative AI while we're still soaking up the scraps because they want to profit off it.
Andrew Allen
user that's the problem, I literally don't have a model that was trained on those vintage sci-fi pictures. Somehow that model of the woman's face is tying into my prompts and not completely fucking them up. That being said, I'm working on retraining a model with the original photos and as soon as I have that, I will absolutely post it here, but it won't allow you to replicate my images since again, I'm producing them using the face model, which is absolutely hilarious
