#🏞|general-with-images
1 messages · Page 65 of 1
Andy Dick
tried to get a younger emilio estevez and it's like, it just made the old version look young. it doesn't know the youthful face of emilio we knew in 1985-1993
thats not gonna work, hoss
it sees Steven Segal the way he sees himself
is that him
what
he was in The Breakfast Club too
ew
Resident evil
aw
also hates Ben Stiller 
ur just using bad mdels
models
i'm not interested in SD1.5's garbage outputs
marvel
that's not ben stiller
i think ben should do some surgery and catch up to AI's version
Artius has a different view of him
im i allowed to post this
why would you want to
Unleash Da Fuhrery!
table connisseur prompt?
christ even when i tell it "emilio estevez as marty mcfly" it just makes old emilio
ur model sucks
no, u suck
🥰
the weight of the woreld
literally professor x
prompt?
toddler-- schwarzenegger++
like those
voldemort smoking weed
i love voldemort and snoop together
they're starting a business
spoiler alert: snoop changes voldemort's life
lol that face
spongebob working the front desk at a dispensary selling cannabis in california, 1996
Does anyone have a good source for real 4k faces? I want to use them as realization images.
Typeo
All high res faces?
controlnet scribble
Do you have the server link?

yo
i tried looking at civit for upscalers
but i cant find any
i'm looking for a good sauce for models
you mean wiki upscale
https://upscale.wiki/wiki/Model_Database ?
np
Common L

Jk
punk-ass orlando bloom
punk-ass frodo (lord of the rings)
punk-ass thug santa clause
santa: union buster mode engaged
aw shit, santa's using kids for soldiers
ultimate upscaler works best with tile resample
just did one more with more feather
300 cheeseburgers later
woulda been a wild twist to Breaking Bad if we just watched him chow down on mukbang and get fat
💀 walter been eating at mcdonalds too much

ginger walter white?
Zovyas model is so odd
sounds like it wasn't trained to convergence
or it was over-trained and led to early convergence. honestly there's a lot that can go wrong
i mean atleast i found an area where it works best for me
so as long as i keep getting good consistent results on 3.5 ill be fine
i forgot you don't care about the details, nvm
wym dont care about the details
Is current a1111 version working properly?
It's been some time since I updated it
Yeah
I sure hope they fix open pose, or upgrade it as there are a lot of poses it fails with as it lacks the joints.
that needs more training, lmfao
what's the most any of you have further trained SD1.5 on?
like 10,000 images maybe?
LOL
why is it so uncanny
it's probably better than what my current models do but there's something.... off
short face and almond eyes make Woody look uncanny af damn
that's a real photo of Woody fwiw
it has hyperrealistic proportions
Better?
can you make it so harrison ford (indiana jones) and jack black (woody) are on an adventure?
I'll try
mine got hella confused
Say hello to Kakaroto Jack Black
👀
@oak osprey alright, so, I did have a transient spike last night that was my PSU shutting down for over current, which doesn't cause any damage to the PSU, it's just triggering it's surge protection
Since then I talked to another friend who is a high rank competitive overclocker, and he was able to help me undervolt the hell out of my GPU with no performance hit
Now I am running it with no power limit, and even in the hardest peak load I was able to get (Subnautica below zero, which is what shut my PC off the first time, and I was able to replicate the over current there 3 times) it no longer has any issues
Not only that but my GPU is now extremely efficient, having dropped from 1.245mv at 1830mhz down to only 815mv
Now the highest normal power load I see is like 230 watts, and even cranking it in Subnautica (1080x2560 ultra cranked) I get 180fps and only use 150 watts.
It's a thing if beauty
I get full stock time spy scores for about half the power draw. Looks like I got lucky as hell with this GPU's silicon
That undervolting in Afterburner is so time consuming.
I just set it with the slider not the graph but the big boys use the graph
I wish AB allowed us to use the slider for core voltage to go negative too.
I did my undervolting in after burner in like 10 minutes
Each time I just hit it with a heavy load to make sure it was stable, then at the end I threw Subnautica at it, and it worked without tripping
just, as someone who has built off-grid electric systems. you will do well to note. P (watts) = C (amps) * V (volts), and it's possible that undervolting will result in equal or greater power use because there's current limits involved. i know that undervolting the Vega allows it to pull more power, and qualifies as an overclock.
run the fuzzy donut benchmark
so the reason these cards don't come undervolted like that OEM is because of the silicon lottery, you will likely have better luck with a 3090 undervolting than a 3080 because of superior silicon quality, but a 3080 will do better than a 3070 - and by better i mean operate more stably over a larger range of possible options. that's why the 4090 has a higher range of power limit options than 4080.
I did, it drew way less power this way
sweet
i believe i mentioned that to you at first when you were remarking about the card's high power draw
my PSU can handle a 400 watt consistent load no problem, its just those massive transients that trip surge protection
well yeah but if you get no shutdown, good to go
yep. NVIDIA cheaps the fuck out on their card power management
fuzzy bench isn't something that would shut down, as its a non transient load, but it IS a good test for clock and voltage stability
it's also implemented in this black box thing they call the GSP now. no one knows what it's actually doing
the real transient trippers were subnautica below zero and tile rendering in blender
both of those nuked my PC before, but now they run just fine
I actually have a test thats more intense 😅
btw, i ran my 4090 on a generator once and LOL that thing is transient power spike CITY
I did 3D rendering for 4 years, and I was working with a friend who is a light transport researcher
Stable Diffusion's usage profile of the 4090 is whack
torch.compile fixes that but is not widely compatible with pipelines yet
you can hear the generator surging when the 4090 runs as it tries to keep up with the transient spikes
this bench I made is my GPU cruncher
it's a 2kw generator
i'm just glad i didn't bother trying the 4090 on a 800w PSU because that would never work
that bench I make hits everything, diffusion, reflection, volumetric, transmission, emission, metallic, rough, specular, alpha, everything
the secene is 4.8 million faces, if that's what you mean
great, blender crashed trying to process it
(not a GPU problem lol)
um hello
it was using 11.2 GB for a seco
wow, ok, found the offender
jesus
the subnautica wiki was using like 7GB RAM lol
weird, teh V2 of the bench won't run, but the V 1 runs just fine
tho V2 is way more intense
@oak ospreyV1 of the bench
V2 looks like this, but its not running in full render for some reason
But my really intense bench is my micro film glass statue render
Its a micro glass film render of this 3m polygon sculpture
it is intense
@dense tapirDid you see the new new controlnet model?
I wanna play with it in just a sec
What I just did if you mean the reference function
Needs finer control though as it is removing limbs on me
Yeah, reference is gonna be insane for consistent character creation
theres a new controlnet thing?
yeah, and this is probably the coolest model so far
where can i get it?
Just update control net
oh
a gigantic step forward for character consistency
ill update it in a bit after these gens are done
1.1.162
@oak ospreyThere has been a development
it is not my PSU turning off
Cause I was just chill completely idle, and everything turned off on my breaker
my monitors, PC, router, modem, everything
They are all running off and super old surge protector, so I am wondering if its the protector that's faltering, not the PSU
cause I was doing legit nothing
I have never seen a surge protector act like that due to how a surge protector works UNLESS your house is getting voltage spikes.
For me I get spikes and my UPS will trip but on a suppressor it would just shut down
It keeps chopping off limbs :/
@dense tapir try generating the base image, then I painting the face with a really high denoise
watch an AvE video on how junk surge protectors are
and your breaker actually flipped?
those age out and need replacement
or you are actually just overloading the circuit.
No, not the breaker, but the power in this room did falter
ah
But also all of the power in this room runs off this surge protector
wtf
:D my controlnet updated

I really just use it cause it's an incredibly robust power bar
that sounds like a fire hazard. are you the dad from A Christmas Story?
Oh, total power draw on this thing is like 500 wats max max max
if thats the reference i cant blame it for cutting arms off
Source image is blurry so that went into the dest image
It's rated for 2750 watts.
It was my great grandfather's tool surge protector, he was an electrician and he gave it to me cause it's always been reliable
Well, now don't you look absolutely stupid with that remark after my previous showing with arms.
no
Yeah, you do
y you gotta be rude
Because you said that in a mean derogatory tone when I had just showed it did arms without the source image having them.
no i didnt
Otay, this will only escalate so moving on.
@oak osprey ok, so I have my monitors, PC, and a couple LRD lights running off the surge protector, and on a different outlet, I have the model and router power
Something tripped both cause the little power box for them (has a red/green light for OK power) flipped as well
You got a spike
Ok, turns out power is down in the area next to ours, somebody hit a power pole
I get them even in calm weather so without a UPS I was constantly just going down and losing files
One of our clocks reset in a different room
There ya go
Alright, so THAT TIME it was not my PSU lol
I was doing deadass nothing. Like I had discord open, and nothing else lol
Yeah, it's our general power
All of the neighborhoods around us have no power or had a brown out. Looks like our internet infrastructure is out in one of those areas
probably any connected to that pole as I am sure it destroyed the wires
Yeah, just glad we have power. Would much rather have power than internet lol
Yep
It's hot AF today
Got cold now hot
We live in the middle of an arid high altitude desert, hope everybody else can get their power on to stay cool
All my garden is whacked and dead due this weather.
humidity here is around 80% in this room 78f 60% rh
It's 81 in my room at 25% humidity
not bad
But my room is always way hotter than the rest of the house
Same
Haven't been able to install my AC unit yet
My room at our other house was uninhabitable without an AC window unit
Legit 100+ during the day cause that house was designed so bad
Even needed AC in winter, that's how bad it was
Damn
the dev of ControlNet is toxic 😦
Love his ass
Yeah, it was a second story house on a corner, with a porch underneath, so I had direct outside on 4 of the 6 sides, and then it was at the front of the house so it got sun all day long summer and winter, and then it was also last on the AC chain
It was effectively a greenhouse
Had to buy an AC rated for 2x the square footage in order to keep it habitable
No trees to help with that?
None, it's a dry ass dead ass desert lol
TBH, I love the high desert and can tolerate the lower desert.
You're in Cali, right?
probably better to avoid the CN update until this settles down
No, grew up in South Texas so I know the desert well
Ah, gotcha
uhh why?
it's got bugs in it
Ok, so it appears as tho the power outage is actually not from the pole incident, they didn't actually hit it hard enough to damage it like that
Not too sure why it's down
Very strange
My power company is ran by the city and they wouldn't tell you and they would get to it on Monday as they don't pay overtime
Also seems as tho that crash was an hour ago
So considerably before the surge/dip
huh... reference is weird
Welp, at least I don't need internet to use SD lol
seems like the ControlNet author has bumped the version like 8 times today, and each time he bumps the version, it's an attempt at fixing an issue he should be testing locally, what the hell is happening
A few years ago I was down for 2 weeks in the Summer but my mom wasn't so I had to go crash on her couch due to how hot it was. Lost all my food and 16k residents did. What was the cause? No idea and suck it up. :/
like he's pushing releases out to his test systems 
i have solar
Damn, I didn't get to download the control net update before the internet went down
Sytan, Discord is Internet
I'm on my phone's data lmao
💀
I'll use my phone as a hidden tether like I did when we first moved in
Can't have that here as this is the city of trees and by that I mean most homes are under canopy of 150 foot tall trees.
You can't even see my home using sat photos
Makes it hard for the govt trying to watch me, lol
laughs in infrared
can only get a gist not the actual. Good joke though
IR, or even NIR can't see through foliage. Heat maps can though.
Alright I have internet back on my PC using PDA net hehehe
(undetectable USB internet teather)
thank goodness we have good signal here, I get like 200/50 on data lol
haha it's foilage 🤪
oh how confident you become on PDA net
lmaooo
It seems work now
this is a midjourney image as input (https://www.reddit.com/r/midjourney/comments/11wsoy8/random_people_on_the_streets/)
this is before fix (1.1.162)
this is after fix (1.1.163)...
so this kinda thing is why i'm saying to hold off on updating controlnet
if you want to experiment with the bleeding edge, go nuts. but keep a backup version
OMG, I had this earlier where my image was whiteish like that one
Looked painted like and shouldn't have been.
trying to get it more pixar
underwater steampunk owl in the lava kingdom of the hinterlands, bright colours exploding out of dark crevice
reminds me of that owl image I got a while back that was somehow almost perfect just from txt2img, nothing required after
found it
i've been wanting to make a steampunk lora
Though it has already been done, though not for 2.1 I guess
i recommend using a 1.5 model with CTU to generate some images for fine-tuning 2.1 with
that steampunk owl came out of CTU which added all of the detail
i generated the basic owl in 2.1
so you could actually just fine-tune 2.1 based on CTU'd 2.1 images
yep
Yep
I'm impressed with this skin texturing
12 votes and 24 comments so far on Reddit
man, this thread hurts my brain to read
you know that girl with the popcorn bag? getting that kinda vibe from those legs
from this concept art
hand legs
add words related to movement
I have some
heavy footsteps, clanging armor, swaying shield, jangling chainmail, clattering gauntlets, rustling fabric, grinding armor plates, stomping boots, marching strides, metallic clinks, rattling armor joints, thudding armor plating, creaking leather, jingling spurs, clanking sword scabbard, clomping war boots, resounding clanks, thunderous footsteps, clinking armor scales, clashing metal.
ok, ill see.
maybe also smooth surface to your negative prompt
walk, run, gallop, trot, charge, stride, march, clomp, stomp, step, pace, saunter, meander, amble, wander, traipse, strut, trip, skip, hop
[heavy footsteps] [clanging armor] [swaying shield] [jangling chainmail] [clattering gauntlets] [rustling fabric] [grinding armor plates] [stomping boots] [marching strides] [metallic clinks] [rattling armor joints] [thudding armor plating] [creaking leather] [jingling spurs] [clanking sword scabbard] [clomping war boots] [resounding clanks] [thunderous footsteps] [clinking armor scales] [clashing metal.]
as a Bark prompt for TTS
kinda sorta added all of those and got this
and add film grain, noisy, noise, light leaks, iso 100, f1.4, shallow depth of field, RAW photo for le epic picture
the poop stance
I am using 2.1, might be a bit different to prompt
not just all just a few 
(color photograph, a medieval knight, motion, dirt, medieval village, overcast, moody, cinematic, armor, running, action movie, movie still, action scene, particles, motion blur),
heavy footsteps, clanging armor, swaying shield, jangling chainmail, clattering gauntlets, rustling fabric, grinding armor plates, stomping boots, marching strides, metallic clinks, rattling armor joints,
that's now what I havbe
try replacing action movie with action shot and remove movie still and just do cinematic
because still also is related to still images like paintings of fruit and all
I don't think large prompts work very well with my model, but the action is definitely there
dynamic, velocity, energetic, agile, quickness
negative prompt like resting, still, boring etc
whats your negative prompt look like
low quality, smooth, smooth surface, low res, deformed, messy, warped, distorted, qualityzoov21, structurezoov21, photozoov21
ugly, amputee, fake, CGI, render, rendered, cinema4D, octanerender, drawing, painting, anime, cartoon, monochromatic, oversaturated, undersaturated, disabled, disfigured, deformed, gross proportions, poorly drawn, compression artifacts, low-res, framed, caption, wax, bad anatomy, extra legs, extra arms, out of frame, tiling, bad art"
i use this general prompt
for negative
it just removes a bunch of categories from the model like digital art
meh it's not that important, I will just make it easier to prompt action stuff in v3 of the model once I get around to making that
because your image looks like half photo half cgi
well I started getting ok stuff, but its kinda messy in some parts
yeah looks better
you can get more detail on the armor by pulling keywords related to specific armor types
you can use https://www.pizzagpt.it/ pizza gpt 3.5 (just free gpt version)
name 20 keywords that describe details of a random medival metal armor type separated by comma
helm, gauntlets, breastplate, greaves, sabatons, pauldrons, cuisses, gorget, vambraces, tassets, coat of plates, visor, aventail, gauntlet-polisher, tunic, chain mail, scale mail, leather straps, buckles, rivets
i use it a lot to unfold an abstract concept in the prompt to make it more concrete
and this one has wayyy too much motion blur
I mean I know I can get some decent motion, though this was with v1 of the model a while ago
Dark souls
try this
A stout, muscular, and valiant man, clad in glimmering silver medieval metal armor with articulated gauntlets, greaves, and cuirass, sprinted with relentless vigor and determined dexterity, holding aloft a massive, razor-sharp two-handed sword as he charged fearlessly towards the viewer, his armored boots thundering across a rough, rocky terrain with each stride.

ok
first result

im going to try a really simple prompt now
Medieval Batman
alr im just gonna generate some villages then im done for now
Mick jagger as Luffy
mick jagger luffy?
Dman, our intrernet is still down here
IS THAT LUFFY LMAO
Mick jagger as Luffy
Its brilliant
i love 2.1
2.1
what model, or lora? @stone cipher
Illuminati Diffusion
it is on HF as stablediffusionapi/illuminati-diffusion
thanks as well @oak osprey
Vegeta
This controlnet feature is really buggy
doesn't it need a number of embeds?
Yes but I don't use embeds
at least bit similar, worse it they dont know bud spencer
They are making headway as 2 versions ago compared to now at least it works but is ugly as if too much cfg
you probably need the stablediffusionapi/illuminati-diffusion version of the files, tbh
he baked in the embeds
yep, .5
I can't use it since the damn model is not on 2.1 yet
That is blurry
If you say so
did your VAE change or something
Did you installed the negative embeddings?
I don't use VAEs as I am on 2.1
i use MSE on 2.1
MSE?
the 2nd VAE from StabilityAI
hmmm, all I have are for 1.5
Put this in negative prompt: nrealfixer,nfixer
already there
For Illuminati Diffusion
I think CN fixed one thing and broke yet another
crunchy
this is the exact image i was thinking in my head and i got it at the first attempt
weight 1.0
I sort of think this CN feature is not ready for even beta testing.
0.5
i told you not to update to it yet lmao
It fixed a lot of shit that didn't work before
for 2.1?
the author of controlnet honestly fucking sucks at development
fix
fix
fix
fix
fix
each one bumps the version of the project and is merged to the main branch. he doesn't know you use a test branch and then merge at once, squashing, with GOOD COMMIT MESSAGES
i can't tell what the hell he's fixing, changing, or anything, from this. and when you ask for details, he is very guarded and tells people essentially to fuck off and just use his code "for research" in A1111
Sounds like almost every damn "dev" attached to anything SD.
it's actually pretty rare, in my experience
that's why i am so outspoken about how frustrating he is
Not for me, or a few others I talk to and we are fucking disgusted at them. Most times we just deleted their shit and move on.
he's a pretty important person who has no idea how to do this work properly. or worse, does, and doesn't care or want others to learn from him
I think it is the Auto disease tbh.
was asking for basic assistance understanding his implementation of ultimate upscaler, tells me "just use auto1111"
Sounds like that Dickhead Vlad
even GPT4 has trouble understanding Automatic1111 code and it feels like it's purposely obfuscated to a crazy extent
maybe it's just the garbage known as Gradio
BINGO
but it honestly feels like "this is an exclusive club, and you are not in it"
it's why i'm pretty persistent at attempting to rewrite ultimate upscaler as a standalone library
no reason at all that a website that wants to upscale stuff for users should need ALL OF AUTOMATIC1111
I went to gradio to tell them about symlinks and why is it their shit doesn't use them. Basically, I was told I hurt their feelings and this is a tight community, etc... Not one answer for it.
basically, "works for our needs. PRs welcome"

those projects should be deleted from github when they do that
I get that a ton, or worse devs that never answers any tickets at all regardless who wrote it.
yeah, SAG Pipelines are broken in monumental ways and i traced it down and figured out how/why, and did some A/B testing and provided test grid to the developer. no response in 3 weeks, lmao. i deleted all the SAG Pipeline code from my project
i don't need to maintain many layers of broken garbage, i keep what works without too much futzing by myself or users
I did that a few weeks ago myself. The dev must have died irl
or died inside, after SAG ended up not taking off the way they hoped
That too
it's so fucking obscure
Damn good when it worked though
it was okay but only worked for square images unless you used 1.4 based models
it wasn't as profound of a fix as the new VAE was, and Karras sigmas (or Euler A)
No, I actually had it working in 2.1 then auto updated and that was it goodbye SAG
well i mean, 1.x - i wasn't messing with 2.x at that time
i don't think my 2.x pipeline worked very well at all, that shit really needed Karras to work properly.
so i wouldn't have cared to test it
I can never get these to do anything
i have no idea what those do. it seems like more unnecessarily exposed knobs in Automatic1111.
Control net
wasn't it meant to be a simple tool?
i think a fork of that, which simplifies stuff to like, pre-selected models and a combination of predetermined VAE/scheduler/steps/cfg would just take its place
The main problem with all of this shit is that everything has to hijack Auto to work. No front facing API was created for extensions so it sucks.
programming. Auto was never made for extension, such as Lycoris, so to get it to work they have to hijack a certain point in the stream and it diverts to their code which then does its job to revert the flow back to Auto.
Like hijacking a game to cheat/exploit only the game updated so now you have to find the new hook points.
This is why auto updates it can, and usually does, break at least one extension until they update.
Lycoris devs was so pissed each time Auto updated.
"Oh, Lord, Auto just updated, so I wonder what he broke this time?" That is spoken by many extension devs.
well i mean, hooks are generally how extensions work in a handful of platforms i've worked on
having a stable API is another issue
Then they were shit platforms. You need to give extension writers a front facing scheme so the changes you do does not break extensions.
automatic1111 seems like it was a fun toy that's grown too big?
i guess they could add new APIs but how many devs would even trust them enough to switch 
this is why LUA was made so even common people could make mods without fear of things breaking, or they breaking anything. Plus not everything is exposed.
I agree.
Plus Auto is a 4channer deep into Anime so gives two fucks about anything. Only recently did he finally cave to having a license.
He can disappear months on end too so now we have forks to go to.
robo groot?
How to make image here
Currently, there is no public bot on the server that generates images. There is an experimental bot available for early server members & Stable Society. You can obtain the Stable Society role by winning our weekly events #⭐|pow-info #1087493421209485393! However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
@dense tapirGeneral question for you
I have been having some issues with my GPU
wanna see if maybe you have an idea
what exactly is the desktop window manager?
it keeps randomly slamming my GPU for no reason, and it completely freezes my SD gens until it stops
did you completely remove the nvidia drivers and install the latest?
it switches from GPU 0 (IGPU) to GPU 1, and it uses 100%
you have two gpus?
Pretty sick waiting my 2GB VRAM GPU to upscale 960x1520 for 20 minutes
and my gens get paused for like 2 minutes while it does that
( i copy from someone else. )
CPU iGPU and DGPU 3080
can you even run integrated GPU on SD?
no, but you can run it on CPU
if there is any way to disable igpu in bios I would
well, maybe you can on AMD
I leverage my iGPU a lot to save energy and performance
like all of my youtube videos and all of discord is my iGPU so there is 0 hit to my SD gens or rendering
then I suggest a clean uninstall (using DDU and safemode) of nvidia drivers then install the latest
I am just trying to figure out what this bug is that keeps causing desktop windows manager to just thrash my GPU
DDU is your last hope so don't rely on anything else to take care of uninstalling
driver issues and shit
see!
its doing it again
only when I do high res fix, and my videos all get glitchy and stuff
like my youtube video dropped to like 3 fps
wait what
yeah, as I said until you DDU in safemode so it really cleans it all out it is akin to upgrading from a HDD/SSD to a nvme as the boot drive. good luck.
and its also showing 9.9GB VRAM usage
most people just start fresh like that as it is easier to do
1536x1536 shouldn't max out my VRAM
hm...
alright, I will try DDU
I have done it before, I just remember it being a pain in the ass
a pita? Easiest thing.
hold down shift key while clicking restart then tell windows you wish to go into safe mode. They took all the short cut ways away from us years ago. Once in safemode boot it and easy peasy
the whole disconnect internet thing and all of the backups and stuff
I said in safemode not normal mode it is shit doing it like that
it can't delete the bad things if you are in normal mode
at least the generation came out really good
https://github.com/Coyote-A/ultimate-upscale-for-automatic1111/issues/69
@dense tapir seems like ultimate upscaler also has the tiling seams issue
LOL
the example images for ultimate upscaler
original vs upscaled
that shit hallucinated doors and windows all through the wall
little ghost
I got beer
seeing how shitty the ultimate upscaler results are once i look closely, i don't really feel the need to get it working anymore.
Damn. Well, I can't use it being solely on 2.1 but that still sucks.
yeah mate. i'm disappointed because it seems like such a neat idea
It worked so well working in latent space
it's working in pixel space to do this tbh
really? shit
it keeps converting everything to pixel tiles and then upscales them and then converts them back to latent space to fix it with inpainting
Don't we have anything that will take a pic, shove its ass into latent space then work on it in there to finally bring it back to pixel space?
it goes, image -> tiles -> each tile to latent space -> upscale tile, denoise tile -> all tiles are stitched back together in pixel space -> the image is possibly tiled again in an offset grid this time -> those tiles are inpainted to fix seams -> back to pixel space, stitched together? i have no fuckin idea at that point
and yeah that's the x2 latent upscaler
it's terrible iirc
terrible due to implementation not because it is a horrible idea.
the problem with latent space is it will do, at min, small changes to the image.
Im debating starting to work on a steampunk lora, but I might just play some mordhau
I wouldn't have time to make the lora

That wouldn't look too good, it's a top-down pixellated game, so it would kinda be repetitive
now, an industrial/factory like lora could be cool
in conjunction with a steampunk lora would look even cooler
yep, fuck styling in 2.1
It burnt me out it was so bad
tried 1.5 with the same dataset and instant perfection. I have not touched training since.
Factorio LoRA
Emad knows his shit sucks too and knows SDXL for training will suck too unless they changed their clip or give us proper tools to train with made for 2.1
the CLIP is the best part of 2.1 🙂
not for training though it fights you
Awesome!!! Love this!!! 🙌🏽🙌🏽🥰
love what lmao
The TE is the CLIP and it fights any changes. uunet is fine
the data. you get fucked up 0.4-0.7 losses
fine-tuning the diffusion model rather than the text
as i understand it, loss refers to the input-output no longer being 1:1 through the autoencoder.
I think for DD I fully trained the text encoder on 100% of the steps
no one has managed a lora/locon with less than a 0.29ish loss in 2.1. Same dataset in 1.5 is 0.0x.
and it wasn't destructive? what your learning rate?
I can tell you at 0.5, or more, throw the lora/locon in the trash
Im not sure... might have been the default
I tried many learning rates though
slower is better when you're training the whole model
It's whatever tyhe default is in the dreambooth extension I think
Actually I could check hold on
slow learning on an unfrozen model helps prevent destructive changes from causing catastrophic loss
I have a friend who just did his first full on 4k images, etc... finetuning on 2.1. He was not happy at all, but I warned him.
Here was the whole config
gradient checkpointing, good job
that monumentally improves results when training 2.1
Which DB is that?
pretty sure it was on by default
the dreambooth extension
for auto1111 but used on vladmandic
people in here i see turning it off because "it saves memory, and i have enough memory" not realising it helps improve coherence too
I always used checkpointing
https://huggingface.co/docs/diffusers/training/dreambooth < this is what i use
Finetuning with prior-preserving loss.
That is like Joe Penna. Good stuff but takes too much mem
well a month ago they apparently fixed DB so that the vanilla one works with the same settings for 1.5 and 2.1, some-fucking-how
Besides, I don't get how people can sit there through 300k+ tags manually fixing them.
If it isn't automated I want no part of it
easydream2 people working on one FT 300k+ images, tags/captions manually fixing them for 3-4 months
I never had success with TLB, or at least anything to be proud of
oh, @dense tapir , I don't think you were around when I asked this the first time. If I were to make a v3 for DD, what would you like to see? So far I am thinking about adding more amateur photography, just to balance it out so it's not so "perfect" or "commercial". Also, add more motion shots since it seems to struggle with action scenes sometimes. And according to @oak osprey I need to add better caustics like SDXL. Do you have any suggestions?
I also need to add monkeys
national geographic archives in general, tbh
cats, of course, more animals in general, and flash photography.
can you set up a cropping pipeline that pulls a "best" 768x768 from an image?
I actually did try finding those, coudn't get any I don't think
eg use a feature extractor to determine where someone is, and then, make a crop
with them centred
there are shit tons of varieties of house cats
I mean... BIRME kinda does it? It wasn't great most of the time but it's really easy to edit the crops in batches
for instance Himalayan it sucks at.
@hasty nova https://magzdb.org/j/1806 < all Nat Geo
shit, scrape that
Woah, thank you!
i gues those are just the covers 😦
i will DM you
I doubt Nat Geo is going to offer up their hundered plus years of photos
No definitely not
be freaking sweet if so
I mean just taking their entire dataset and captions and stuff would be awesome to train on
omg, yes
apparently i had to add you as a friend first
just describe a nat geo image and BAM
great photography
it would even do lots of diverse people really well
I think I accepted it
Hi
I am buying a new 970evo nvme for my boot drive and taking my current 500gb for auto.
getting SD onto a nvme, even if gen 2 x4 will be so much nicer.
I noticed all my models, and stuff, are almost 500 gigs so hmph.
meh I am just gonna play some mordhau for a bit, I don't want to make the lora tonight. I did a lot today and i need to get up early to make breakfast for my mom since it's mother's day tomorrow
thanks for reminder.
Aye
and I think we are gonna go see GOTG 3 tomorrow, can someone remind me what happened in movie 1 (I remember 2 well enough I think)
LOL, 1 was the best. 2 was good, but 3 I dunno.
tbh, none of the marvel universe stuff is worth watching more than once.
Im not a huge MCU person so I like the movies/characters that are a bit more standalone. I dodn't really like all the infinity war/end game stuff because I didn't understand a lot of what was going on
I understood it alright, I just like more to my keepsake movies. Good movies/films, but more than once is not for me
I can't even think of a movie I have seen a bunch in my lifetime.
I haven't seen anything more than like 3 times
I used to own over a 1k 10 years ago then the wife passed and that was our thing together so not much into that now.
Sad is the films she had just ordered arrived after she passed.

I never opened them.
what's a 1k?
like 1000 movies?
Ah ok, I was just confused by the "a" like it was a singular object, mb
that was a typo
yep no worries
we were big time into films, especially head trippy ones
if it made you think we loved it and most time audience scores bad and we loved them
There are a few shows I watch, I am not big into many movies
since wokeness arrived I don't like any non indie stuff.
National Gallery of Art
oo cool ill have to check that out
that's cool
holy cow, wikicommons is fucking cool
yep
my nvme and heatsink for it was ordered. 970evo plus
I priced my wd black and shit. wth WD?
Jensen pricing
Waiting on my case to come down to pre MIR rate.
then I am ready for the card
https://commons.wikimedia.org/wiki/Category:Seney_National_Wildlife_Refuge_photography_competition < i can make a zip of this 😄
those are good, might sift through them for some nice training images
though, i think adding photorealism to DD made it worse at it's original purpose, which was digital art. maybe I need to fix the dataset and add more digital art (though it already mostly is)
can use them as regularization images
stuff that keeps it from nuking the stuff SD already knows
it might generate them before training starts
because you can use images from SD as regularization images
I don't think it does for m









