fallow marlin
Hello, guys I am working with SD upscaler "stabilityai/stable-diffusion-x4-upscaler" and find some unexpected case.
I am trying to upscale different types of images and when I upscale images with large amount of black color model produces some random artifacts, maybe some of you know how to fix it?
I am also providing example of it 1st is basic image and 2nd upscaled image totaly filled with artifacts.
celest sigil
steep sleet
Flux.Dev.FP8 - 20 iterations - 1024x1024
languid pebble
nocturne oak
fleet goblet
celest sigil
languid pebble
cobalt isle
nocturne oak
nocturne oak
nimble mason
nocturne oak
latent kraken
I made this diagram with flux to help my ai's in case they start hallucinating...
nimble mason
warped night
Flux
latent kraken
a bird in the hand is worth two in the bush but 2 cat birds? ...priceless...
tiny quail
nimble mason
pallid flax
languid pebble
Good morning coffee
leaden lava
it might be a dumb question, but can i download one of these images and extract the workflow? I'm searching for an ultra realistic model for professional photos that is not too "smooth" and some of these are awesome
languid pebble
it might be a dumb question, but can i download one of these images and extract ...
You can open them in your browser, save them and sometimes the workflow might be included.
leaden lava
You can open them in your browser, save them and sometimes the workflow might be...
I Tried but the images are jpeg so comfy does not detect the workflow, should i change them to png?
languid pebble
Click on the picture in discord, click the link "Open in Browser", save it ... some people don't include the workflow ... so it doesn't always work.
I am one of the bad ones 😄
clever oar
unique condor
accidentally ran the prompt "abnormal" because i had text already selected when i typed and this is what it spit out lol... (flux.dev.nf4)
apparently, this image is the definition of abnormal
vagrant dust
I am one of the bad ones 😄
Still not quite sure if not embedding the workflow is better then embedding one of clownsharks with all his custom sigma nodes... 🙂
languid pebble
Clownshark is our Sheldon Cooper 😄
fleet moth
wispy nest
this was flux image for the control net input, kolors image for the IP adapter and then SDXL for the generation
celest sigil
frank drift
Gorgous! May I ask your prompt?
wispy nest
you can get prompts from images using stuff like Florence 2 by the way
it makes a caption
the difference is that they suddenly got good enough to make the image from
frank drift
you can get prompts from images using stuff like Florence 2 by the way
I've not heard of Florence 2! But you're right! Thanks for the tip. 🙂
vagrant dust
I've not heard of Florence 2! But you're right! Thanks for the tip. 🙂
Florence2 won't help you. Getting a detailed caption or regional object definitions will not get near to the same result. The "more detailed caption" of the angle image is:
"The image is a digital illustration of an angel with long white wings. The angel is in mid-air, with her wings spread wide and her body slightly tilted to the side. She is wearing a long, flowing blue dress with a sweetheart neckline and a corset-style bodice. Her hair is flowing in the wind and she has a peaceful expression on her face. The background is a starry night sky with pink and purple hues, and there are planets and stars visible in the distance. The overall mood of the image is serene and dreamlike."
Neither FLux.Dev, PlaygroundAI, Pixart-Sigma or a mix of them and SDXL will nearly get to the result of the image from GalaxyTimeMachine. The Amount of settings, parameters, models, seed, cfg, loras, ... is limitless. So if you just want a flying angle in front of a night sky, florence2 can describe the image. If you want to learn how the workflow was done to achive such a piece of art you need to hope that the usual creatives share a part of their intellectual work with you.
lavish garden
Prompt: https://civitai.com/images/24127520
"Your Reality" - I1 (oc)
candid surge
great prompt showcase
lavish garden
💜
languid pebble
You 🙂
languid pebble
You are up 🙂
clever oar
this?😃
languid pebble
Cool
celest sigil
Gorgous! May I ask your prompt?
Thanks! These were both made using Flux. I think the angel one was using a fantasy style template with the prompt
abstract shore
flowing magic light, blue purple teal red, fantasy warrior in epic battle
The second one was:
render, human (advanced, hyper, detailed) holding a sword (twin blade) looking down his fallen enemies surreal hyperrealistic (SSS)(stable ray tracing), full wide shot (subject in focus), vast war arena
coarse echo
flux is kinda crazy ngl
candid surge
thick tiger
stone pulsar
style cyber punk
frank drift
Thanks! These were both made using Flux. I think the angel one was using a fanta...
They look tremendous! Thank you for sharing!
hearty violet
@royal monolith I generate all controlnets with each preprocessor manually, is possible to genereate it automatically?
royal monolith
<@407561236339752981> I generate all controlnets with each preprocessor manually...
probably, but i have no idea how to do that without some sort of scripting. you might as Matteo on the L2 discord
hearty violet
probably, but i have no idea how to do that without some sort of scripting. you ...
I will take a look on the scripts... what you means with Matteo and L2 Discord? I never hear about it
royal monolith
I will take a look on the scripts... what you means with Matteo and L2 Discord? ...
i'll DM you the link. he writes a lot of stuff for Comfy
hearty violet
i'll DM you the link. he writes a lot of stuff for Comfy
thanks ^^
hearty violet
execution inversion
what is this? 👂
wispy nest
imagine writing scripts using logic like for loops
using comfy nodes
its gonna be the biggest change to comfy in a long time
hearty violet
will be great 🙂
wispy nest
it will be an improvement over current comfy yeah
So flux has a term in their licence that prohibits the training of AI models on flux output.
users generate millions of images and spread them on the web
web scraper aggregate data
datascientist go to scraper and get image data packet
datascientist train AI model.
isn't this totally unenforcable and also unreasonable?
the NY Times lawsuit against Open AI is one of the big ones that will decide
imagine having your model taken down because flux devs claim its trained on flux model outputs
you cant know an image is generated by a certain model if it has been recompressed a few times over the net
yeah the legality is rough
gonna be an interesting time for law
half verge
tree
spring mist
azure stone
gril
spring mist
young pasture
Forest
steep sleet
Flux dev.bnb.nf4
steep sleet
Flux.dev.bnb.nf4
steep sleet
NF4@Flux
steep sleet
steep sleet
NF4@Flux
coarse echo
are these with flux, also are u using any loras for it? mine come out too realistic
celest sigil
are these with flux, also are u using any loras for it? mine come out too realis...
Yes, all Flux and no Loras.
Use SDXL style nodes.
steep sleet
NF4@Flux
dull flame
flux seem incredible with fabrics
merry ironBOT
cinder arch
frank drift
Guess where they're from
languid pebble
vagrant dust
Some resemblance with your home town. Did you trained a Lora.
languid pebble
Not really 🙂 But it fits to a canal here 🙂
Looks like flux doesn't know "hidden picture puzzle"
Or maybe my translation is wrong. In German we have the word: "Wimmelbild"
vagrant dust
Yeah sadly some concept where not trained for the model. But as far as I know the community is already adding some concept. Guess po*n will be first.
languid pebble
Yeah sadly some concept where not trained for the model. But as far as I know th...
Sad but you could be right ...
vagrant dust
Sad but you could be right ...
Hidden Object Image is the translation at least that is what google says.
languid pebble
I'll try that one again, thanks
vagrant dust
Maybe „in style of where is Waldo“ 🙂
vagrant dust
Looks like Leo.org isn't uptodate any longer 🙂
It seems someone created a Lora for SDXL
languid pebble
Cool! I'll have a look for that one!
steep sleet
NF4@Flux
candid surge
fleet goblet
candid surge
coarse echo
celest sigil
I had to find it after you moved it from #🧣｜comfy-ui 🤣 !
coarse echo
i posted it in the wrong channel 🤣
celest sigil
Nice image though
coarse echo
thick tiger
nocturne oak
coarse echo
languid pebble
mystic marlin
I can't wait for winter to come back, I'm sick of the heat.
Made with Flux.
boreal hare
Give froggy a hug
thick tiger
thick tiger
languid pebble
Good morning coffee! 🙂
brazen sand
candid surge
candid surge
steep sleet
NF4@Flux
celest sigil
wispy nest
I should make more pink images I forgot about this colour
candid surge
"Hi Barie!" says realistic cgi fractal art
im happy with the primary colors red green blue, though in a bit dark Bulgarian style
steep sleet
NF4@Flux
languid pebble
Remember these guys? 😄
languid pebble
Beaves and Butthead ^^
coral veldt
Thought this came out good
languid pebble
Strange neck ...
coral veldt
I’m still learning how to use it.
languid pebble
For that it's OK ... can't really tell you what the problem is ...
Just continue ... I'm learning 2.5 years now and it's still fun!
Stable Diffusion might be the wrong name for it. I bet you are talking about your UserInterface.
You want to reinstall?
If you want to delete the old U.I. you will want to keep output and models folder ... the other stuff can be reinstalled
Not that much ...
Sometimes the model folder has a subdirectory "Checkpoints" and sometimes "Stable-Diffusion"
I would keep those, yes
I don't really know that one
But not a lot inside there on my computer
lavish garden
Prompt: https://civitai.com/images/24467451
"Respite" - Kabbu (Bug Fable)
coral veldt
Okay I used this model https://civitai.com/models/139391/jin and used a different website/app and this is what it came up with. I suck at trying to describe what I’m thinking but the face came out good
languid pebble
Okay I used this model https://civitai.com/models/139391/jin and used a differen...
Most of us use local generation ...
And what you call Model is just a Lora ... kinda AddOn but needs a fitting model ....
coral veldt
And what you call Model is just a Lora ... kinda AddOn but needs a fitting model...
Ah ok
languid pebble
Try a photorealistic 1.5 Model with it ....
coral veldt
Try a photorealistic 1.5 Model with it ....
I’ll try that thanks
languid pebble
I’ll try that thanks
Best luck!
coral veldt
Try a photorealistic 1.5 Model with it ....
That is a checkpoint right?
languid pebble
Yes ....
languid pebble
SD 1.5 is a bit outdated but many still love it cause of so many LoRas and Finetunes. If you don't need higher resolutions you can still use it
crimson surge
coral veldt
SD 1.5 is a bit outdated but many still love it cause of so many LoRas and Finet...
Ah ok didn’t know
languid pebble
Ah ok didn’t know
SD2 changed the way of prompting, that'S why many didn't like it. SDXL is pretty cool and followerd SD2. SDXL is good for 1024x1024. SD3 isn't really finished yet ...
gray ore
How can I get results like this?
coral veldt
yeah im wondering what to use for that for mine lol im trying to figure out how to word it making a xmen theme lol
languid pebble
yeah im wondering what to use for that for mine lol im trying to figure out how ...
It will take time and become better!
languid pebble
Clownshark made into an Latent Vision Youtube Video .... congrats!
gray ore
or if someone could explain the steps to get somthing close
coral veldt
created jin as a zombie with pixai
coarse echo
lookin like the next big chinese mobile game 🤣
warped night
Is there anything made after liveportrait update for lip sync
fringe fog
Generate anime animation expression package
royal monolith
Generate anime animation expression package
read #artisan-faq
languid pebble
Good morning coffee!
celest sigil
sterile kiln
guys, can someone tell me if its possible to train a LoRA for Flux?
And with what is this possible? (Kohya_SS?), and what are the perfs required to create one?
Best regards
flat sand
yes
I use SimpleTuner for it
I think kohya_ss should support it now, too, but I haven't tested it yet
Ideally you have 24gb of vram, but I'm pretty sure you can also train with 16gb vram
lower is probably not possible, but I don't know exactly
sterile kiln
i have a 3080ti (12 GB) 😦
flat sand
I mean, there is probably some crazy setting - but it will be horribly slow
I would rather rent a gpu then
sterile kiln
I bought my 3080ti less than a month ago xD
flat sand
machine learning is vram-hungry 🤷♂️
sterile kiln
machine learning is vram-hungry 🤷♂️
Yeah, but 3090 and 4090 are out of my reach. And their monstrous consumption just as much, the 3080ti is already a big whim financially speaking x)
nimble mason
if you own a car, ebay some of the non-drivers doors and passenger seats and stereo and hubcaps
might be able to raise enough for a used 3090
maybe even two
sterile kiln
if you own a car, ebay some of the non-drivers doors and passenger seats and ste...
Not when you drive a 30-year-old Honda Concerto.
x)
languid pebble
😄
nimble mason
so you could kep the drivers door and still raise $600
lol
bet if you sell the seats and just drive around in a lawn chair you'd be golden
languid pebble
In one week I'll be at the GamesCon ... hope to get some news there ^^
5090 I'm coming 😄
nimble mason
kind hull
Crie um monstro de roçadeira
sterile kiln
#200 for an old door
Yes, after all, during heavy rains (called “episodes cevennols” here), they won't be very useful xD
nimble mason
hey, opportunity to sell your clothes too, then you can just wear a bathing suit and pretend you're swimming in the ocean
kind hull
Create a brush cutter monster
nimble mason
the wheels are ready to fall off my car too
i could get another one, but then i realize i could buy like 15 4090s for that
languid pebble
Create a brush cutter monster
You can't create here ... look here: https://discord.com/channels/1002292111942635562/1237461679286128730
nimble mason
which i would much prefer to a new-ish car
but, that would be stupid to buy, so i just do nothing
sterile kiln
And for a new PSU 1000 Watt +, I could sell my fridge, maybe my bed too.
nimble mason
yes lol
languid pebble
I'm buying souls ... ^^
sterile kiln
I'm buying souls ... ^^
Jeez, what a idea, It's true that instead of giving it to the good Lord, I could get an RTX that's worth a bit more. After all, a good GPU against eternity, why not?
languid pebble
sterile kiln
Trade done !
everything went very well, I don't see what the problem is.
Well, the salesman is a bit weird.
languid pebble
languid pebble
Cheaters 😄
sterile kiln
nimble mason
coral veldt
Can’t seem to get the correct haircut and style at all it keeps giving me this look can anyone help me if I send a photo of the hairstyle in a dm I want to use and how to word it in the prompt to get it correct?
marble slate
Its all Setup for the best fight
languid pebble
vagrant dust
Can’t seem to get the correct haircut and style at all it keeps giving me this l...
i got you. here is the hair cut he needed.
coral veldt
i got you. here is the hair cut he needed.
not the hair i wanted lol ill send the link of the hairstyle
vagrant dust
Just saw it in the prompting help channel
coral veldt
yeah thats the hair i was trying to use in the prompt it keeps giving me the same one over and over in the one i just posted on here
vagrant dust
yeah thats the hair i was trying to use in the prompt it keeps giving me the sam...
this is a pixie haircut
coral veldt
ill try that thanks
i have to wait to get more credits lol
nvm still didnt work still giving me the same length not the length shown,,
languid pebble
clever oar
languid pebble
languid pebble
my gpu is a blackhole
Did it work?
languid pebble
It's all always beta 😄
errant zenith
😭 what is it doing!? I only have 32 gb ram and 6 gb vram
languid pebble
6GB is dang low
languid pebble
Will depend on the Model, Resolution and Extensions you use ... but happy if it works for you!
nocturne oak
errant zenith
actually it's 95 seconds now
languid pebble
I'd think Flux is to big ... but call me stupid!
errant zenith
I'd think Flux is to big ... but call me stupid!
😔 it's totally too big
languid pebble
Need a 5090 🙂
wispy nest
wispy nest
i have 4tb ssd, but 32gb ram, flux gets read from ssd each time since i cant keep it all in ram. need more
errant zenith
damg
wispy nest
the ai stuff is getting crazy. to run the latest image models you need 2k+ setup, while for llm ai models may need 15k+ setup
errant zenith
There are some smaller Flux thingies ... have you used those?
🤔 only schnell from the official repository
languid pebble
5k setup here ....
wispy nest
😃 I'd have 1038gb
id give fooocus a try, its well optimized for small vram
wispy nest
5k setup here ....
what gpu u got
errant zenith
id give fooocus a try, its well optimized for small vram
🤔 what's focus
wispy nest
🤔 what's focus
like a1111
https://github.com/lllyasviel/Fooocus
languid pebble
🤔 only schnell from the official repository
I'm not 100% into that topic, but you can use a reduced LLM and I think there's a smaller version of Flux, too
languid pebble
what gpu u got
4090
languid pebble
nice
I don't spend a lot of money ... but I don't save money with the tools I work with ....
errant zenith
shouldn't vram get less expensive as demand goes up?
languid pebble
shouldn't vram get less expensive as demand goes up?
NVIDIA let us pay to much for VRAM .....
wispy nest
shouldn't vram get less expensive as demand goes up?
Not if the market has a monopoly
languid pebble
AMD has way cheaper VRAM but doesn't really work perfect with A.I.
errant zenith
NVIDIA let us pay to much for VRAM .....
they be making of creators jealous
didn't intel start making gpus
wispy nest
they did , lots of companies make gpus, but only nvidia has gpus that shine on ai related workloads
cause they have monopoly on the cudnn software which makes it possible.
languid pebble
Bad luck for us ....
marsh drift
hey, I don't see a section regarding flux, so I will write here - what version of flux model should I use on GF 4080?
languid pebble
Try dev ... if it's to slow go to schnell
vagrant dust
hey, I don't see a section regarding flux, so I will write here - what version o...
Depending on the ui you use, you could try the new gguf versions.
marsh drift
using forge, but thanks for help guys, will try dev version then
languid pebble
It's just my thing ... try the best first and only go down if it doesn't work for you ....
And I have unlimited download and a 4090 ... that's why I have a small problem giving advice
But Lord of the Weed said there's a gguf Version ....
marsh drift
well I dont really know what is "gguf version"
but will check that also, i hope it will work with forge ui
languid pebble
I'd think ComfyUI ... should become best stuff for first trials....
I think they are at least a week ahead ...
marsh drift
one more question in this case, the more advanced version of flux the more resolution it can generate? or version does not matter?
languid pebble
I don't think they can produce higher resolutions ...
An upcoming model maybe able ...
Now it's more about finetuning ...
vagrant dust
well I dont really know what is "gguf version"
Sorry, it is a concept of reducing the size of large language models to increase the speed or eneble large models to run on consumer hardwares. For most of the image generation models the quality loss was high, so other concept where used.
But it seems that the flux model does not suffer much by the quantization so people created 4bit, 5bit 8bit models.
And Comfyui got nodes that enable the usage of these models without any trouble. BUT for now for example the canny controlnet or the flux loras does not work.
errant zenith
about right
yea my pc got 8 gb ram (task manager apparently doesn't show all used ram)
vagrant dust
one more question in this case, the more advanced version of flux the more resol...
I would focus on upscaling, tile upscaling etc. to achive higher resolutions.
Needs less Hardware and achieves good results
The 4q_0 Version of the gguf model is about 6.8 Gbyte large. So it should work good on Hardware with 8Gbyte of VRAM.
errant zenith
I would focus on upscaling, tile upscaling etc. to achive higher resolutions.
Someone said not to use xformers with flux, is that true?
Well.. if you know
vagrant dust
Someone said not to use xformers with flux, is that true?
Did not try it. Can't say if it works or not
errant zenith
I see
royal monolith
Someone said not to use xformers with flux, is that true?
it's an older package and probably doesn't work well with flux - pretty sure it was @limpid lichen that mentioned something about it
unique condor
So this is interesting.
You can actually just say f it and run the q8 versions of the ggufs instead for maybe 20-25% slower generations. It's what I'm doing with my 8gb gpu
Still faster than fp8 by a landslide
And higher quality
Just make sure to use his node and update it
fleet goblet
Still faster than fp8 by a landslide
Not faster for me.
These GGUFS are slower for me than just fp8 quantized variants
I assume this is because I'm using the load diffusion model node, which loads the weights as fp8 e4m3fn
unique condor
These GGUFS are slower for me than just fp8 quantized variants
make sure the text encoders aren't stay in vram, what gpu do you have?
fleet goblet
I'm an arc user. Also, the clip encoders are unloaded once inference begins
unique condor
oh that's probably why
unique condor
yeah that shouldnt be happening, the gpu probably has weak support for one of the precision cast types then
fleet goblet
Actually, I think I know why
this gguf node seems to still be upcasting
for me that is
to fp16
unique condor
yeah like i just said
unique condor
there might also be some things city96 can do to mitigate it though. 99% of addon creators only bother coding for cuda
unique condor
you'd need to run hundreds of double blind a/b tests between it and fp8 to actually find out. seeds aren't going to match up, so you can't compare like this
fleet goblet
That's very true, yes.
unique condor
i've seen a ton of garbage fp8 gens and a ton of amazing q4 gens
fleet goblet
Why do the proportions on his face remind me of pixar?
YAAAAAAAAHO
lol
@unique condor I'm currently also using --lowvram to ensure that the clip encoders are cpu inferenced in Comfy
Which is like 2-3 seconds slower for me
Now we just need T5XXL 4_0 GGUF.
unique condor
<@204182589303554048> I'm currently also using --lowvram to ensure that the clip...
i think the only way you can actually do that is to use the force clip node from the extra models pack and set it to cpu. and no, with llms, pretty much absolutely nobody uses anything beyond q8
oh you changed it
fleet goblet
Yes.
unique condor
yeah, q4_k_m is the best jack of all trades
fleet goblet
Also, I wouldn't use 8_0 with LLMS
ye
4_k_m all the way
At least until we get better matrix quants that preform better
4_k_m it is
unique condor
well personally, i always just pick a whatever quant can fit in my vram+room for context
depending on the model, it might be a q6 or something
fleet goblet
6_K be like
unique condor
usually, at most, i only ever need a context size of 4096, even if some of the models go up to 128k now. so that would have been wasted vram room can just be used to up the quant level
unique condor
6_K be like
one of these is fp8_e4m3fn and the other is q4_1, i'll let you decide which is which. both have identical settings, seed, etc, just swapped the models
"A druid with a glowing face stands in a ruined city, surrounded by eerie, glowing plants that cast an otherworldly light upon his face. Behind him, the skeletal remains of once-great buildings stretch towards the night sky, their forms silhouetted against the faint glow of distant stars."
(kinda dumb prompt now that i'm looking at it, just used llama3 to expand some basic tags)
unique condor
see what i mean about a/b tests when you aren't aware ahead of time?
fleet goblet
Yep.
unique condor
but yeah, left is fp8 and it actually mangled it more than i thought it would
the weird glowing dildoids on the ground
fleet goblet
frank drift
Pikachu!
fleet goblet
I'm so amazed how little the quantization really affected the quality of outputs...
They really aren't affected in the same way an LLM is.
anomalous sponge detected
fleet goblet
errant zenith
I'm confused, is fp8 better or worse than 4_0, what does the numbers even mean?
languid pebble
looks awful
errant zenith
😭 what's the 5001 input
I'm looking for a queue extension
almost questioning why stable diffusion doesn't have that-
languid pebble
It's not stable diffusion you are talking about. It's your UserInterface ...
ComfyUI has a queue
sterile kiln
I'm so amazed how little the quantization really affected the quality of outputs...
Quantization?
candid surge
errant zenith
ComfyUI has a queue
yes but I don't think I can run flux on 6gb vram on ComfyUI
languid pebble
yes but I don't think I can run flux on 6gb vram on ComfyUI
It handles everything pretty well ... should be worth a try
errant zenith
well- saw someone else who said that 8gb vram didn't make it run properly but I think it would be outside my comfort zone to get comfy ui working
vagrant dust
Well the so far smallest Version is the 4q_0 Flux gguf file with a size of 6.7 Gbyte. So with 6gb it will definitly swap between RAM and VRAM so it will take longer. As Flux.Schnell only needs 2-4 iterations it should still be kind of acceptable
sterile kiln
I don't understand guys, what are you talking about q_4m, t5 and other gguf files about flux??? Sounds like LLM file, what the point with flux?
wispy nest
there are some similarities with LLMs yeah
Flux is a DiT architecture rather than a Unet
and this can make it more suitable for certain methods like this
the point is mostly to reduce VRAM requirements
and also to make it faster
sterile kiln
But, t5xxl is not clip file?
sterile kiln
T5 came from text LLM world
but, with comfy, i have to set it in clip folder to get Flux working.
And another question, wiche is better?
fp8 or fp16?
steep sleet
well- saw someone else who said that 8gb vram didn't make it run properly but I ...
My 8Gb VRAM gpu runs nf4 an image every 110 seconds at 20 steps
errant zenith
nice
candid surge
Well the so far smallest Version is the 4q_0 Flux gguf file with a size of 6.7 G...
thnx for the reference 4q_0 Flux gguf, gonna try this!
wispy nest
but, with comfy, i have to set it in clip folder to get Flux working.
And anothe...
I haven't switched to flux yet so I don't rly want to advise on flux workflows
boo
hi
lol
wow nice
duel RTX 4090s
ah I see yeah
ebon ingot
$$$
wispy nest
do you split inference between them ?
wispy nest
$$$
years of working in the industry unfortunatly
on Vast ai you can get 4090 for about $0.20 per hour
and the cheapest GPUs are $0.035/hr
3 cents
I run 3 webui at the same time across 3 monitors
ebon ingot
years of working in the industry unfortunatly
do you use AI for profressional work / make $$$ ?
wispy nest
I would never use GPU sharing . way to many exploits hidden in them
Yes i do im the Dev of Dream Diffusion gen models
wispy nest
lol just dropped the new hyper trained Flux checkpoint 2 days ago
wispy nest
I would never use GPU sharing . way to many exploits hidden in them
I agree but I think if you stick to verified data centre GPUs the risk is lower
its not the verified centre where the issues are. But its not my place to say here. I only just joied this server lol
have you guys used the flux checkpoints yet?
there's always Azure/AWS etc, you don't pay enormously more there
yeah I used Flux a bit its really nice
wispy nest
there's always Azure/AWS etc, you don't pay enormously more there
yeah I used Fl...
im loving it right now
errant zenith
😭 do you think someone could make a chart of all the llm terms or can we stop inventing new ones
ebon ingot
whats LLM ?
errant zenith
wispy nest
Large Lan Models
errant zenith
actually.. this isn't.. mang
errant zenith
at this point it's a language model too XD
wispy nest
😆
languid pebble
wispy nest
he has two red ties
double bubble
languid pebble
I have noticed ... since Flux came out all the good pictures at CivitAI have been produced with Flux 😄
wispy nest
lol yeah
Ive been working with Flux to get some awesome transparent renders today
on this server too
languid pebble
Using it 90% of the time ....
wispy nest
Only issue i didnt like was the Flux 1 and Flux Dev checkpoints wernt full trained . So I retrained the dev version and now it does things that the base models cant
errant zenith
Ive been working with Flux to get some awesome transparent renders today
you could fool me by saying it's real autodesk renders 😭
errant zenith
I think flux is more flexable than sdxl.. but I am not sure
wispy nest
I'm still on SDXL to get looks like this
What SDxl model are you using bro?
that was Helloworld
its got the comfy JSON embedded so if you drag it into comfy you can see the workflow
Try this sdxl model the base is turbo but 100% trained model https://tensor.art/models/751438781070140461/SD3-LIKENESS-Dream-Diffusion-By-Dice-v1
released 2 months before SD3 was dropped
all my renders with prompts are here https://www.shakker.ai/modelinfo/0ab32c16310a4b6bb52a065e09cb5017?from=personal_page
I don't think this license is for me. It looks cool though
languid pebble
Another 15 gig to download (Flux)
wispy nest
thats Stabilty ai licence bud
best 15GB youll ever use lol
I meant this bit, I don't really know Tensor Art so I don't really know what their terms are
Thats the rules on Flux atm
Dev base is a bit confussing to be fair as shows as a non commercial license
untill thats made clearer not much we can do about that
its a confusing site cos every model seems to have a different combination of perms
for example this one is an SDXL checkpoint
in this case they prohibited selling generated images 🤔
yes that could be something like a trained SD3 model
or a turbo model
for example
languid pebble
I wonder how they want to find out the used model ...
languid pebble
wispy nest
tasty
languid pebble
Hungry now ^^
wispy nest
languid pebble
I love these kinda Pictures!
errant zenith
looks a bit moist-
wispy nest
that was rendered on my hyper flux
errant zenith
what does it mean for something to be hyper?
languid pebble
wispy nest
few more there
languid pebble
We need kinda Ultrapixel for Flux 🙂
wispy nest
im big into the sphere detailed renders
languid pebble
im big into the sphere detailed renders
Dang! I'm not the only one who likes to put things behind glass 😄
wispy nest
Dang! I'm not the only one who likes to put things behind glass 😄
Glass is the way forward lol
languid pebble
Hal-1?
languid pebble
Nice results ...
sterile kiln
Trying my 1st LoRA on Fal.ai (Flux)...
Lets see what happen...
just to confirm, and avoid wasting time with it, LoRAs work well with Forge and Flux?
languid pebble
just to confirm, and avoid wasting time with it, LoRAs work well with Forge and ...
At least there are alredy some LORAs
sterile kiln
At least there are alredy some LORAs
Yes, I know, but they could have been functional only with Comfy and certain workflows, for example.
wispy nest
Most flux loras are trained incorrectly and will on hinder your gen more
sterile kiln
15 min generation estimation, its super slow. Without LoRA, i can generate in 1min.
I don't know if its Forge, or...
Do you have a workflow for Flux with LoRA?
languid pebble
Do you have a workflow for Flux with LoRA?
There are some on CivitAI
I've seen a picture generated with 2 LORAs yesterday
They didn't understand the concept of a Pizza yet ....
sterile kiln
jesus christ !!
Never seen a LoRA that works so well !!!
(Leah Dizon)
wispy nest
jesus christ !!
Never seen a LoRA that works so well !!!
(Leah Dizon)
dude render that same prompt exactly with the seed, but with the lora
see if you actually need the lora
rain gazelle
honest gate
The clothing in the picture is generated into a 3d virtual clothing effect
steep sleet
SD3 Gold model by @wispy nest
wispy nest
SD3 Gold model by <@456226577798135808>
nice cheeky renders there
did you notice i made sd3 gold to run in auto1111
steep sleet
nice cheeky renders there
Not yet - I don't very often use A1111 any longer - but I shall certainly give it a try!
steep sleet
Good - say Should Have Been - and not Should Of Been - much more professional 😄
wispy nest
lol true
steep sleet
Nice to see my favourite SDXL model 0.9 behind this SD3 Gold
wispy nest
the sd3 likeness model is actually slightly better https://tensor.art/models/751438781070140461/SD3-LIKENESS-Dream-Diffusion-By-Dice-v1
thats a SDxl turbo base
steep sleet
I'm finding it quite difficult to do text in SD3 Gold
wispy nest
yes thats just SD3 all over
steep sleet
I tried to get it to say "Bless These POMGRANATES!"
wispy nest
did you try my top trained SD3 checkpoint, thats better than gold https://tensor.art/models/753956516668680191/SD3-PLATINUM-Dream-Diffusion-By-Dice-V1
steep sleet
the sd3 likeness model is actually slightly better https://tensor.art/models/7...
Running out of storage space with these ginormous - but tremendously exciting - models!!! 😄
wispy nest
I found SD3 to be so annoying for text i actually made a lora for it https://tensor.art/models/751649208697700869/SD3-Text-Fonts-v1
steep sleet
did you try my top trained SD3 checkpoint, thats better than gold https://ten...
Bookmarked it!
wispy nest
I was going to train one more SD3 checkpoint then Flux came from know where so i dropped SD3 out
its a very hard model to train , but does actually have the potential of being someting really quite nice
candid surge
wispy nest
D/loaded SD3 Text Fonts - I wonder if its a tad like the Harrlogos2 LoRA?
the fonts lora was made for my SD3 platinum checkpoint rather than my SD3 gold
clever oar
They didn't understand the concept of a Pizza yet ....
I recently saw a photo of gremlins from AI and decided to re-watch parts 1 and 2 of the film Gremlins😎
old but very good films
steep sleet
the fonts lora was made for my SD3 platinum checkpoint rather than my SD3 gold
Note taken ... will try later 😄
unique condor
testing out making some 1080p wallpapers with flux. using a 1344x768->upscale to 1920x1104 with 4kultrasharp->second pass denoise 0.3 with 20 steps->center crop to 1080. seems to work surprisingly well. it can also handle doing it in one ksampler at 1920x1104 as well, but i kind of prefer to do a two pass workflow on stuff like this
smoky vigil
Nice to see vehicles starting to look decent in AI
Well defined parts that seem to fit together, not much melting together. Nice.
valid creek
Young cartoon male avatar, 30 years old mature male, wearing black T-shirt, waist-up, blind box style, Pixar animated movie character style, Disney cartoon animated movie style, c4d high definition rendering, realistic pvc texture, 3d rendering, oc rendering, 8k, white background, gloss, front, best quality, --niji 6 --ar 1:1 --no glasses
languid pebble
I recently saw a photo of gremlins from AI and decided to re-watch parts 1 and 2...
Cool, should I rewatch, too?
clever oar
Cool, should I rewatch, too?
100%
languid pebble
Downloading 🙂
clever oar
i dont know how but film without new effects looks nice
languid pebble
Loved them when I was young 🙂
clever oar
classic never die
languid pebble
Some do .... like the old Bond Movies 🙂
clever oar
and bladerunner
languid pebble
and bladerunner
You could also watch: batteries not included
clever oar
i will check it
clever oar
You could also watch: batteries not included
ufo robots?)
languid pebble
Yes ... kinda 🙂
clever oar
looks nice
clever oar
wispy nest
gotta love a bit of plate art
languid pebble
wispy nest
lol
languid pebble
As long as you don't serve an open monkey head ...
clever oar
errant zenith
the drift wood gibson les paul
I don't know about you but if your guitar looks like that then I think it's time to contact the seller about a wood eating bugs
wispy nest
I don't know about you but if your guitar looks like that then I think it's time...
lol true i have a few les pauls and pleased to say they are not infested lol
clever oar
candid surge
errant zenith
lol true i have a few les pauls and pleased to say they are not infested lol
😔 they start to crumble after the hole phase
errant zenith
I'm sure there's better models for this
wispy nest
candid surge
just trying smallest flux model
wispy nest
boogie time
errant zenith
just trying smallest flux model
😭 I don't think it's worth while then
candid surge
that is a quick conclusion based on only 1 image
errant zenith
true..
candid surge
here's another
errant zenith
It's doing something strange to the people, like there's an outline to them, you sure you need to go this low?
candid surge
yeah i know flux is SUPER real, making stuff more real than actual reality
languid pebble
It't can not be realer than reality ... 😄
candid surge
but it can do other styles and is quite prompt snappy, and the smallest model runs on my 12gb vram setup
errant zenith
smallest which is?
candid surge
it's like photoshopped smoothed and polished reality
errant zenith
that's.. a render right?..
candid surge
wtf
errant zenith
it's like photoshopped smoothed and polished reality
bro I can see the ai
it's so obvious
candid surge
i know its flux
errant zenith
😭 is it flux? the consistency is too consistent
wispy nest
thats sd3 likeness model https://tensor.art/models/751438781070140461/SD3-LIKENESS-Dream-Diffusion-By-Dice-v1
candid surge
no, not that video, idk probably not flux
wispy nest
all my work is renders
errant zenith
all of it?
languid pebble
You have to much time 😄
wispy nest
You have to much time 😄
i wish i did lol
languid pebble
i wish i did lol
German?
languid pebble
To many tools to play with ... it's starting to become confusing ...
wispy nest
love a free tool to mess with
sterile kiln
dude render that same prompt exactly with the seed, but with the lora
I don't need to do this to realize it. It's not a star that's been highly digitized for at least 20 years (or even not digitized at all). I've done enough trains since 1.4 to know that it's quite efficient x)
languid pebble
I always forget the name of the tool for these videos ...
wispy nest
candid surge
wispy nest
batch editing bro
languid pebble
Ohh... I thought of something I've had on my google collabs using video input ....
Deforum ....
languid pebble
I've used it a short time before SD2 was released ... trying to do a music video. But than messed up VAEs ... and stopped cause 12 hours rendering has been to long for not learning fast enough ....
wispy nest
languid pebble
Cool!
wispy nest
old tools but ive been in the game a long time
languid pebble
Looks like ... the first time with deforum was pretty hard ... they've improved it a lot!
wispy nest
yes got way better
languid pebble
I had to calculate picture numbers and so on ....
wispy nest
https://youtu.be/jiDmi9E_890?feature=shared this is my latest LLM the promt for this explanation was , Give me the 10 stages of ai growth in depth...... This is what my trained LLM gave me as a response
he will tALK AFTER THE INTRO
languid pebble
See it ...
rain gazelle
languid pebble
It's a crazy time 🙂
sterile kiln
Hey Omnia, nice one
Guys, found solution for LoRAs with Forge.
LoRAs works only with flux_1Dev (not with nf4 ones), then, you have to put clip_l, T5xxl and ae in VAE/TextEncoder field, and boom, it works.
greatly faster than Comfy
wispy nest
If you want a better flux checkpoint than flux1 and Flux Dev then here it is Hyper Trained Flux https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1
rain gazelle
Hey Omnia, nice one
thanks 🙂
rain gazelle
bathroom break? xD
It's still beautiful
it's supposed to be pensive and determined 😐
sterile kiln
it's supposed to be pensive and determined 😐
one doesn't rule out the other x)
rain gazelle
one doesn't rule out the other x)
touché
wispy nest
sterile kiln
wispy nest
sterile kiln
wispy nest
sterile kiln
@wispy nest
Left: Without LoRA, Right: with LoRA
(same seed and conditions)
my dataset
completely obvious
languid pebble
Big brother ... ... ...
steep sleet
Flux Dream Diffusion checkpoint - trained to make excellent text!
languid pebble
But hell??? 😄
languid pebble
errant zenith
greatly faster than Comfy
how fast are we talking
pallid ruin
Its beautiful 🥺
deft bison
Its beautiful 🥺
Thanks a lot ! ...
sterile kiln
LoRA of myself with Flux, it's just fucking incredible.
Its... me
pallid ruin
LoRA of myself with Flux, it's just fucking incredible.
Its... me
How many images did it take for training? 👀
sterile kiln
How many images did it take for training? 👀
24 images, ~15minutes (1000 steps), WD14 Caption, on https://fal.ai/
But cost... 5€ per train
fleet goblet
sterile kiln
how fast are we talking
~2 minutes for a 1920x1080, 30 steps
fleet goblet
nocturne oak
The Restaurant at the End of the Universe
fleet goblet
honestly with how effect these quants are
I wonder if anyone has decided to do gguf clip encoders lmao
Saving as much ram as possible should be prioritized if perplexity is manageable
candid surge
im running a gguf flux version
errant zenith
~2 minutes for a 1920x1080, 30 steps
10 minutes. 896x1152. 16 steps. What you do to your forge
candid surge
sterile kiln
<:bobagirl:1045338327605321768> 10 minutes. 896x1152. 16 steps. What you do to y...
Maybe we have diffenrents GPU, i have a 3080ti with 12 Gigs
and using this
errant zenith
yea- I got half of your vram
but I'm using the dev version from black forest labs
not sure if it's less optimal for my setup..
though I don't think I'm getting out of this 10 minute generation
sterile kiln
6 gigs is very low, I changed from a 2070S (8 gigs) to a 3080ti, because of this
fleet goblet
though I don't think I'm getting out of this 10 minute generation
Nope.
even the 4_0 variant needs 8+gb
errant zenith
perfect text 😎
I think it's xformers or flux text is a lie
though photo realism is pretty good
candid surge
fleet goblet
perfect text 😎
4_0 can write text fine
it did "Required" instead of "required" because of my own text lol
candid surge
yeah i like how it just takes imperfect texts and renders it completely as it is
fleet goblet
4_0 is really no different than fp16 at text gen from what I can tell
It's really no different at all. The perplexity loss from this is fascinating
errant zenith
😭 vram shouldn't have any effect on the result?
fleet goblet
Though I wish we had 4_k_m variants
fleet goblet
😭 vram shouldn't have any effect on the result?
Except it does when it has to offload and reload constantly.
errant zenith
uhuh..
errant zenith
Well I sure hope that's not the issue
fleet goblet
I'm on Intel Arc.
And I'm running it at 9GB of VRAM.
Alongside OLLAMA and T5XXL being onloaded and offloaded in order, then purged by that layers node at the end
errant zenith
_> well if I could put it onto ram then that would make things a bit easier
candid surge
do you have enough ram, or does it need swap hd space?
errant zenith
I'm not
errant zenith
Why?
fleet goblet
because it allows proper memory offload.
and it also runs the clip encoders off sysram
and not vram
candid surge
broken lens effects, cracks in the image, dirty text on the wall "FLUX is SH1. %%*!~lol", cyberpunk city, garbage can, papers on the floor, metal scarps
errant zenith
do you have enough ram, or does it need swap hd space?
🤔 32gb
candid surge
should be enough, I run flux on 12gb vram and 16gb ram
errant zenith
and not vram
sounds more like nothing is gonna use vram
sterile kiln
fleet goblet
You only have 6gb of VRAM.
Even the T5XXL FP8 cannot fit fully onto that.
So yeah
🤷♂️
fleet goblet
fleet goblet
Is there --novram XD
--lowvram's your best bet
errant zenith
so do I want it all to just be loaded on my ram?
fleet goblet
so do I want it all to just be loaded on my ram?
Wait question
Have you tried the nf4 models
fleet goblet
Wow. No wonder why you're imploding.
You're trying to run the full 22gb safetensors?
errant zenith
👀 guess so
candid surge
fleet goblet
Nobody here can do that. Not with 16gb of VRAM, or 24.
They have to use methods to decrease it otherwise the model just doesn't fit
errant zenith
I've always just let it do it's job, not if it fit or not
fleet goblet
I'm running the 4_0 variant ofthe dev and schnell
you'll get much better performance on that
as unlike the full fp16 models these are 6.7gb large
lmao
errant zenith
isn't there noticeable change between 4_0 and 8_0?
fleet goblet
Perplexity between 4_0 and 8_0 is incredibly hard to tell between.
candid surge
flux1-schnell-Q4_0.gguf 6.4G
sterile kiln
i can smell a Gamecube player
errant zenith
i can smell a Gamecube player
I bet you don't know how much vram the gamecube has
errant zenith
Well you just have to crunch the model down a bit more
errant zenith
if you divide the resolution does it use 4 times less vram?
sterile kiln
I bet you don't know how much vram the gamecube has
not sure we can talk about Vram with the Gamecube's architecture x).
Anyway, I've had it in the past.
fleet goblet
No.
It plateaus the lower res you go.
But is exponentially more if you go higher.
errant zenith
I see
fleet goblet
aka its just exponential
errant zenith
is the vram on a gpu different from ram on the motherboard?
fleet goblet
Yes. It's both a different type, at different bandwidths and speeds.
VRAM is meant to be ridiculously fast.
for example
DDR5 is able to be clocked upwards of 6-7000MT/S
errant zenith
what's a mt?
fleet goblet
while GDDR6 is like double/triple that
clever oar
fleet goblet
what's a mt?
Megatransfers per second (MT/s) refers to the number of million data transfers that occur per second in a RAM module. It provides a more precise representation of the actual data transfer rate of RAM
errant zenith
while GDDR6 is like double/triple that
I don't suppose we gonna see DDR6 in terabytes anytime soon..
fleet goblet
royal monolith
limpid lichen
I don't suppose we gonna see DDR6 in terabytes anytime soon..
covid delayed a lot of factory building. ddr5 is spec'd for 512 dimms but they haven't got modules for it developed yet. ddr6 actually just got talked about this year and specd out a bit. no more dimms. cam 2 is the future.
also that means motherboard manufacturers are going to fk up a lot and the first wave of ddr6 boards are going to suck since camm2 is so new. it'll be fun. i'm still on ddr4 yet. i think ddr5 will have a long future yet.
oh but thinking about it a little, camm2 is actually addressing a lot of EU right to repair regulations too, so the spec requires that it's repairable by end users and that might mean crappy manufactured boards will be more repairable and saveable. it's all meant to reduce ewaste, so maybe it's not a big bad deal. we'll see. no more dimms is a big change however you look at it
coral veldt
Anyone want to help make him look better? I suck at describing a zombie
pure monolith
Anyone want to help make him look better? I suck at describing a zombie
I am sure there is a zombie Lora out there to help with making (turning anyone or anything into) zombies.
rain gazelle
clever oar
resident evil police station!
fleet goblet
clever oar
flux and sd3 cant generate nemesis from re
gentle blaze
24 images, ~15minutes (1000 steps), WD14 Caption, on https://fal.ai/
But cost......
I just trained it, and with flux, it's almost perfect; it gets the proportions wrong, making the head larger than normal. I suppose it's because of the training photos
fleet goblet
That's better.
Uh...
Are graphic images allowed?
What I've made is something I was not expecting.
Eh this one's better.
This one isn't as gross looking.
This model sure is good at making real horrors.
2560x1440
torpid cove
beautiful girls
coral veldt
Finally got it right
nimble mason
errant zenith
oh but thinking about it a little, camm2 is actually addressing a lot of EU righ...
that's awesome but I don't think I get it
languid pebble
The Good Morning Coffee has an eye on you ^^
jovial tiger
steep sleet
Flux@nf4
tight silo
can use a 1 terabyte ssd so that I can get more memory
grand sapphire
