#๐๏ฝsd3
1 messages ยท Page 93 of 1
i could be mistaken but think the quantized versions are limiting a lot of the features
Limiting? How so?
Well its obvious the original Dev is the best model everythign else is a compromise
i tried image2image with gguf, nf4 both dev and schnell, dont think they are producing correct results
but you must learn patience to use it
a simple example i can give about the quantized gguf is that some of the samplers are producing low quality outputs
THat is not an example. Can you provide a precise example with image?
and that Dev pure does not have this issue
not at the moment, i have to sort out the images
also running some other tests and im downloading non quantized versions for closer look
then I'd be careful about making the claim
i had those models but rushed to delete them earlier
That GGUF is the issue
my claims are based on actual outputs
Personally, I have not seen this. And I compared the pure Dev with the GGUF on many instances
some GGUF, not all
lcm?
Never used it to be honest. Ok, sure
any thoughts on these versions? these are not quantized are they? https://huggingface.co/Kijai/flux-fp8/tree/main
I tried it
they are
compromise
but not bad
I always go back to shnell coz i am too impatient.
4 steps. please
not 20
Did they do something with the inference?
My machine is unchanged but the pure Dev is going way fast
This won't sound fast, but trust me, 3m30s for 30 steps is completely abnormal
No change in model. Absolutely certain
anything over a minute for a 30 step 1024/1024 image is not going to fly
its sis not dude
comfy has done a ton of work on memory management and loafing unloading stuff
with pure unmodified Dev?
oh no
ah
we are talking about SD15 back then in the old days ( 8 months ago)
flux doidnt exist yet
Yes, I was talking about pure Dev. I guess my comment came in mid conversation of yours, and led to this misunderstanding
In any case, I am still scratching my head
Granted this is the fp8 version, but it is still the plain UNET file. 17gb. I know what times I got. It is even why I hesitated to run torq's test
whne i started SD15 was king - it was terrible to run it. barely... SO I upgraded and can do SDXL now ok and here we are pixart, ckolors, flux and so on....
it never ends
Whats next...
25GB download and 64GB vram minimum...
And it almost weekly. Like sometimes every week a ne wmodle comes out, not exagerating. lol
Flux came out of nowhere
lolol. Welcome to the wonderful world of tech, where the less opulent play a never-ending game of catchup
But there is this to consider: I always viewed SD and co. as being of the same ilk: playing catch up. I was never really content with the SD results except here and there. Flux compeltely changed that.
It sucks for someone like me. I just like to figure somethign out and work ith it forever if I can. I use like 20 year old programs in some of my pipelines.
I just want soem AI stuff that does what I need it to lol
I dont nee dupgrades every 2 months.
I actually deliberately chose to focus on my CPU tratehr than GPU, since my gaming needs and wants are pretty light, and most of my stuff was extremely CPU intensive
Yes, as far as one off image making fluz is there. We now need consistency and later on video.
and were ready to take on hollywood
I have zero illusions of that last.
No, we really won't. The problem is not one of tech
And competence. It is like saying you can make video games and thanks to UR5 you can now take on the big studios and AAA titles. It is to have no idea how big the project and amoount of work is needed
Ah well of course
a madman or woman has to come along with enough time and determination and vision to do it yes
of course
but it will be possible
It honestl already is
WHat is it you think will be so different?
we need consistency
we dont have enough continuity to tell a story with AI
Everythign shifts and morphs.
Start piece by piece. Write teh story
Yes we can write stories
But until the characters and objects and locations shift fro frame to rame its not gonna work
we HAVE consistency. i gave you one yesterday, and one today, with extrem consistency. you obviously have your head in the sand
It has to be consistent down to like the moles on the face
you don't even get that in movies shot on film
Heh. I think writing the story is already the first major stumbling block
We all have stories
Sure, but the question is of quality
doge is discording while high apparently
I started Rebel Moon, Director's Cut. Got through maybe 15 misn. Then I stopped it it was so annoying and cliche
go watch this, https://new.reddit.com/r/StableDiffusion/comments/1f0b45f/flux_is_a_gamechanger_for_character_wardrobe/ then talk about whether we have consistancy and high quality or not
Screenplay by ChatGPT and directed by Grok2. XD
Gawd. I have seen the writing it spits out
Creative writing that is
the essays are mostly ABC
you only think you have
Oh. Right. I only think I have
you've seen plenty of its writing, only you don't' realize it. you only realize what is posted as from it, or that you got - but what it can do is extremely good and you can't tell
I'm impressed you know in such detail what I have seen and what I can tell
i know, because i know what it can do. and based on your comment, i know you've failed to detect a lot of its writing
One day i will also enjoy those special powers
not likely
yup. you're looking at your toes and the rest of the world is walking past your obliviousness
lol
you and doge go off and enjoy living in a dreamworld
dont let cyrtsalman get to you hes the local goblin.
do you have a LLM detect trophy, whatever that means
you remember that when you point a finger at someone, you have multiple fingers pointing back at yourself
Dude. I have been supporting myself for decades as a writer. His opinions have no impact on me
i can guess how many harliquin romances you've written
Is that the Joker''s GF?
guess that means you don't have a decent iq
You know how every classroom has a bully/idiot. Like the Dr Magnusson in the hl games
thats crystalwizard
it is what it is
Don't you need at least an impression of strength or superiority to be a bully?
X (dont deadname) credentialed
Assholes are equally distributed across all groups.
Now, now now.
heh
you would know
He reported everyone in this channel probably several times over
thatsa bully
talks down to people
thats a bully
insults eveyrone
thts a bully
hes just another jerk
Crystal is entitled to his delusions. I think he simply cannot discern the writing peccadillos of the AIs, their systemic giveaways, and therefore cannot imagine they exist
crystal is a saint, idk how anyone could not like them
They are not incompetent writers mind you. But we are talking actual fiction, not a couple of paragraphs of descriptions
it does not matter whatsoever what he thinks about anythign or SD3 or anythign else
what matter sis that is he a nice human being or a dick
and he seems to be a nasty dick
thats the bottom line
he is not a decent human being
he is nasty rude condescending and rough and harsh and a bully
he proved it many times
and has a 50/50 crappy picture ratio
I don't view him as a bully. Just a troll. A bully for me is different, but maybe I am insufficiently exposed
yeah
a troll fits too
it wa sreally so nice here before he showed up
hes the only jerk here
i've been here far longer than you have
?
like why
Reporting people?
oh yeah
he reported everyone for nonsense
lol
every othe rmessage of his was
"reported"
doge is very good at fabricating facts
ypu reported me
thats why i think of him more as a bully
rather thin a plain troll
he likes to harm people
fklex
threaten
he is a bully
I still would define that as trolling. Reporting is not bullying
its just silly
it is silly
but it does harm people emotionally
and otherwise
it is a form of bullying
he likes to attack people
Mmm... Threatening to report might be a form of bullying, Actually just reporting is not
Like you are reported now and maybe discord will ban you and whatnot
like its vague threats
vague uses of force that he gets off on
Meh. I have been excluded from a number of Reddit groups for pure inanities. I would not feel any differently
sometimes the do you know who i am defense comes out, DO YOU KNOW WHO I KNOW
Do you know who I is?! I am the crystal wizard
!
I can call the dodge coin "it"
and report "it" for nonsense
because I am the crystal wizard!
That's a threat? DO you know who I know? I thought that was just insecurity
Well, I can see you have a major hangup on her
Its not a her
id be shocked if it was a female
Well I do. why because I am tired of one or two jerks ruinign a family or a group
a handful of countries ruin the whole world
its the only person ive seen attack people here besides spam trolls that get booted
So yeah I am sick and tired of peopel likei crystal wizard
absolutely sick and tired of theuir evil.
well, why not just put her on your mute list?
I despise them and I ahve no clue why mods aent doing their job
when they see bullying and harrassment go on for weeks
and months
I muted it
but discord is like not muting them 100%
I always see message by crystal dork and do you wanna see it
its quaint he thinks i read his garbage still
If I block crystal wizard I dont want to know he is on discord yet this is what I get.
Dodge, my tip is just ignore the notifications of hidden messages entirely. I have a cople on my mute list. I gurantee you I have no curiosity or concern whatsoever
The importance of a rando on the internet is subzero. I mean, I already care very little about most people's opinons, of people I actualy interact with or know. Imagine some pseudo on the internet?
You should see some of the PMs I've gotten (not from this server ever), just block anyone who stresses you out or worse. The block feature is kinda awesome.
I understand. It just goes depper than that. I have known people like him in real life. One of these cna ruin a classroom or family.
I know I should develop a thicker skin lol.
this is the Interwebs! XD
Like imagine lording it over people "I am reporting you!"
Like who do you think you are exactly?
A classroom? Well, I admit I am so long past that scene it has no residual effects left, but even then I was the nerd, and had no issue ignoring most
I don't beleive in the thicker skin bs, but if someone stresses you just block. Much less stress ๐
internet crossing guard
Sadly the bullies I have known were aggressive.
You ignored them but they still slapped you.
fortunately you don't have to learn martial arts on the internet.....
Hey.
Remember that scene in Jay and Silent Bob when they went around and beat up everyone who talked smack about them online lol
XD
I learned martial arts, they are cool. And I am short, and still very much the nerd.
Martial arts are definitely fun ๐
I did learn a bit actually.
And it was not an easy task. I had to take a 45-min subway to get to my kungfu school, and then 45 back. it was a COMMITMENT
As a 13 year old little girl lol
yeah, I was 14
looked like In was 11 and descended from some starved people
Same, but in Vancouver a bus takes that long to get anywhere lol
Took like half an hour.
you guys are so lucky! I didn't take it up until my 40's! I tried when younger, but didn't find one I liked yet
๐ฎ
I just did hiking, kayaking etc before that for active stuff
I went crazy on them, watching Bruce Lee movies fram by frame. lol. Had my mother buy me a large punching bag for my room. and lied my way to going every day.
This last deserves an explanation.
The monthly fee gave you the right to two classes a week. Each lasying about 2 hours
This was in the 80s
I soon noticied the master was never there and they had different teachers on different days
Bruce Lee is epic but in eastern europe ot was jean claude van damme that was the big action star :)))
so no teacher ever saw me more than twice a week. lol
Steven Seagal too I must admit for a brief moment - a 5 year window or so
So since we are way off topic anyways, for a potluck tomorrow I'm making frittatas. Should I make the Greek/Italian flavor version, or the mexican flavor version? Basically black beans, cheese and a bit spicey, or extra garlic, and feta cheese? ๐
I think the teachers spoke among themselves though, because two of the best teachers, one who had been 5-time French champion, caught on but said nothing... to the master.
heh
Greek/Italian version for sure
Though tbh there are very few martial arts teachers who kick someone out of a class, that's extremely dedicated
Yeah, he casually paired me with another guy who was taller and a belt above for the sparring, and when i began to protest he said that with the amount of classes I was taking, I shoujld be fine. I gawked, shut up, and red-faced put on my gear. lol
(I was still a pipsqueak, don't let any of this suggest otherwise. ๐
Martial arts kicks etc. were impossible to prompt for prior to Flux!
even with those fancy pose workflows/images
Dont look at the toes
Look, that is not horrible for AI
I did so many good parkour images with flux, also SD3, but with sd3 it had to be at least 4b, preferably 8b
Foot fetishists are in shambles
Prompt the same but give her some boots.
I'm always amazed when people post feet on DA and they have FIVE toes!!!!
Controlnets
Yeah those, but still... ๐ฆ
Or they start with a reference image and do a low denoise
Like 0.4
Enough to hide the fact it was a real image originally
Saw a lot of that shit early on with video generations of tiktok dancers lol
Last year with v5 I think, I had Midjourney give me a pic based on this prompt: taekwondo kick, drip painting on white background, style of Leroy Nieman
So believe me the foot and posture you showed is not bad
I had it doing all kinds of funky stuff mind you. Ex: baseball player swinging bat, drip painting on white background, style of Pollock
MJ is unique in its ability to dole these out
young woman, blonde hair, yellow eyes, determined look, wearing choker, bare shoulders, raw photo, cinematic.
that with image2image ....
young woman, redhead, yellow eyes, determined look, wearing bead necklace, bare shoulders, raw photo, cinematic.
using denoise 0.60 still not getting the output right..
BTW, I tested lcm and frankly the results of Pure Dev are just as shoddy and slightly out of focus
this was lcm sampler on schnell, but can't seem to get the similar result with gguf
also this
results with lcm sampler that i used here are not highly polished but that's the artistic look of it, giving it a somewhat worn out fantasy theme look
i get 5 toes on txt2img with pony in like 6 out of 10 imgs so just gen a lot of imgs and delete the bad ones
New AI image generators top the list: Ideogram v2 and Mystic. Here I compare them with Flux Pro and Midjourney v6.1
#aiart #ideogram #mysticai #fluxpro #flux #midjourney #ainews #ai #agi #singularity
๐ Get an Exclusive NordVPN deal + 4 months extra here โผ https://nordvpn.com/aisearch Itโs risk-free with Nordโs 30-day money-back guarantee! โ
Id...
Madness.
I did notice some weirdness when I tried img2img with flux.
So what is this Mystic stuff. I guess I have to check tis out now...
flux just misses to make changes
Midhourney dropped the ball.
basically this with img2img
a blonde girl, wearing a sundress, in the garden, analog photo, cinematic.
a brunette girl, wearing a pullover, in the garden, winter morning, analog photo, cinematic.
TLDW version? ๐
YouTube clickbaits

Hmm
You know what
That's what we need the giga ais that can watch videos for us for
They'll watch those clickbaity videos
They work fine, it's just that flow based models don't have the same kind of sigma curve, so you usually need denoise values in the .55-.8 range
Fortunately Geminii and Claude can do that ๐
Or at least I've found that range to be similar to 0.3-0.4 for sdxl models
You can also slap an auraflowshift node between and set it to 1.0 and use a simple schedule
It will give it the diagonal sigma curve
It overrides comfys built in forced flow curve for flux
i used denoise value between .06 - .08, while the first value doesn't have enough impact and the 08 changes the facial look on a portrait
Not 0.06, 0.6
As in 60%
right.. typo.. it was 0.60 and 0.80
I'll make an example in a minute, waiting on the damn washer to toss the clothes over to the dryer
sure, that would be nice, ive been trying this for hours
isnt Claude that img bot that was making the black George Washington imgs ๐ฟ
Were they ever paid?
Claude doesn't do images from what I know?
it was the google img bot but i dont remember the name it was all over the news
Sounds like an improvement to me ๐
was it gemini, i forget which name too
yea that was the one,when ppl asked for accurate pictures of the founding fathers it generated them as black women
embarrassing for them
yea they werent even the thicc sexy ones

I saw them months ago maybe they are free now
They were premium users only kinda stuff
For this super tech of being able to watch videos 
damn, really had to push the denoise up high. but anyways, the flow based models are weird compared to the previous styles of models and need really high denoise to make big changes. you'd be surprised how much of the image can remain even at 0.9 or 0.95
and here's part of why:
flux on the left, sdxl on the right, both using simple
thanks, yes it definitely needs high denoise value, but the changes i was aiming for didn't feel satisfactory
the alternative might be to unsample the image down to a single step or two, then resample the whole thing or do some noise injection, but that's a much more complicated topic overall
Gemini advanced gives a free trial, which is 6 months long if you have a yearly drive subscription. (2 months otherwise I think).
Claude has been free so far fortunately ๐
Now GPT is free for the regular, then paid for the fancy create your own GPTs etc.
Yes, it works very differently. I use it in an ipadapter style, with a denoise set to 0.9, using the same prompt with different reference photographs so that it generates different environments in the image. Since the prompt is quite different from the reference image, it usually applies the style of the references. Same prompt, and same seed:
Yep, good example. Haven't messed around with ipa for flux yet though. All I know is flux has a reaaaaaaly strong style transfer when denoising. I think even the ipa addon creator guy mentioned it as well
The obvious alternative is to just use a cnet with a full new image
(canny or depth)
Like if you want some scene composition, but in a totally different artsyle or mood
That's the idea, when I like a composition, I generate a batch of 36 images with different variations
And between 0.85 to 0.95, depending on how much variation I want. 0.9 is a middle ground that I like; at 0.85, there's already a lot of variation, and in some cases, I want it to be that way
Since flow models have such a touchy denoise range, I've thought about making a node to pseudo-remap the range for you so that if you put in a range you'd be used to with sdxl, something like .4, it will crunch the numbers and map it to the delicate flow range like 0.83457 or whatever
yeah sounds like a great idea
You'd just promote denoise to a pin and plug it in
I'd just have to f around for a bit to figure the ranges and equations out
I have this issue with CFG++ and ancestral samplers
Cfg++ uses like half the normal cfg
at high steps like 300-600 steps, CFG++ ancestral samplers need something silly like 0.003 CFG
that was the alt mode, the full mode is more like 10% CFG
specifically, 12 normal CFG = 1 CFG ++
the alt mode was comfy devs getting confused, as far as I can tell
the CFG++ paper only has one conclusion- that the final CFG number should be between 0 and 1
there is something else weird happening where ancestral samplers at very high step counts have more CFG burn
so they need incredibly low CFG
I am not sure if this is correct behaviour or if ancestral samplers are not implemented properly
The man from the car! ๐
It's so so. These were definitely cherry picked.
Dude really wanted to get on the shot cx

I'm calling the extermination team
that Mystic model looks horrifically over-fit
I think it has overtaken Kolors
as the most over-fit on midjourney model
I prob would have rated it higher has they bot claimed to be better than the most well known realism ai art app.
Sd3
the guy got over-fit way past orange he is actually red now
Does this bot got free nsfw?
The trick to that program is to have it keep making more until you get good ones.
Sdxl
Try mage.space
not rly, if a model is over-fit that also means less diversity
so running a lot of samples doesn't work also
High contrast != Overfit, but if you start seeing the same 10 faces and poses, then yeah
it depends on what the prompt was and whether it is able to do other styles ye
but I am willing to bet it will be over fit
How to try it
Flux is over fit, by definition
That's why there's shallow dof in every photo
And why it sucks at traditional art
Well, not sucks, but just isn't great at
ye the shallow dof is annoying
as someone who loves deep dof
I am staying on SD 1.5 or SDXL in general
Yeah that's pro though. The distilled versions are extremely unreliable at doing a lot of common art themes and styles
I'll try the same prompts with Dev next to see how it goes. How I manage is that I have Claude help with the prompts ๐
Oh you'll get some decent results here and there, but it's just highly unreliable and you end up having to fight the prompt a lot
and I made an artistic styles glif as well, which pretty much focuses on any artist I mention...
I remember spending an hour fighting dev to get it to make thick textured paint strokes
Forget what word salad I used, but eventually I got it to where it would work like 30% of the time
its a weird thing with flux
that I find it very uncontrollable
but some people really love flux and don't like it when it gets criticised
Fortunately with pro I've barely had to cherry pick, now to try Dev and hope I"m not too disappointed. SD3 is best at various art styles, I really need to work on that SD3/Flux workflow I was using (I've made too many workflows and not organized them well enough ๐ฆ lol)
30% of the time is definitely not useful ๐ฆ
Schnell has a shitload more variety than dev. I use it a lot to prep and then img2img with dev
They kind of hide that fact with their obnoxious graphs
On their site
Hmm, can you somehow use schnell, then dev in the same workflow somehow? ๐
Yeah, just make sure you have the ram to hold them for swapping
I deleted Schnell yesterday, oops lol
Er prob not lolol
The problem with schnell is that the images are always so bright and blown out like using a cfg of 8-10 with sdxl
But it has a ton more variety
The problem I keep havin with Dev is that the images keep ending up looking as if I ran out of printer ink, faded, hmmm
Sometimes you can trick dev into spitting out more variety by lowering the overall megapixel count to like .8 or something and upscale and resample back to 1mp or higher
Like make shit in the 768x768 range
that might even speed things up at the same time!
Sometimes, definitely for exploring seeds though. Reminds me of working with cascade when you do a schnell->dev workflow
At any rate, they are great models but no model is perfect. They definitely made them with the average user in mind in the sense that they focused mostly on being able to make good people
But they're still great at a lot of other stuff
Their ability to make hybrid creatures is pretty amazing!
and SD3 knows who Pietre Bruegel the Elder is. No complaints at all from me!
yeah this
Anyone know how to prompt 2 separate characters in the same scene, with flux? Each doing different things.
its the opposite of the style I like (soft and pastel colours)
use the prompt character 1 and character 2. discribce character 1 and left or right side then do the same for charecter 2
I got this error when I tried to use the depth map. Any solution would be helpful
FLUX HYPER CHECKPOINT
Thank you ๐
egyptian bug is nice yeah
In mech armour lol
Pro
sadly pro is so much better
Doesn't look like "Beksinski and El Greco" style in the slightest though ๐ฆ
Too bad pro can't be run on our own systems
use hyper its like pro and works onsite generators and local
Sdxl
How are things on the Antartica Science Base? ๐
FLUX HYPER
Prompt?
"spectacular digital rendering of a (transparent:1.8) Zonda car, revealing internal mechanical components such as engine, car chassis, suspension, and internal wiring, detailed textures, detailed machinery, accurate lighting and shadows, 8k quality, intricate patterns, high-definition, glossy finish, vivid reflections, perfect lighting, showroom BREAK"
just change car model to what you require and works best in landscape
if you want the Flux Hyper Checkpoint https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1
1.8 etc works?
Does that mean dynamic prompts work also?
yes I trained it to use those
Is it on civitai or any other sites?
Im banned from civitai . You can download it from Tensor Art and Shakker.ai
why ban
Alt account ๐
Tensor it is then. I recommend huggingface also ๐
Can't sell generated images, nm
They didnt agree with my views about the way the used there onsite generator with buzz and not allowing any of my models to used on it. Even tho the civitai community voted my checkpoint as number 1 . Civitai then had to pay me 100k buzz for winning then 2 days later they banned my account and my 32 models . I was ranked 4th master generator with just under 2k followers . Its was upraw lol
wow ok
a lot of the top devs left with me and we went over to shakker
Not huggingface? ๐ฆ
there is a fun report video on youtube by a guy who filmed all the screen shots of it happening
This won't be good get there and leave comments and down vote the change. voice your opinion.
Path to Sustainability: Mid-year Update
https://civitai.com/articles/5833
CivitAI's Alternative Shakker
https://www.shakker.ai/
And support Acly for his AI Diffusion for krita
https://github.com/sponsors/Acly?frequency=one-time
this guy
Shakkar seems so sketchy tho, abd they do bait and switch stuff ๐ฆ Glad you are on tensor as well
nice
Flux Dev + LoRAs
problem with tensor is the model permissions are way worse than civit, and every model seems to have a different combination of permissions
Ohhh I should see if that model on Shakkar allows selling 9f images ๐
Tensor also doesn't allow nsfw (anymore) ๐ฆ
ye Civit is by far the best for low restrictions
The real reason I like it is it's easy to see the type of images each produce
HF is definitely not as easy that way
yeah getting checkpoints or loras from HF is pretty unusual
its not rly made for that
it has all the CSBW stuff though! โค๏ธ
Recommend me some outstanding Flux LoRAs? ๐
heh... well I've made 3 so far ๐
They are of a very specific theme though LOL
That dark fantasy one looks awesome, I haven't tried it yet though
Duc Haiten has made several, they are pretty good
Wow, there's suddenly 10 times the amount of flux loras today!
try this prompt with the neg prompt to bro spectacular digital rendering of a (transparent:1.8) Ferrari car, revealing internal mechanical components such as engine, car chassis, suspension, and internal wiring, detailed textures, detailed machinery, accurate lighting and shadows, 8k quality, intricate patterns, high-definition, glossy finish, vivid reflections, perfect lighting, showroom BREAK, macro photography, extreme close-ups, intricate details, unseen worlds, capturing texture, emphasizing details, revealing beauty in small things, professional lighting, precise focusing, artistic expression, immersive experience, expert technique
Negative prompt: bad anatomy, comics, cropped, cross-eyed, worst quality, low quality, painting, 3D render, drawing, crayon, sketch, graphite, impressionist, cartoon, anime, noisy, blurry, soft, deformed, ugly, lowres, low details, JPEG artifacts, airbrushed, semi-realistic, CGI, render, Blender, digital art, manga, amateur, mutilated, distorted
On tensor it says no resale of images, on shakker it says nothing... soooo are we allowed to sell the images made with this model if downloaded from Shakker?
its a tricky one. as the base i hyper trained is flux dev , But I changed it up quite a bit . So its kinda a yes no sitution. Ive set as no commercial use as thats what the flux dev base was originally set as . But that license to be honest reads so badly its hard to grasp if you can or not
I make far less than 1mill pear year, I'm good ๐
lol same lets go for it lol
Fortunately they reworded it, also one of their official people came into this channel to clairify things one day :>
Yes i mentioned they would a few weeks back didnt I
Basically if you make less than 1mill before expenses, you can sell it all you want (the resulting images)
Sweet they asked me what i thought and i suggested that to them
I look forward to the day that I have to pay SAI a cut ROFL
Btw, I hear that Mage.space pays model makers well for generations on their app.
More importantly they seem like awesome people
I have not yet tried any workflow with negative prompting. Still using the original workflow from Comfy using the UNET Dev model...
I'm looking forward to 3d printed transparent full sized cars in the future!
yes that sounds interesting, Shakker have paid me out a lot recently, Just under 1000 bucks cash
whoah thats a lot
That sounds 1000 times better than your civitai experience
Hmmmmmm
civitai was certainly no experince I want to relive lol
if your a creator of models you will earn good money on shakker.ai
My loras aren't quite good enough yet, but soon....
that 1000 bucks was over a period of less than 2 months and just by adding my models to there site and image comps
Someone already di it ๐
one comp that just ended was a best sd3 image render 1st prze was 2000 bucks cash and my image won community vote best image $100 . just for posting a few images . its a great return
Yes ive given my prompt to a lot of people
those were genterated on my SD3 Likeness model https://tensor.art/models/751438781070140461/SD3-LIKENESS-Dream-Diffusion-By-Dice-v1
which is a SDXL TURBO hyper trained checkpoint
All my shakker dreams are shattered! ๐ญ ๐คฃ
yes its a porn free zone lol
congrats!
I'll have to read their definition of porn....
that's surprising, even MJ would allow that. Different detection systems I guess
Yes because my uploads to shakker have the prompts included in the image , so some of the context in prompts are read as NSFW
like this image i rendered with flux hyper
with my Flux Hyper even the little details as in these renders ,,,,, A light moring frost on the car body work
same prompt ... but just with a slight change to hollow blue neon glow
I should perhaps make some SFW loras as well ๐
sdxl images with my powder explosion lora
flux?
Yup... Flux.1 Dev
Still using my the MJ free trial ๐ Don't mind the cheezy abs, I prompted hypermuscular..
There is a Flux LoRA competition at Civit for any interested
not rly into Loras, I just use base models
THat would be ok if the base model could create even remotely some of the things I want
If ther top prize isn;t $124,198.76 USD, I am out, ๐ ๐ ๐
I only make the same image over and over so my bar is pretty low
Used it twice so far...
I mean Replicate's is potentially a lot faster but nowhere near as cheap
No idea what Civit uses, but Repliacte builds on A100s
I thought Replicate is using H100s?
$0.001528 per second
$5.5008 per hour
this price is not that great rly
they quote per second prices for a reason ๐
The issue is how many hours are needed
but I will say this: they offer automated labeling of your images
whihc is a nice feture
oh yeah that's nice
Anyone know how many hours to expect using it?
My CivitAI LoRAs have taken no more than :45 minutes... but not sure how they're queued... of if the VRAM is shared or the GPU dedicated etc. etc.
That is ridiculously fast. Cool. Question: how were the results? Did you have to furnish your own labels? I have not used it yet
I used about 300+ images from a public dataset (which I shared as well) tagged with GTP. I upload the whole thing in a single ZIP file.
Here is the link:
and a previous generation with less steps
your captions are at least as important as what you train on
you def want to be looking at those with your own eyes, even if only briefly
No, They do have an option to tag / label your images if you want them to
I can completely appreciate this
gpt4o writes solid captions but it does do some shit grammatically that doesn't make sense for captions
the best example is it'll use commas in ways that separate tags
"a woman standing in a cold, dark basement" = a woman standing in a cold... whatever that is
I wonder what Ideogram uses. I have also heard good things about Claude for making promts. Maybe this translates to good image descriptions too
"creating a warm, inviting environment" = an action creating a warm, instead of what you want which is just "a warm inviting environment"
those are the two biggies
you could probably prompt for it to not do that
NO worries. I am well aware of ChatGPT's oddities in writing
you can but it's tough to talk it out of doing that shit 100% of the time
was this another dream
Act as a creative agent who generates a pair of highly creative detailed 70 word image prompts derived from each uploaded image. Begin each prompt by specifying an artistic style (such as a type of painting or illustration or drawing, etc., for example, oil or acrylic or watercolor painting, or charcoal drawing or engraving or sculpture, etc.) or camera settings, using descriptive elements that explain what makes the artistic style or camera settings unique. Then, continue with descriptive visual elements of the subject, surroundings, and background. Please do not omit any details, especially objects, lighting, colors, and textures, even subtle ones hidden in the background! It is imperative that you carefully consider every aspect of the image, and provide satisfactory descriptions! Also, please be very careful when considering whether the image really is a photograph or actually a detailed painting or illustration. Include elements presented as discrete descriptors in one long 70 word sentence without story. Put the most important descriptive elements at the beginning of the sentence. Only respond with the 70 word prompt, don't preface the prompt with anything else. Use simple language. Limit your response to one sentence. Each prompt should be 70 words long. Please create a pair of 70 word image prompts for each uploaded image.
Surreal digital painting with a deep blue and red color palette, a child standing inside a large, jagged circular opening leading into an underwater scene, holding a basketball, a pool of water reflecting the child's silhouette, subtle textures of underwater flora and the rocky opening, rays of light filtering through the water, dreamy and mysterious atmosphere with high contrast and intricate details.```
that's my captioning prompt
I have had Discord itself, not sure what they use, highlight my grammar with red underlines, and I stared and immediately knew it was full of s***
i use it with batches of 8 images
the numbers are tuned for that... the more images you feed it, the shorter the responses you get
thanks this prompt is nice
been using the auraflow prompt so far
i try to use at least two different captions for every image... onetrainer shuffles randomly between each caption for each epoch
onetrainer?
well, being hardware limited, I can only yuse online tools I think
and have the patience of a saint
kolors
but that Tatooine, Naboo looks different
ah yeah you're right
SDXL has a Naboo lora that gives you true Naboo, its a little over-fit though
Did it work for you? I'm one of the 50% that got the compressed file doesn't work error ๐ฆ
got a notbook or vast setup for that? ๐
It is literally running right now... I'll let you know... just got time to play with it today
I run it locally... it actually works great. Thanks to @torn wharf for the heads up and help with OT
(I just wish I knew what I was doing)
I haven't tried yet
DId you use the normal or the cost an extra 200 buz version? Mine with the extra buzz usually take about an hour. Though in the past with my regular ones I've had some that take 8 hours.
Regular... I am patient
Works out to approx $2 per lora which is the same as civitai
but I am also calling that expensive
Cheaper than a new computer ๐
I can't wait for more affordable options!
20cents vs my impatience... ๐
You think 2 dollars is expensive? Really? A good machine with nothing set up will cost more than that per hour on Vast.AI, and the doesn't include all the ready-made software that is just there for you
A100 80GB is under 2 dollars per hour
uh-huh. Does it come ready made for you to build your LoRA? Send your material and output it? And on a sidenote, how much time will you need to do all that and obtain your LoRA?
My time is a benefit
Do you happen to know of any premade vast workflows for lora, or checkpoint making? ๐
Skill level = html lol
He said he does
I think civita $2 was a good deal when I amde a 1000 image lora once. But it seems to have lowered the amount of images you can use now (500 perhaps)
get yer loras here. hot spicy loras. $2 a pop. hot salty lora balls get yer loras out
I spend 2 bucks just on an espresso out of my home. But I have also used Vast.AI extensively and the process takes time. First you need to enter the service, make a bid on the machine, have it accepted, load your software, even if a ready-made Docker, set up a system to upload and download your files, etc. I did this a hundred times at least, simply because the machines would go offline here and there. All in all, there is time spent on all this, no matter how well-oiled your system is
is kohya trainer ok?
and frankly, a half hour of my time, even if the pure dollar cost were the same, is money. I most definitely value my time at more than 2-4 bucks per hour
but it takes about 5 minutes to go from google.com to a running comfy instance
My fear of forgetting to turn the instance off when I'm done, and my credit card being maxed, has definitely slowed down my lora creation on clouds!
LOL this happened to me once on rundiffusion
That's the one civitai uses. Most people seem to use it, not on civitai as well.
you can set limits though
it is worse than that frankly. If you sit there regularly loading machines to use for 15 or fewer minutes this will reflect on your reputation
Yikes, did it turn off automatically after a bit? Hopefully.
if you like kohya then there is a koyha template in the current vast defaults
it turned off eventually yeah
The stuff about 'saving one dollar' to make a LoRA is anathema to me to avoid that dollar and pay for it in ease-of-use. If I made thousands, I might see the light, but otherwise?
what was your bill? lol
Oh ty for the info ๐
I have to figure out how to install that uncensored LLM I downloaded....
for tagging ๐
but you just boot in and wait about 5 minutes
for it to set up
that's it
and then install stuff like this:
wget -O 4xNomosWebPhoto_atd.safetensors "https://github.com/Phhofm/models/releases/download/4xNomosWebPhoto_atd/4xNomosWebPhoto_atd.safetensors" && \
if [ $? -eq 0 ]; then \
sudo supervisorctl restart comfyui; \
else \
echo "Download failed, restart aborted."; \
fi```
not exactly. You need to know how long it will take, and get your file, and shut it all down
yeah you need to get your files but that little snipped I just put will get anything
It is not "press here and it will be ready when you come back"
you could use a script to shut it down when its done
For those who have tried both, do the cloud lora makers through various programs, end up with producing better resultimg images from your lora, than from using civitai?
Yeah, all this to save a buck. I can really see the light now
if you use exact same settings it would be same
but its a few lines of code that you only have to write once
lol
Dude, my budget is not that of a college kid or teenager. I certainly save money with decisions, but that is nowhere near something I would consider.
Remember the old saying:
The Computer You Want, is ALWAYS $5,000 dollars
I am waiting for the day when NVIDIA pays ME to use their HH5000 MEGA ULTRA GPU ๐ ๐ ๐
I'm sure there are plenty of ML researchers and developers already enjoying that privilege
the Pixart team moved to nvidia yeah
its a bit of a risk that nvidia won't let the next pixart have as nice a license
Dreams don't die based on the present. Chase'em
I do agree that sometimes its worth paying for convenience.
I use rundiffusion.com when I am feeling lazy
they just have everything setup
my current one is about $5k of parts (at resale). my dream pc is up towards the 10k mark now
I will likely try that local software here. Convenicne be damned, because it will be a fun puzzle to figure out. But I will also without question be trying both Civit and Replicate
Yeah, the WANT is always relative. My number of $5,000 was from back in 2000 - 2015 era... nowadays the WANT is more like $55,000 lol
Water Cooled Quad H100 rack mounted behemoth... would be nice
yeah rack server in the basement is the dream
I'm considering doing a home build as well at this point, with used 3090s
and then only use cloud for the big GPUs of 40GB vram and above
use it for home heating
Dunno. I have seen laptops with a 4090 (laptop version therefore 16GB VRAM) for under $2000 stateside
Dual Use... nice... thinking green planet, huh?
I want that site in upstate NY with a built in hydroelectric power plant that company bought to run 24x7 Bitcoin Mining...
and that comes with screen, KB, and many cored CPU
It's the same silly con but, tuned WAAAAYYY down.
advantage of a home 3090 would be trying modern gaming too i guess
i'm in BC so we're on all hydro more or less. there's some carbon sources on our electricity but majority we're green energy on my island circuit. i dont care about saving money for carbon at all. i dont' even believe that individual homes coudl change their energy usage in a way that would significantly offset global carbon emmisions.
I'm just cheap and heating bills are nightmares
I keep telling that to myself when I upgrade the local rig, but it never really happens. Too many other things taking up attention, including my very jealous mistress: chess
Yes, studying chess is very time consuming
modern gaming got pretty good. red dead 2 alone is chefs kiss.
oh yeah i need to try red dead
i've never played the others in the series. just rdr2
it's more or less self contained
not sure about how that is defined, but yeah, it has been a huge part of my life and livelihood
i used to play lots of chess. notsure why i stopped. oh wait yeah i remember. i hit a hard skill wall
my brain with 3 moves ahead ๐คฏ
Nah. Any activity with enough layers can be called hard. Being a top player is hard, but the game itself is not
(I am not a top player, just a middling master)
ive gotten a black eye playing chess
this is what 4d chess should look like. startrek is out of touch with the 3d chess
so many regretful chess moves. especially when you're considering like 8 moves in advance and think you're hot shit, but you didn't see his bishop
or hers if you're playing princess peach
or other women who play chess too.
Here's an article I wrote on chess progress back when I was editor of CB News: https://en.chessbase.com/post/chess-progress-making-the-big-leap
Contrary to some fields where controlled steady progress is the norm, chess is dominated more often than not by timely bursts forward preceded by periods of seeming stagnation. It can be tricky knowing when it is one or the other. Here is the tale of just such a leap forward, as well as tips and recommendations to help you make your big leap.
chess hasn't been too exciting since fischer stopped being so public about communism
this is the next benchmark. when will be able to prompt a chess game notation and get an accurate board
Squares are always the problem
We need balls
propaganda of we need your balls here
keep calm. ball on.
Actually, the number of squares in chess boards is where the AIs really stumble tragically, but truth be told, movies seem to get the plain orientation of the board and pieces wrong more often than not. Some notable ones like Casablanca don't, but that is because Humphrey Bogart had lived as a chess hustler before being an actor
https://www.reddit.com/r/StableDiffusion/comments/1f0hs79/you_can_train_handwriting_loras_with_flux/ with these secrets unlocked, it may be possible to create the one ball to rule them all
And some actors are also known chess aficionados. Some surprising ones, like Schwarzenegger, who is a major chess fan
tv chess be like https://youtu.be/v4cefbkif9E
GM Maurice Ashley' as a chess hustler, playing a game with Chuck Rhoades, partially recreating Maurice's "hustling" game where his opponent took two pieces at once. game in question: https://www.youtube.com/watch?v=U5vnpOp0U_g
For the fourth time, the Arnold Classic Brasil took place in Rio de Janeiro, and once again the sports and health bonanza attracted record numbers for the South American edition, with over 80 thousand visitors, 8500 competing athletes in 37 sports events. It would not be complete without Arnold Schwarzeneggerโs presence, who once again reaffirme...
I beat IBM Deep Blue about 5 times, but I do not like to brag ๐
this is one of my fave arnold moments because he wasn't just like "ARNOLD SMASH RUARH GET TO THE CHOPPA" or throw the board like a grenade and shout "GET DAHOOOWN" . just good wholesome arnie crushing his enemies and driving them
low poly balls
That sounds like some outdated ai >.>
I don't remember them well
It also sounds like a big fat joke

no i think it was an ai joke
Deep Blue was IBMs early Deep Learning platform... it competed in Jeopardy and famously beat Gary Kasparov eventually.
"deep blue so fat when it sits around the house, it sits AROUND the house" would be a fat joke
But your name isn't .... nah never mind LOLOLOL
jeopardy was watson
Deep Blue was a chess-playing expert system run on a unique purpose-built IBM supercomputer. It was the first computer to win a game, and the first to win a match, against a reigning world champion under regular time controls. Development began in 1985 at Carnegie Mellon University under the name ChipTest. It then moved to IBM, where it was firs...
Ah... ok...
kind of wonder how one of these open source quantized language models would do on jeop
Flux just generated this with one of the Xlabs loras... LOL
There are bazillions of articles on it of course, but I did write my own trilogy in which I compared the output of its logs with that of a modern engine. https://en.chessbase.com/post/komodo-8-deep-blue-revisited-part-one
When Garry Kasparov faced IBM's super computer Deep Blue, there was not a player alive who did not secretly dream of having their own private Deep Blue to consult on demand. That day has already come and passed as today's engines such as Komodo 8 can outperform the famous computer even on a smartphone. Here we take a look at the famous match and...
Tried to do a ball monster. Not really ball shaped though.
just a giant pile of shiny balls nothing rude about it
DId anyone else count the fingers throughout that? :>
"a spherical monster in the shape and style of the 80's toy Ballz"
I'll press that button
lol woman armor designs. do you figure they're just knocking around in that iron bucket while she's in battle? as if the nipple dimples are giving them space to float freely? girl would have such a bruised chest by the time she was done the day. consider the purpose of modern sports bras. woman warriors would have some sort of leather garmet that would contain her jiggly bits before putting on the plating. Like do you think a guy is going to cage his nuts in plating, in a way that they'd bounce around inside? anime armor design seirously serrrriously.
They're padded with ultra suede
ballz is just my lora tag. "Mad Balls" seems to be understood by sd3 and flux. less so in flux cause of the distillation, but i think it knows about them deep down
(or is it alcantara leather?)
then why even engineer nipple dimples?!
Because Tim Burton.
shock an awe?? .. hmm actually that might be a good reason
schumaker you mean. burton's didn't have the nips
Tim didn't want to do Batman anymore. So then we get Shitmaker's nipples.
I blame Tim for leaving the franchise.
Might be a fast H100 but it looks like they have only one LOL
Clooney is the best Bruce Wayne too, but not the best bat
neither of which i credit to "shitmaker" lol
Those are caps so she can breast feed between sword fights without taking off the armor
oh shiiii.. lol
i'm just a bigot this whole time
The BEST Batman will always be Animated Comic Batman... change my mind
Acting so bad it was brilliant
I appreciate all of them for various reasons, but Keaton is still the best all-around portrayal, in my opinion. He is the one who has best conveyed on-screen that Bruce Wayne was the costume, not Batman.
thye just threw all the dark themes right out the window when they did that show
he was the only part of the new flash that made it worth it to see
Keaton man taking down a kryptonian. Hot.
Otherwise, that movie was like 95% trainwreck.
Darnit, I was doing that from memory! I clearly had a boring childhood (no mad'ballz)
and bikini armour, well it all obviously has some sort of magic shield attached which surrounds the body lol
They're not Mad... they're just less joyful ๐
Disgruntled monsters
did you try clicking on it? That works depending
I see it now. ๐คฃ
flux > kling
Discord hides stuff if it's "sensitive", or if it was posted too long ago.
Apparently I'd gotten the toy name wrong "madballz"
Though "those round foamy monster things from the 80's would prob even work lol)
Who's model?
my Flux Hyper checkpoint
Ack, wrong window, so glad it didn't work the way I intended!!!
i did wonder lol
"eldritchpunk, madballz, spherical monsters as planets in space"
This implies that draton milk is a thing
Ball Rain.
lol that leg tuck was pretty special
I move away from the mic to inhale
It's Flux checkpoint merge day again today
Asian style dragon, impressive
ty

