>>103922297
>My gen didn't make the collage
Damn, I guess this means I need to do better.

Blessed thread of frenship

>>103922353
based carrot enjoyer

>>103922359
Yes

>>103922372
Fuck, anon that's insane.

Need Noobai for Touhoes.
Need Flux for inpaint comprehension
Need Gimp for everything else

>>103922327
Same

>>103922417
>>103922372
Sick

New default musubi tuner settings are shit, AVOID

>>103922667
>4chan, LDG, ComfyUI in the background
kek

>>103922649
Trying to get noob vpred working, but all my gens look like this. I'm using SwarmUI with a Comfy backend

Euler A, 4 CFG, 30 steps

>>103921404
I really want to know what kind of loras / prompt you use to achieve this look

>>103922782
Something went terribly wrong then
Try euler cfg+++, 40 steps, 0.5 cfg scale

>>103922816
Same thing basically, here's a catbox

https://files.catbox.moe/79gt24.png

>>103922854
>NoobAI-XL_-NAI-XL-_-_V-Pred-test_Version
that is the 1st version, get the vpred 1.0 model they released on in December, that is the fully trained model.

>>103922854
>NoobAI-XL_-NAI-XL-_-_V-Pred-test_Version
What the fuck is that model even

>>103922887
>>103922890
Oh shit I'll try that then

King in Yelough

>https://huggingface.co/Laxhar/noobai-XL-Vpred-experiments
>https://huggingface.co/Laxhar/noobai-XL-Vpred-experiments-2
Anyone have access?

Can you lads help out a 72 year old man please? I like to think I am fairly tech savvy for my age, I am probably the oldest person here afterall, but I am struggling with creating anything of note with comfyUI + noob model. I am new to AI image generation, but my wife loves animation from the 80s and I wanted to recreate our Korean style wedding, but everything I make looks like shite, just far too simple, unlike anything ive generated on websites. What do I need to do on top of installing the gui , importing noob model and running positive/negative prompts?

>>103922926
One more on epilipson

>>103922985
I assume you're already using negative and positive prompts?
Try different artistic medium (like traditional media, album cover, watercolor, oil painting, etc) and composition (dutch angle, backlighting, foreshortening, etc) prompts, it usually leads to far more interesting results. Just throw a bunch of schizo prompts into the machine and look at the results, sometimes it may surprise you.
Also maybe try playing around with different samplers and schedulers, for example i moved to DEIS beta for epsilon model and cfg+++ beta for vpred.

>>103922985
post wife

>>103922985
Depends on exactly what kind of outcome you're looking for. Do you just want to generate a generic anime bride and groom with a retro 80s art style? Or is there a very specific photograph that you want the AI to recreate?
Either way, if you're ultra-new to this AI stuff, I reccomend using Forge instead of ComfyUI. ComfyUI is a little bit more advanced because of the overwhelming "click and drag nodes and connect them with lines" interface.
I also recommend the CatTower checkpoint of NoobAI. I think that base NoobAI kinda sucks.

>>103923011
Vpred is alright, but I'm not seeing that extreme contrast and brightness

>>103923078
this. i cant help anon if i dont know what hes working with.

>>103922783
Man it's too much effort due to i2i and loras that only I have

>>103923066
I liked the paladin in Kingmaker

>>103923127
What do you do to get that painterly look?
Sorry for using a faggy term like "painterly" but I don't know what else to call it

>>103923146
I kneel

>>103923147
Not him but "watercolor" plus some artist style i assume, like Kazue Kato

>>103923046
Thanks.

>I assume you're already using negative and positive prompts?

Yes, although how well written they are, I don't know. I've attached an image of what I've been trying. I guess I expected more from the model, given the example images. Although its obviously a user error. Sillu question, but am I expected to add more than just this model? Its felt far too easy. All I've done is download the comfy windows portable and throw a model in a directory and hit go.

>>103923108
Ultimately there's a specific photo I'd like to recreate, but for now just something like yours would be lovely. I will look at Forge, thank you. And I quite like the nodes interface, it reminds me of a software I used a very long time ago called Houdini.

>>103923147
Here's the catbox for that one
https://files.catbox.moe/ehue24.png

>>103923181
Part of it appears to be your latent image size. NoobAI along with all XL models are meant to be used with at least 1024x1024. In addition, for Noob, anon seems to prefer euler cfg ++ for the sampler and beta for the scheduler. In the future, once your workflows start getting more complex, a better method to help anon diagnose is to upload your generation to catbox.moe and post the link. That way we can see the metadata.

Read the model page on Civitai as well. They tell you "how best to prompt" which serves as a good starting point.

>>103923181
I think for vpred you need to use a node called ModelSamplingDiscrete.

And use something like euler_cfg_pp and put the cfg at like 1.0-1.5. And you can append/prepend some quality tags like (masterpiece, best quality, very aesthetic).

>>103923181
You're using a vPred model. They need a bit more tard wrangling. Try 40 steps and cfg at 4, or euler cfg++ sampler and cfg at 0.5. Although i'm not sure why your output comes out this way, it shouldn't be this low res and blurry.
Try finding finished, basic noobai workflows for comfy, that stuff is all over internet, just google it.
>Ultimately there's a specific photo I'd like to recreate
Then you will have to get into image2image and controlnet once you solve the generation issues.

>>103923271
>I think for vpred you need to use a node called ModelSamplingDiscrete.
that was only for the beta test version because the author forgot to mark a flag.

ComfyUI repo turned two years old today ~~<3~~

>>103923181
You need to tweak your sampler settings. Noob uses different ones from default. See the quick guide from the author for those settings and how to prompt effectively and etc. about the model.
https://civitai.com/articles/8962/noobai-xl-quick-guide
I know from what you posted the CFG is too high. The above guide suggests a CFG of 3 to 5.5
>>103923271
You don't need it anymore, ComfyUI automatically will detect it for you.

>>103922985
>72 year old man
Does this shit blow your mind?

>>103923181
A few of the words in the prompt might be confusing the AI.
Terms like "paint bleeding effect" "wet media" "traditional ink splatter" could put a big wet inky blood splatter onto your wife.
Start simple
>man, woman, standing, korean clothing
then gradually start adding more words, and keeping note of which words inadvertently have a detrimental effect on the image.

>>103923304
I don't care if it's worse than it was last year

>>103922947
No but speaking of which, did anyone every try any of the illustrious vpred stuff that someone posted the link of? This seems to be the latest checkpoint from a couple of days ago
https://huggingface.co/AngelBottomless/Illustrious-v0.1-vpred-test/blob/main/20250101_convert_H100bs256rate2e_2debiasoffset/20250101_convert_H100bs256rate2e_2debiasoffset-step00056350.safetensors

>>103923320
Not him, but i've made heavily edited and stylized posters for all my older family members over the last year, and they were all speechless about it.

>>103923181
since noob is trained with images from danbooru you can use that site along with https://danbooru.donmai.us/related_tag to find decent tokens to prompt with
anything with at least 700 or so images shouldn't be too difficult to illicit out of the model

>>103923181
sometimes less is more
>masterpiece, best quality, 1boy, 1girl,
>80s, retro artstyle,
>couple, standing,
>man, woman,
>hanbok, traditional korean architecture, wedding
Overwhelm the AI with a bunch of other words and phrases, and you might get some weird unwanted results

>>103923078
My wife is far too old for you to be interested in. But in the 90s, there was a Korean actress that every one was crazy for, called Kim Hee-sun. I think she look(ed) like her. Of course, she wouldn't agree, and nor would you probably, but I'm fine with that. She is in a truly terrible film called The Myth, with Jackie Chan, that I would suggest to watch if for nothing else than a laugh with a drink or two.


Thank you all, I appreciate the help. I'll return the favour by posting increasingly less shit AI images of a wedding you don't care about.

>>103923320
Yes, it does, but I'm glad I got to live my life during the age of reality. Not that you asked, but I think the future is going to be one based entirely on artifical narratives and fakery, and I'm happy to have lived in a comparatively simpler time.

>>103923397
> I'm glad I got to live my life during the age of reality
What the hell is this guy talking abou -
>the future is going to be one based entirely on artifical narratives and fakery
...shit, he's right.

>>103923397
Based grandpa thank you anon

>>103923336
just revert to a previous commit tho

Finally decided to try out Illustrious after sitting on Pony for ages, might make the switch. Illustrious has pretty impressive artist and character recognition and the prompt adherence is definitely better. Also feels like it has better 'aesthetic sense' though it still has the usual AI problem of perspective issues.

>>103923453
would it still be 2 years old?

>>103923467
I haven't tested out Illustrious yet, but I upgraded from Pony to CatTower and I'm so glad I did
So far it has recognized every single artist name I've typed into it
It feels like artist loras are a thing of the past

Had my Rapunzel lora drafted and was about to upload but someone uploaded a synthetic slop (I can tell it's been trained off base PonyXL gens) Rapunzel lora literally minutes ago. Not sure what to do, I don't want to seem like an asshole

supreme gens ITT

>>103923564
I remember hearing an expression about this type of situation
"two cakes is better than one"
even if someone else baked a cake, go ahead and serve your cake anyway, even if it's redundant

>>103923567
agreed

>>103922478
cute, stylish

RIP David Lynch
your art, and the way that you used young girls and women in your art, is a large inspiration to me

>>103922720
You just know when the black goo leaves the card it's dead. Rip

>>103923564
>Not sure what to do, I don't want to seem like an asshole
If you can do better then you're helping the community.

>>103923659
Twin Peaks had the prettiest women. Lynchian is also a great term.

>>103923374
Looks cool, sorta reminds me of King Death: Monster.

How do I use a controlnet?

>>103923842
Fucken sick anon

>>103923540
How does it differ from noob?

>>103923976
thanks anon
>>103923914
Depends on what UI you are using and which controlnet you want to use.

dude
what if
like
godzilla...
...fought king kong
like
who would win, dude?

>>103924242
I now just noticed the perspective on the arch is completely fucked

>>103924178
I always get weird shit with Noob
Left is Noob, right is CatTower
Same prompt, same settings
Noob generates le artsy fartsy stylized images, CatTower produces images similar to what I'd expect to get

>>103924242
could nitpick but it's cute

Thought I would post the final product of trying for a few days on and off to get as close as I wanted the gen to be without altering my workflow further.

First proper gen I've taken time with, next time I post I'll have done so with a Lora based on the Armored Core: 3 design. Might do more Loras for each leg type. If anyone has suggestions on ways to make my slop look better I would like some advice please (besides inpainting which is the next thing I'll be trying out after I do the Lora).

Noob keeps giving me shit that looks like an alpha mask instead of an actual image
am I just a fucking retard?
the answer is obviously yes, but like, specifically how would I un-retard myself here?

peekaboo
i
see
you

>>103923375
Gonna try it later

mochi unfortunately does not know who david lynch is (or maybe that is a good thing because he is safe from low effort slop)

>>103923702
i can tell by your gen that Officer Diane was a personal favorite of yours

>>103924314
>specifically how would I un-retard myself here?
steal someone elses comfyui workflow
>im not using comfyui
try to setup and use comfyui
>im using comfyui and it didnt help
go find and use some random meme webui with a support for noob made by an esl or install gentoo

>>103924335
is this mayli

trying hunyuan prompting in an itemised fashion (Scene: outdoors, empty plaza, Subject1: male, 35, skateboarding, SUbject2: female, sitting, staring at subject1, etc) and it works just as good as natural language
is that llama being clever or the model

>>103924285
>>103924314
your noob vpred baseline should look like this.
"1girl, masterpiece"; no negative.
euler a cfg ++, 1.5 CFG, sgm uniform, 28 steps, rescale cfg 0.7, model sampling discrete "v_prediction" & "znsr" true (for comfy), I also had PAG activated at 1.0 strength because I am a muppet vut with that vague prompt it really stabilizes the image.

The pope.

>>103924681
I tried your settings and got this
I appreciate your help, but I'm just too stupid to figure out Noob. I've grown really accustomed to CatTower, so I think I'll stick with it

>>103924725
cool gen actually lol
mess around with it until you get something good - it is worth it. you on comfy or reforge/forge?

>>103924285
left is leagues more sovlful desu

Anime sloppin' is pretty chill, I get why anons do it.

>>103923240
thanks bro
it'll be a while before i can make stuff with as much soul as yours, but i appreciate the help

do you like my image

>>103924951
I feel like I've already seen it a thousand times before, and this isn't something I would ever search for

the quiet girl at school has been bullied for the last time

>>103924964
tanuki bros , it's over

>>103924951
She cute

>>103924951
Should be a .mp4 if you really want (you)'s

h

excuse my retardness but whats the best bang/buck GPU in year of ur lord 2025?
amd included

>>103925634
what for
I can guess considering what thread we're in, but it's important

>>103925638
SD performance of course
gaying is a plus
limit of about 5-600bux

>>103925634
get a 4090 or 5090

>>103925634
>best bang/buck GPU
Ironically it's going to be a 4090, the most expensive GPU

Curious to see how much better a 5090 will be though

>>103925634
I wouldn't recommend it but it's the 4060ti 16gb.
VRAM is king, 16gb is more than enough for image gen and it's viable for current vidgen

>>103925634
https://www.videocardbenchmark.net/high_end_gpus.html

>>103925634
dumbass img

>>103925682
>>103925655
>4.090
anxiously waiting for the price drop

>>103925920
Ye should drop once 5090 gets released

https://blog.comfy.org/p/comfyui-turns-2-a-journey-and-call

Huh

>>103925935
>Ye should drop once 5090 gets released
GPU prices never go down.

>>103925954
SDXL ain't even 2 years old yet

you are jerking off to the images created by an underage model?

>>103925972
2 out of 10

>>103925996
what happens after 10 years

>>103925954
Happy birthday to ComfyUI! The best UI!

>>103926001
She's legal

>>103926002
best backend but far from the best UI.

>>103925954
Happy birthday to ComfyUI! The best UI with the best mascot!!

>>103926026
I don't show up to your birthday and tell you all of your faults. Maybe have a little RESPECT.

Do you think I could play AoE DE while genning a big batch of images (the last node uses SD ultimate upscale to 4k)? I have a 4090 btw

>>103926052
It's going to tip the vram over

>>103926068
looks like starfield, empty and boring

>>103925998
based antisemite

>>103926119
>no blowjob hole in her helmet
for SHAME

Comfy must have fixed some shit recently because cosmos img2vid is kind of...good?

>>103926141
Lets see the results sir

>>103926204
comfy has been doing damage control all day since yesterday. HyVid is the only thing worth a damn

>>103925954
Can someone address his posture, he's basically in the fetal position in the interview

>>103926141
if it was really good, people would post a lot of examples already
>>103926209
this, HyVid won't be dethroned soon, it's good and completly uncensored, will be hard to beat that

>>103926301
You're falling prey to brand loyalty for a free product. The fact nobody is posting here because nobody knows it can be run in comfy yet and there are maybe 5 people in this thread with a 24gb card.

>>103926313
>maybe 5 people in this thread with a 24gb card.
if we follow your logic, Hunyuan would never be hyped because 5 people in this thread has a 24gb card, yet it's not what happened at all
>You're falling prey to brand loyalty for a free product.
the fuck is he talking about? I'm someone who jumps ship as soon as I find something better, I was a big fan of SAI for more than a year and a half and then when Flux appeared I pretend SAI never existed in the first place, you can't make someone as unloyal as me

>>103926326
Have you tried img2vid with cosmos in comfy yet?

>>103926330
I've seen some i2v results on this thread, they don't look that good, and Comfy also added an example of his own with his avatar, just a sitting anime girl with a camera panning left and right, did you try the i2v model anon? if yes then what are you waiting for to show us how good that is?

>>103926326
>jumps ship
these are tools, there's no "jumps ship" here, they all serve their specific purposes

>>103926338
no they don't, no one use SD1.4 nowdays because it's deprecated, and no one use windows 95 too because it's also deprecated, deprecated software exist and I'm tired to pretend it's not

>>103926350
stop being silly, there's a reason people still use sdxl. sdxl can never compete with flux's prompt understanding and vae but like wise flux can not be finetuned like how sdxl can. for me flux cannot replace sdxl for me and wise verse. i have to use them both for specific needs

>>103926336
>I've seen some i2v results on this thread
Yeah. They've all been from me. I said it wasn't so good, but them implementation seems better now. Much less unstable. You can even choose the end frame.

>>103926366
>there's a reason people still use sdxl
because there's a lot of vramlets? if everyone had a 5090 card SDXL would die the next day

>>103926373
I see, may I see some of your new outputs then?

>>103926376
no because, like i said, sdxl can be finetuneable. that is extremely important to me. you cannot finetune knowledge and style into flux, base flux is limiting and loras can only do so much

>>103926393
>you cannot finetune knowledge and style into flux
I really wonder why no one tried to finetune the undistilled version of flux yet

>>103926403
there is some furry group that's trying to finetune the undistilled 8b version, didn't look into it tho

>>103926409
>there is some furry group that's trying to finetune the undistilled 8b version
why is it always the furry that try to save a model? kek

>>103926141
>cosmos img2vid is kind of...good?
the "good" in question: >>103926429

>>103926429
I swear to god if Nvdia will train their autonomous machines with this video model, I'll never use an autonomous taxi ever in my life lol

>>103926429
That's terrible... is this the 14b model?

>>103926429
Where he goin?

>>103926471
he's leaving to fix his fucking spaggheti software kek

>>103926326
>Hunyuan would never be hyped because 5 people in this thread has a 24gb card
I knew there was no way I could run it yet I hyped the shit out of it because what I saw made me audibly say "wow". Now I stay quiet because I can't run it therefore I have little to contribute except this reply.

>>103926429
they turned my xis into spaghetti

>>103926478
>I hyped the shit out of it because what I saw made me audibly say "wow"
so a model gets hyped because it's good, and it gets not when it produces garbage like this >>103926429
that's one thing I like about the image/video AI ecosystem, people know what's good and hype that, and when something stinks, they also say so and won't give the model any strength, as it should

>>103924816
>you on comfy or reforge/forge?
look at his filename btw

>>103926489
>they turned my xis into spaghetti
which is ironic because he's the spaghetti man, now all there's left is Will Smith to eat it, unironically I'd love to see that
>A sitting man turns into a Spaghetti, and then Will Smith comes out to eat it

Where did the kinosovl go?

>>103926429
LTX i2v could do something better than that, that's humiliating because LTX is just a small 2b model

>>103926558
don't tell me that Cosmos is like LTX and you need to write a bible to get something simple, I guess that's another win for Hunyuan, because that chink model works well on small and simple prompts, like Flux too

>>103926558
Why would we blame ourselves? That's absurd. It's always the computers fault and never the operators.

>>103926568
That's assuming it's the operator's fault in the first place, we don't know what his prompt is, and desu if he wrote something like "A man is standing up" and the model produces something like this, then I'll stay in my hunyuan i2v waiting room thank you very much.

i might give swarm another try

>>103926601
lmao that's fucking ASS, Tencent save us!

>>103926601
what are you prompting? this looks like ass

>>103926601
I think Nvdia should keep making just GPUs, that's the only thing they're good at lol

>>103926401
>prompt a god
This kind of thing bothers me.

Depicting a 'god' ought be a serious thing, if it's worth anything. That's not because I like moralizing about """beliefs""", but by calling something a god you declare it a serious thing, unless you're making a joke. And there aren't many jokes left to make about these gods nobody believes in.

If we're not joking, then to undertake to depict it with utmost seriousness, even let's say experimentally with AI, which is fine—we should at least contemplate how to represent this something which we don't see. That's the creative act here, that's where we have choice. What are we going to tell the model to draw? What visuals do we invoke?

To say "god" in the prompt is to say, "show me how a god of this variety is drawn conventionally". Which is to say, I am satisfied with how gods have hitherto been drawn, even in television animes and RPG video games and so on. (Which we can't really be.) And I am choosing not to use the executive power which belongs to me as the artist; I defer to convention. If I'm making background art assets for some conventional scene, where being conventional is the point, maybe that excuses it. But if the image is for its own sake...

To have this attitude while ostensibly being interested in making a picture of a god is really not serious. No one should be this blasé about making images of gods. This tvtropes LotR nerd mindset makes piles and piles of worthless fantasy slop for throwaway paperbacks, forgettable under-$30 steam games, unwatchable television, etc. Nobody believes in these "gods" and nobody has thought of anything interesting to do with them. It's repulsive, aesthetically.

In fairness I should admit that being quite 'serious' about gods usually produces very bad art of a different kind.

^ chatbot

>>103926636
what llm you used on that one anon?

>>103926620
"An interview going wrong as reality fails. The person being interviewed tries to stand up but is caught in a reality glitch and deforms horribly"

>>103926642
that's bullshit, but I believe that

>>103926601
2023 called, they want their i2v Gen1 model back

Police precincts get calls like these at least twice a week in certain states.

>>103926601
>>103926429
What's funny is that Nvdia worked hard with their guardrails as if this model is top tier and would be good enough to fool anyone kek

>>103926640
>>103926641
First time seeing an effortpost? You know back in my day we wrote all the text walls by hand. Some of us still do.

Troubling news: The Hyvid github repo seems to have removed all mention of img2vid from their open source roadmap.

>>103926705
>effortpost
yapping isn't effort, if you can't explain an idea succintely and make it easy for your fellow readers it means you're just a retard
>>103926708
kek you made me scared for a second anon :')
https://github.com/Tencent/HunyuanVideo?tab=readme-ov-file#-open-source-plan

>>103926708
you sneaky anon you

>>103926708
we shouldn't have made fun of that effort poster talking about god...

>>103926708
Funny misinformation

>>103926705
Wait, you're right. Chatbots can't do the schizo triple quotes,

>>103926743
what's up with the glitches at the right of the screen? I feel sorry for Comfy because he worked hard to make Cosmos compatible to his software, but at the end this model is complete ass, he definitely wasted his time

>>103926762
i am pretty sure he prompted for that >>103926642

>>103926771
I did not.

The image is a video of a man sitting in a modern, blue upholstered chair with a black metal frame. He has medium-length, light brown hair and is wearing a light gray t-shirt, gray sweatpants, and gray hiking boots. The man is smiling and has a microphone attached to a boom arm resting on his left shoulder, indicating he is likely engaged in a live stream or podcast. The background is a plain, off-white wall, and the floor is covered in a speckled, gray carpet. To the right of the chair, there is a small, orange round side table with a stack of books and a few other items on it. The table has a black metal base and a black top with a small, colorful light on it. The overall setting appears to be a casual, indoor environment, possibly a home office or a studio. The man's relaxed posture and the casual attire suggest a comfortable and informal setting.

>>103926716
I did explain it succinctly. How long does it take you to read those short few paragraphs?

"Uhhh give me the main idea as a bullet point" No. You're a lazy retarded sub-literate nigger.

>>103926806
see, you made your idea clear in 2 lines, more of that please

>>103926778
imagine making a quality wall of text >>103926778 just to get something as bad as that >>103926743
Hunyuan will remain the king for at least the remaining of 2025, mark my words

>>103926923
>Hunyuan will remain the king for at least the remaining of 2025
if they don't release the i2v model no they won't, this shit is important and I hope the chinks won't betray us (can't believe I live in an era where I have to trust the chinks to be saved but here we are)

Oh, you thought you found a way to prompt found footage? The abandoned house has lights now. Fuck you.

https://www.youtube.com/watch?v=sQObRyR01WA
https://www.reddit.com/r/StableDiffusion/comments/1i2nmas/comment/m7fuxpl/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button
>The full 4'19 video is fully AI, mainly using a finetune of Hunyuan video model
This is really impressive wtf

https://vchitect.github.io/RepVid-Webpage/
that looks like a tool to improve the video output like "Enhance-A-Video" I guess?

The only NAI tool I can't seem to figure out how to replicate is the sketch tool. All the small things like the structural lines and random warmup shapes. Any idea how I should go about recreating this kinda thing locally?

>>103927090
Why doesn't anyone upscale each frame and throw them through a nice anime t2i model?

https://youtu.be/ihDbAUh0LXk?t=42
is this the most useless paper ever?

>>103927166
>replicate is the sketch tool
is that a tool that turns you gen into a sketch or turns your sketches into gens?
>>103927170
it would probably make each frame look pretty inconsistent

>>103927170
>Using synthetic data to train an AI model
when will they learn? what's wrong with training a model with actual good anime screen, like Hibike Euphonium for example
https://youtu.be/Cwe8ht7ACBw?t=249

For those who want a sort of block swap on comfy core, did you try the --reserve-vram flag? that shit is cool when you want to specify how much memory you want to put into the ram instead

>>103927175
Turns a gen into a sketch,
input is on the left, result is on the right:

>>103927207
maybe you could try using a canny controlnet and just prompt for a sketch. there's the lineart and softedge controlnets but those aren't very sketch like

What the fuck, why did my lora self-destruct all of a sudden? And that's while using Prodigy and scale weight norms, meant to prevent that.

>>103927236
Only ashes remain.
(You can roughly measure burn-out by enabling compression on the lora file).

>>103927244
Nta but burnout?

>>103927254
Not sure what you're asking about. A burned-out lora apparently has many weights clipped to zero or max value, thus reducing entropy of the file and making it compressible, unlike normal models.

>https://rentry.org/59xed3
Should revive this resource.

>>103927289
holy bible!

>>103927229
Thank you! I had tried other ControlNets but I always assumed it'd have been Softedge or Scribble but this actually worked quite well. It's not exactly where I want it but this is plenty to start testing around with.

>>103927207
Forge's maker created a controlnet to produce in-between stages of a given image being drawn, that triggered a lot of inkcels back in the days.

>>103927289
go for it

>>103927326
Oh? Any name or something I can look into, maybe there's something that'll help me out.

>>103927368
https://github.com/lllyasviel/Paints-UNDO
I'm not sure it's a controlnet now, i haven't used it myself.

We aren't getting img2vid

>>103927507
the chink promised us we'll get the model recently, don't doom :'(
https://github.com/Tencent/HunyuanVideo/issues/180#issuecomment-2592674405
>We are working hard on developing a text-to-video model and believe it will be completed soon. However, the open-sourcing process involves some steps that are beyond our control, so we cannot estimate the exact timeline. We believe that it will be available to everyone soon.

>>103927520
That sounds like later than January now.

>>103927527
I don't care if it's february if at the end the model is perfectly trained lik the t2v model

Did you get a better prompt adherance by changing clip_l to one of its finetune on hunyuan?

>>103927544
It's barely perceptible.

>>103927289
Do it. I wanna know about restarts and optimal number of cycles

>>103927266
Bet you didn't know that.

>>103927537
There is a problem and they're too afraid to admit it.

>>103927602
they're so fucking retarded, they should've released both the t2v and the i2v model at the same time when no one knew they even existed, now every eyes are on them for the next release, no one was looking at them before

Oh wow he is finally updating this. It's funny how great supermerger used to be.

I think less of gradio UI users. Even lower of people who purposely hold off on upgrades to stick to a gradio UI.

>>103928076
I think less of comfyusers who support copyfaggotry instead of spreading leaks.

How do I interpret these benchmark results? >>103928110

>>103928133
more it/s means faster gens. pay attention to what model is being used, bigger models are usually slower

How is it possible to get speech that isn't random nonsense in manga/comic panels? (There are many other problems in this image I know).

>>103928358
you manually edit them in yourself

>>103928358
Legible text is only really for larger models and you don't have that much granular control over the output.

Thank you, lads.

>>103928488
To add; I definitely need to improve, it feels very 'A.I' to me, still, but I like the composition, I just don't understand how I'm always trying to stop it from using cherry blossoms and japanese motifs in the background. But a long way from the shite it was yesterday.

>>103928519
give me your prompt/workflow, i'll have a quick look and see if i can help

>>103928519
Show me what you consider shite then.

>>103928531
you are not entitled to a catbox

>>103928546
he is entitled to my help

Guys is there some resource where I could find complete high quality art packs?

>>103928697
Sort danbooru by score.
Are you looking to build a regulation class dataset?

>>103928737
I mean classic art, paintings and such.

>Are you looking to build a regulation class dataset?
No. I just don't have a good quality art pack for the classics.

>>103928519
As it was already stated in this thread, play around with prompts that impact composition or camera. Maybe research anime artist styles you'd want to imitate the most.

Speaking about regclasses, are they even required/usable for training XL? I just don't seeany particular effect when using them in XL training. SD1, on the other hand, makes it very noticeable, as artstyle of the regclass influences the training a lot.

>>103928358
Paint?

I'm still trying to figure out what a 5090 can do that a 4090 / 3090 cannot do

>>103929329
the 5090 has 32gb of vram, so you'll be able to make a full finetune of flux, make loras of hunyuan with bigger resolutions and more frames, and you'll be able to run hunyuan at higher resolutions too

>>103929341
Thanks! Glad we got nvidia employees monitoring those threads to reply in timely manner.

>>103929364
the fuck you talk about nigger? I fucking hate nvdia but it's not like we have a fucking choice, the day AMD release a 48gb card I'll jump ship instantly

Upscaling (including tiled) resulting in crash for sdxl vaes still a problem for everyone, or just me?

>>103929341
Next thing you'll say it won't be coub-tier with six seconds anymore.

>>103929414
You aren't allowed to criticize videogen here.

>>103926429
its the fucking trash it goes with all the other hyped censored garbage.

What is currently best boobai/illustrious based mix?

>>103929443
The one that obeys the negatives (I'd be surprised to find one that does).

>>103929414
yeah, I can't really disagree with that, the days we'll instantly have a high quality 15 mn video with sounds from a single prompt sounds like a far away dream, I hope it'll happen in our lifetime lol

>>103929443
have you considered just using a non-mix model? vpred10 or one of experimentals work fine if you aren't a skillet

>takes effort to get basic results
>best model

Might as well suggest to train the best checkpoint from scratch if you're not a skillet.

>>103929514
>Might as well suggest to train the best checkpoint from scratch if you're not a skillet.
so petty, I love it

>>103929514
so much this, i switched back to pony and couldn't be happier. i gen sexiest hentais

https://github.com/kijai/ComfyUI-HunyuanVideoWrapper/issues/19#issuecomment-2591135788
god bless the turkish god, he fixed one of my issues with cuda and triton

>>103929329
Think "Moore's Law".

>>103929438
yeah, I really hope Tencent will release the i2v version of hunyuan, they're our only hope

>>103929329
Since 32GB is easily available to consumers, expect all the new research models target that.

>>103929603
Well, you aren't fucking poor, are you? You ARE buying the newest nVidia product, right?

>>103929597
Fuck this.

>>103929611
Don't have to worry about being poor when you have no chance to beat the bots

>>103929597
I mean, 32gb isn't that big, if you want to run hunyuan bf16 with its biggest settings (1280x720x129f) it'll probably won't be enough

Size matters. Bigger is better.

>>103929636
>Bigger is better
yep, it's true, it's even called the "Scaling law"

>>103929636
Enjoy your 200gb videogame installation sizes.

>>103929661
Seems like they are hitting a wall

>>103929712
for the llm side yeah they're hitting a wall but I mean, they're hitting a wall at fucking 500b+ parameters, we are far from that with our 12b flux or our 13b hunyuan lol

>>103929674
Welcome to the era of big data.

"It's all so tiresome"

Never sure which thread to post gens in, this seems like moreso the discussion one.

>>103929950
fat asuka sex

synthetic and non-heteronormative

>>103929950
ArchitectGODS...

>>103929950
yeah, /sdg/ is the place to post your picture, /ldg/ is more about talking and discussing about the news and shit (to be fair nothing's happening recently so...)

>>103929950
I'm never getting any reactions to my questions there, it's basically a shitpost thread of a single schizo, so why bother.

>>103930047
why do you stay here then?

>>103930047
Isn't it the faster thread...? I don't post here often, but while I'm looking at lmg and such, I always notice this one is higher on the list.

>>103930047
>it's basically a shitpost thread of a single schizo
Literally me
But don't worry i won't fuck up the blessed bread

>>103929628
preach, I have 48 and it feels like the bare minimum

https://xcancel.com/freepik/status/1880251686360670283#m
>Character consistency
how did they achieve that with hunyuan?

>>103923375
hey this model seems pretty decent

>>103930149
>Character consistency
For a whole second! OMG!

>>103930177
this

>>103922947
Why the fuck is a weeb model gated? I though only evil retarded corpos do that.

>>103930263
Can ypu find it on some non-google approved web file server for free?

>>103930102
thank you based schizo anon

fresh

>>103930326
>>103930326
>>103930326

>>103926002
>>103926033
samefagging tranny freak
return to your avatarworm shithole

>>103930390
TRANNIES ON MY MIND CAN'T.... STOP.... THINKING.... ABOUT TRANNIES!!!!!!!!!!!!!!!!

>>103930027
Basically, /sdg/ is your playground and /ldg/ is the library. Though let's be real, neither one seems to have much "recent news." Guess it's just the usual cycle of doomposting and waiting for the next shiny thing.