File: PixArt_00001_.png (3.41 MB, 3328x1152)

3.41 MB PNG

PixArt workflow
https://pastebin.com/tnQKhTLT

SD3 PixArt workflow
https://pastebin.com/AJ9uJrmM

File: 116775249263549015-SD.png (2.34 MB, 1352x960)

2.34 MB PNG

Fresh bred............

File: 116775249263549026-SD.png (1 MB, 536x968)

1 MB PNG

>>101507976
1.5 still decent

what UI should i use?

>no bred blessings

File: 1701157107814667.jpg (763 KB, 1024x1024)

763 KB JPG

New animu pixart model dropped. I think I messed up the workflow lol

https://civitai.com/models/590464/anime-sigma-768px

https://civitai.com/models/583615/dalcefopixartsigma-lucida

>B-Lora
https://b-lora.github.io/B-LoRA/
https://github.com/yardenfren1996/B-LoRA
The savior of vramlets?

File: teaser_blora.png (1.65 MB, 2086x911)

1.65 MB PNG

>>101509375
https://github.com/liusida/ComfyUI-B-LoRA
>Can apply Style or Content, or both.
>Much smaller model files. (~100M for SDXL B-LoRAs)
>One B-LoRA only needs one image as training dataset and 15 minutes to train. (on a single RTX 4090)

File: 39lVwj0.png (120 KB, 1147x673)

120 KB PNG

new to this and going to only coom proomt.
are there any must have extension i should be aware about?

slow sunday

>>101509714
you can try Automatic CFG, it gives you one less thing to worry about (you can use CFG 8 on everything).

File: help.gif (1.56 MB, 400x267)

1.56 MB GIF

>>101510523
i think i downloaded some broken extension or something because it won't load anything on the extension tab and won't let me generate anything( when i click on generate it just idles for some time and then it lights up again)

File: Gemini_Generated_Image_x9(...).jpg (376 KB, 1536x1536)

376 KB JPG

>>101509375
>magic numbers
Of course!
>Checking if your connection is secure...
Of course not, it's Cloudflare...
>

>>101509060
ty will try soon!

>>101510666
Go to the extensions folder and delete the bad one then relaunch

I think

>sdxl base, 1024x1024, 20 steps, dmm++2m sde
>gen time: 100 seconds
Bros, Is this normal? I have gtx1660ti (6gb vram). On reforge github page they show that with the same settings on rtx3070ti (8gb vram) gen takes like 14-20 seconds. I know mine should be slower, but 6x, 7x slower? I'm using linux, on windows I had even worse results.

>>101511495
Try like 15 steps and make sure you have xformers

But yeah that's normal as far as I know

>>101511495
Install cuda, and in your webui-user.bat, open it with notepad and put in --xformers --no-half-vae in the set command line args. But yes, that's pretty normal for older cards. On a 3060 12gb it takes me 30-40s to gen an image w/ adetailer applied.

>>101507621
Hey, I wanna generate some images for a NSFW game. Nothing crazy, just a bunch of one or two girls doing some stuff. What’s the easiest/cheapest way to do that outside of picking up a 3070 of my own?

>>101511583
>>>/aco/sdg
>>>/h/hdg
^ nsfw is that way

>>101510219
How it usually is

File: Screenshot 2024-07-21 at (...).png (29 KB, 418x170)

29 KB PNG

I have a macbook pro with these specs.
Would it be possible to run SD on my machine without it sucking?

>>101513247
>ancient laptop cpu
>integrated intel gpu
>Would it be possible to run SD on my machine without it sucking?
kek

>>101513852
Decent anime screencap.

File: Sigma_06370_.png (1.58 MB, 1280x768)

1.58 MB PNG

>>101513866
ty

hibernation mode

File: Sigma_06306_.png (1.46 MB, 832x1216)

1.46 MB PNG

>>101514304
bigma waiting room

File: Sigma_06265_.png (1 MB, 1280x768)

1 MB PNG

Good night anon

What exactly do I need to do to get images that aren't obviously flawed ai rendering like limbs and fingers? How can I fix up images I like that have small issues? I really don't get it, the generators never quite get exactly what I want and I have artistic visions but I can't draw for shit and need this to help my game dev.

>>101507824
These workflows are messy

official pixart bigma and lumina 2 waiting room

>>101515598
inpainting. You might be able to get away with an ADetailer at a higher denoise value.

Why did imagegen die? All imagen threads on all boards are dead or completely engulfed in schizo bickering. Did SD3 failure really kill the momentum?

>>101517751
>Did SD3 failure really kill the momentum?
Yeah, and the fact that there's no serious alternative base model that would fill the gap, we're just waiting for something better than SDXL base (with a good licence) and it ain't happening

>>101517751
snoozing till bigma or another model that can replace sdxl (with a good licence) like the anon above said

File: Screenshot_20240722_070838.png (65 KB, 326x658)

65 KB PNG

>>101517771
Sigma base and Cascase base are both "better" than SDXL base. So far down the road you forgot where you came from

>>101518135
Also me, but why is Hunyuan DiT so much worse?

>>101515425
SD1.5, SDXL, and maybe Sigma HYPE

Sigma pretty good

>>101508360
1.5 is great

so after the disaster of SD2 and beyond I assume stability AI is outright dead or best case wont release unlobotomized models ever again. Is there anything to be excited about in the future of iamge generation? SDXL is really far behind dalle3 and takes so much work to get a passable result.

>>101519039
For most of the internet's needs, dalle3 is really far behind SDXL/Pony. It also doesn't really beat PA-Sigma or others for prompt adherence.

But if you still have some reason to use DALLE3 like perhaps that you want a ton of unasked random shit added because your smartphone is annoying to type on, just continue.

>>101519417
dalle3 is far better at human anatomy than stable diffusion. if you prompt it right it generates perfect images of complex subjects and poses about 1/4 of the time. to get the same effect ons table diffusion requires agonizing work. try generating soles of a woman's feet for example. SD never does it reliably.

>>101519452
> try generating soles of a woman's feet for example. SD never does it reliably.
well yes most use pony or sdxl for that until some of the newer models are also better nsfw trained

>>101519417
Dalle/Bing, as a service, is certainly far behind SD and other competition, but that's because it's heavily censored and blocks pretty much anything that isn't pictures of cute cats now.
Its comprehension and prompt understanding however are so far above sdshit it's not even fucking funny. You should have played with it during first few weeks.

>>101519712
it was better before it was censored and (I think) probably compute resource "optimized"... limited

It was not really much better in prompt comprehension than Sigma tho. Maybe it was even worse if you assign penalties for random unprompted shit appearing.

>>101519810
>Maybe it was even worse if you assign penalties for random unprompted shit appearing.
Seen some people say this but I don't know, it wasn't any bad or annoying, the "unprompted shit" was usually relevant to the prompt and didn't really feel at odds with it. I mean, you don't really want your gens to ONLY have shit you prompted, right?
I haven't really played around with sigma much so I guess I'll try it.
>it was better before it was censored and (I think) probably compute resource "optimized"... limited
Yeah it's just bad now. Ugly outputs, censored to hell.

>>101519885
Too much random stuff showed up when people compared prompts. I guess if the NSFW training and so on had been better it could have been excused as a quirk but it's still counting against prompt adherence.

Sigma has pretty good prompt adherence but the issue with not enough NSFW and other content is actually the same so far. People are trying to train it in rather than out tho.

>>101519712
I was using it when it first came out. So sad that I was only lightly using it and didnt make goated prompts back then. Even still the smut it generated was insane.

Is that one thing better than that other thing? Personally I don't think so

File: 00_3.jpg (391 KB, 1400x1400)

391 KB JPG

>>101519791
Nice

>>101508686
I'm literally crying

>>101517751
last i checked /degen/ is still going. anons need to JO is strong.

>>101517751
Because who generates variations of the same picture every day for months? We're really at the point where the images people post are just repeats and boring. Normal people got bored and moved on and image AI, as it is, is now the new normal and no longer that interesting.

>>101520860
>We're really at the point where the images people post are just repeats and boring
We reached that literally 6 months after SD 1.0 bro. All that's happened since then is people making new and finetuned models to generate either anime chicks or porn. That's literally it.

>>101520759
This, actually.
The novelty wore off long ago unless you are a newfag. In the end, you're either interested in the tech or the coom.

>>101520976
>That's literally it.
New tunes are one of many things that have developed since then are you trolling? Kek.

>>101521230
Reread the post

>>101511495
>6gb vram
Most likely you are running out of VRAM and it's going into shared GPU RAM, which is slower by roughly a factor of 8x.
I used to run SD on a 1060 6gb and would run into this constantly, i just had to give up on SDXL for the time until i could get a better card.

My only recomendation (other than get another GPU ofc) would be to try to find a SD1.5 model that can do the same resolution and see if that works.
I can't think of the names off the top of my head but i know there are a few 1.5 models that can do 1024 reasonably well.

File: Sigma_06811_.png (1.59 MB, 1280x768)

1.59 MB PNG

>explores the world while nogens seethe

>>101518135
you forgot to read the part where I said the licence is also very important

>>101521772
what's wrong with sigma license?

>>101522788
nothing wrong with that, it's just that sigma is small as fuck, SD3 would crush it if we had it uncucked and the 8b version, don't expect to improve shit if we keep having the SD1.5 size

>>101522821
>if we had it uncucked and the 8b version
Do you think that will ever happen?

>>101522821
welcome to the bigma waiting room

>>101522849
no it won't happen, that's why we need an alternative that is big, has the DiT architecture, uncucked, and with a great licence, and I'm afraid we'll never get this shit, that's why people left this ecosystem in the ecosystem, they feel it's hopeless

ill munch on the undercooked for just a bit while occasionally taking a big bite from the old stuff, maybe mixing the two as well
just how it be

>>101522561
are these base or 900m or

File: 00011-2710379579.png (712 KB, 656x1048)

712 KB PNG

I have a new absolute hate sampling method, which is LCM.
In a significant number of cases, with high steps (45+), it gives me facial artifacts, but such that make "sense", like random pieces of cloth. But the fucker dares to often paint masks on my waifus, despite this line

(mask:1.2)

being included in my negatives.
The model is AnyLora, but what does it even matter? LCM is literally the ONLY sampling method that gives me these results! What the fuck! I mean without looking at the actual math I can guess why it happens (it adds a HIGH amount of random noise in the face area at a comparatively late steps stage, which then gets smoothed to something that semantically makes sense in the face area).

Picture ENTIRELY unrelated, this is just a random cute result from a img2img I did of fanart of Gudako in her boyscout uniform.

>>101523115
>LCM
Didn't the author say he never uses it himself and that it sucks?

File: Sigma_07227_.png (1.74 MB, 1280x768)

1.74 MB PNG

>>101523059
bunline 1024 v0.8

File: Sigma_07240_.png (1.66 MB, 1280x768)

1.66 MB PNG

>>101523059
>>101523446
Oops, lost count. bunline 1024 v0.9*

>>101522860
how will anon wait thodesu

>>101523515
>v0.9
hands look better

Is pony prompting all booru tags, or does it understand descriptive tagging as well? For example, 'A warmly lit bar at evening', rather than tagging it as bar, evening, warm lighting.

>>101524609
find out you self you fucking idiot. not spoon feeding you.

>>101524592
nice

Sigma chad going hard ITT

>>101524592
that looks good, I thought it was a MJ output at first, sigma has a lot of potential

>>101523515
>v0.9
Where is it?

File: Sigma_07293_.png (1.08 MB, 1280x768)

1.08 MB PNG

>>101525279
ty

>>101525478
ty, my mind hasn't stopped being blown by this thing

>>101525452
Making sure new bunline wasn't a downgrade

>>101525600
Give me a few to upload. Work just ended

>>101525788
Bunline is the best sigma checkpoint hands down
Hype

>>101525788
>>101525871
mind sharing me a civitai link?

File: Sigma_07072_.png (1.71 MB, 832x1216)

1.71 MB PNG

>>101525600
It's live

>>101525871
ty, glad you like it! Hoping my efforts bring more talented tuners

>>101525888
https://civitai.com/models/435669?modelVersionId=665122

>>101526015
Thanks!

>>101526015
whats new with this version? more data or something else?

File: Sigma_06977_.png (1.45 MB, 1280x768)

1.45 MB PNG

>>101526152
* Recaptioned all ~60k images and using both sets (random choice)
* Lowered CAME lr to 5e-9 from 1e-8

>>101522876
If they left then they were never built for this to begin with desu

>>101525478
even base sigma is miles ahead of SD aesthetically

How's the double exposure tho

>>101526641
nice

>>101526015
I wonder why he doesn't continue the 2k version, that's the interesting one

SD3 is unbanned
https://civitai.com/articles/6221/sd3-unbanned-community-decision-on-its-future-at-civitai

>>101526811
Sigma bros WW@?

>>101526811
>The Open-Source Alternative
>This is why a truly open-source approach makes so much sense for our community. Models like AuraFlow (Apache 2 licensed) and the upcoming Generative Models from the Open Model Initiative (OMI) offer a more equitable model where everyone – from individual creators to large companies – can benefit from and contribute to improvements.

>no mention of sigma

File: Sigma_07501_.png (1.18 MB, 1280x768)

1.18 MB PNG

>>101526724
>City double exposure with a wolf
Struggles to draw a background but nails the embedded image

>>101526746
ty

>>101526801
Noted

>>101526811
Woo!

>>101526878
kek oh well

File: 00067-1132515665.png (381 KB, 512x512)

381 KB PNG

need an upscaler that doesn't airbrush the living shit out of my subjects skin

File: sig_0074.jpg (176 KB, 1216x832)

176 KB JPG

>>101527965

File: Sigma_07847_.png (2.04 MB, 1280x768)

2.04 MB PNG

>>101528134
nice

File: sig_0076.jpg (93 KB, 1216x832)

93 KB JPG

>>101528134

File: Sigma_07857_.png (1.97 MB, 1280x768)

1.97 MB PNG

>>101528152
also nice

>>101528484
Very nice.
Prompt?

>>101507621

/ldg/ bros..

>>101525545
>>101525545
>>101525545

File: Sigma_07918_.png (1.76 MB, 1280x768)

1.76 MB PNG

>>101528812
What the vendor lock-in is that?

File: Sigma_07933_.png (1.1 MB, 1280x768)

1.1 MB PNG

Good night anon

>>101526811
Is there an easy way on Civit to monitor the adaptation? Very telling that they can't afford the licensing. Slightly surprised that it wasn't done at heavy discount to encourage a userbase.

>>101528812

      message: app.getLocale() == "zh-CN" ? "本程序仅允许单实例运行,确认后本次运行将自动结束" : "This program only allows a single instance to run, and the run will automatically end after confirmation",

That is an impressive localization effort. Intel needs to get their shit together.

Anyways, outside of this being yet another UI does this do anything special. There is some LLM integration, but it looks pretty weak and forge has some better stuff.

It was nice of them to credit vlad. We won't have another ComfyAnon issue. We probably wouldn't without the credit, but I digress.
https://github.com/intel/AI-Playground/blob/main/service/xpu_hijacks.py

>>101528961
very nice

>>101528812
maybe they'll make a good gpu with lots of vram eventually? it doesn't seem to do anything interesting for now

File: 00970-1553354129-masterpi(...).png (1.25 MB, 832x1280)

1.25 MB PNG

Reposting this from the /sdg/ general:

I haven't used Stable Diffusion in about 2 years. Is it still really hard to get dynamic poses and perfect hands / feet, even if you're using a controlnet?
I used to be really into Stable Diffusion, but I got bored and quit because I couldn't do any advanced poses with good anatomy.

Another question, why are there two generals for Stable Diffusion now?

File: sdxl_15.jpg (237 KB, 1480x1328)

237 KB JPG

>>101529460
>Is it still really hard to get dynamic poses and perfect hands / feet, even if you're using a controlnet?
Hands maybe, ask debo about poses, he was doing some girl with a bat I think awhile ago with some pretty cool poses, he should know
>Another question, why are there two generals for Stable Diffusion now?
There arent

>>101521045
>you're either interested in the tech or the coom.
There is a third reason

Hey has there been any better models dropped since I checked a few months ago?

>>101529679
>>101526015
>https://civitai.com/models/435669?modelVersionId=665122

>>101529685
Can you give me the name and I'll just look it up in Metastable? Wants me to sign in

>>101528812
The left most gen is mildly interesting. The other two suck.
Verdict: I don't care.

>>101520706
there there, anon

>>101513247
LOL No. It's going to suck. CPU only with just 16 GB is barely enough for SD 1.5.
You don't need the latest and greatest hardware for simply generating images(you need for training though). Just save about $1000 and buy last gen parts and you should have a decent rig for generation.

>>101527314
lanczos

>>101529674
Yeah, ez money

Keeping it warm for Anon

>>101531250
cool gen, how'd you get the pixel art to look so nice?

sitting on thread to keep anon nice and toasty till bigma day

>>101531324
I dunno, the prompt is simple, maybe it's the lora

score_9, score_8_up, score_7_up, high resolution, rated_explicit, solo, 1girl, azula, topknot, black hair, amber eyes, toned, small breasts, seductive smile, aroused, from above, erotic pose, nude, anus, dripping pussy, (pubic hair:0.7), sake glass, choker, torches, fountain, asian palace hot spring, steam, night, fire nation banners, suds, bubbles, flower petals, dynamic lighting, best quality, d3p1x3l, pixel art, detailed face,<lora:d3p1x3lXLP:1>
Negative prompt: signature, score_6, score_5, score_4, ugly face, mutated hands, low res, blurry face, black and white
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 383375689, Size: 1066x600, Model hash: 821aa5537f, Model: autismmixSDXL_autismmixPony, Denoising strength: 0.7, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 24.5.1, Hires upscale: 1.4, Hires steps: 12, Hires upscaler: R-ESRGAN 4x+ Anime6B, Lora hashes: "d3p1x3lXLP: 7da88233d5c8", Version: f0.0.17v1.8.0rc-previous-1-ga9e0c387

File: 1695440359683715.png (74 KB, 238x138)

74 KB PNG

Any good models to make "icons"?
I need some small representations of the ideas from my papers for my presentation.
Like a phone with houses and roads coming from it or a syringe in a corona virus and so on.

>>101531535
sd3 might be able to do it, you can search the op for a huggingface space link so you can try it out online first

>>101528744
Prompt?

>>101528812
Had me excited for a second. This is literally just dreamshaper.

File: 00089-398910205-score_9, (...).png (917 KB, 960x1440)

917 KB PNG

gonna post this slop i made from 2 months ago and see if janny didn't like it

File: Sigma_07945_.png (1.11 MB, 1280x768)

1.11 MB PNG

Good morning anon

>>101529382
ty

>>101531824
> A woman is seen sitting in a garden, her hands occupied with knitting. The artist has used soft, warm colors to create a peaceful and comforting atmosphere. The woman's expression is focused and content.

File: ENHANCE.png (987 KB, 1271x1650)

987 KB PNG

'sup AI bros, I wouldn't normally do this, but I feel like there's a big potential to employ your tech here. Please see those two posts (now archived) for a general idea:
>>>/pol/475312298
>>>/pol/475313316
And the overall threads:
>>>/pol/475307020 (archived)
>>>/pol/475319446 (current)
>in b4 /pol/, nypa, schizo, etc.
Like I said I wouldn't normally do that, but I think there's a better way around the problem than just tweaking contrasts knobs and shit.
Do with it what you will, but please have a look without prejudices first. I'm not sure if this is the best thread (of all AI imagery related ones on /g/) to ask this desu, but I won't bother you or the others again. Thanks for your time.

File: Sigma_07951_.png (1.85 MB, 1280x768)

1.85 MB PNG

>>101532645
Decoding secrets is fun stuff! No time for me to do that today. You might be able to employ https://github.com/csguoh/KD-LTR

>>101507621
Hunyuan doesn't seem to work well without the ddpm sampler, I got poor results with extramodels comfyui (though with ddpm from the GUI it works fine). I guess one would have to install the diffusers/ddpm version thru their github since extramodels doesn't have support for it. Also https://www.shakker.ai/modelinfo/87e2cc2169934523a2ff82fb12e7206b?from=feed waiting room. More LoRAs have been added too for those interested.

File: Sigma_07978_.png (1.48 MB, 1280x768)

1.48 MB PNG

>>101532754
Try translating to Chinese first. Not joking

>>101532696
Thanks, I'll pass this on!

File: Sigma_08017_.png (851 KB, 1280x768)

851 KB PNG

>>>/g/ldg is a page 1 thread. Sigma grindset

>>101511495
>>101511525
Also try --medvram alongside --xformers

File: 00090-3442032018-score_9,(...).png (1.47 MB, 1024x1536)

1.47 MB PNG

migu

>>101532769
With Chinese it tends to perform better, I wonder just how much we don't know about the model. Maybe I should leverage the use of Yi 1.5 34b for accurate translations.

File: Sigma_08032_.png (1.37 MB, 1280x768)

1.37 MB PNG

>>101532898
When it's good, it's better than Sigma

Ready to roll, early morning freshly baked...
>>101532879
>>101532879
>>101532879

Is it possible in comfyui to reliably JUST run e.g. facedetailer again, instead of everything? I know I got it to work a couple of times by going back into the history and then changing the seed of the facedetailer, but it is not working now and I don't know what I'm doing differently. Muting doesn't work because then facedetailer is missing an input.

New bread this way
>>101532879
>>101532879
>>101532879

How would running local on an RTX 3080 compare to NovellAI's image generation? Trying to decide if it will be worth the time setting it up.