Dubs = bbc board

comfy should be dragged out on the street and shot

>>107386240
...with love

chinks do not deserve to walk this earth

my two rank 4 character loras turned out way better than my two rank 32 character loras. just saying

>>107385996
sure but it's writing me a fucking research paper before it gives me the prompt
also it needs to be jailbroken

坚持习近平思想

File: ComfyUI_temp_jqhpi_00015_.png (2.76 MB, 1344x1856)

2.76 MB PNG

whats the latest meta with the z-model, I was out the last couple of days, is anyone using facedetailer? hires?

Can someone pls post the chink pasta system prompt for Zimg?

>>107386273
>it needs to be jailbroken
very minimal one, or if local you can get a Derestricted version

File: BWCslave10.jpg (571 KB, 1800x1200)

571 KB JPG

Reposting from last thread:

Nofap is really dangerous. Almost every time I develop some fetish or something I've never felt before. This time around I've developed the strong urge to rape Asian women. I don't even find them attractive, and I swear to god I never have before in my life. And now every time I see one that isn't overly hideous, which most of them still are, I think to myself, "Made for BWC". I don't see them as humans, but rather as property made for the pleasure of white men. I want to slap them, choke them, bruise them, and impregnate them. I even want to fuck kids. Asian girls need to learn from a young age to worship white men.

asian ladies with large posteriors

>>107386230
Yeah most people are like that, to them AI image gen is a hobby and they aren't gonna shell out 3k for a new GPU and RAM if some fat 24GB+ model drops even if they could. For a model to get traction it needs to work well on current midrange GPUs.

File: speeeed.jpg (5 KB, 430x150)

5 KB JPG

is this a good speed for training a Lora?

>>107386309
if you're training on 256x256px images its very bad. if you're training on 4096x4096 then its very good.

>>107386297
kys

so what is the ideal image size for z image, is it 1024x1024 or more?

>>107386295
ill have to wait for a quantized version

>>107386356
way more

THE BASE
THE BASE IS REAL

Z-image or noobai?

>>107386297
She looks underage

>>107386417
sd 1.5 sir

>>107386417
depends on if you want realism or illustration
respectively

File: YunyunStore.jpg (2.22 MB, 3072x3456)

2.22 MB JPG

my wife Yunyun looking at the store for food to prepare the meal for tonight's dinner after i took her after school, is summer.

>iphones have paid and free image generation apps that run locally on the gpu
>android phones don't have anything like that despite high-end android SoCs being better at it
Why is that?

mmmmmmmmmmmmmmmmmmmmmmmmmmmmhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh i know for a fact this one is underaged

>107386297

How the FUCK do I get crossed legs on zimage?

>>107386417
Go for Noobai if you want better quality for anime style.

>>107386451
Asians girls have no ass or tits and their vaginal canals are very shallow. They're literally built for microdicks

>>107386393
THEY BELONG TO US

Alibaba here. Base has been canceled due to >>107386297

>>107386428
apparently she's 25

>>107386488
loraloraloraloraloraloraloraloraloraloraloralora
or just gen enough so that it listens and does it

Man it feels so good being able to gen batches of 4 images and not feel the card dying from oom.

so are tokens separated by commas or periods for naturalistic prompting

>>107386442
I really like the shading and lineart.
Can you share the catbox or model and prompt?

File: ComfyUI_temp_qnyyp_00069_.png (1.77 MB, 832x1248)

1.77 MB PNG

>it keeps the top buttons closed
kek

File: WifeInTheStore.jpg (2.71 MB, 3072x3456)

2.71 MB JPG

illustrious:
#style
(realistic anime lines:1.5), (flat colors:1.7), 8k, anime screencap, masterpiece, soft lineout, very detailed, highly detailed,

no idea what catbox is, i'm just a humble stable diffusion user.
also i want to turn my wife fantasy prompts into videos for maximus schizo.

File: StoreTimeWithYunyunWife.jpg (2.46 MB, 3072x3456)

2.46 MB JPG

>>107386672

>>107386287
Everyone's upset because Base isn't out.

>>107386513
>no ass or tits
Perfect.

>their vaginal canals are very shallow
I'll manage.

Does the prompt enhancer increase the seed variance or is it just bloat?

File: qwen_zit.jpg (726 KB, 3860x1080)

726 KB JPG

Amazing. Qwen Image (left) is worse than ZiT (right) despite being larger and two times slower (with quants). Qwen is worse at making realistic-looking things that are very rare in reality (space station interior), it makes concepts bleed through (water for "floating", the korean flag for "korean girl"). Or is it just issues caused by quants? I'm using Q6 for unet and Q8 for clip. What do anons who can run full qwen get?

>>107386758
>That's easy, Anon. It's apple!
>Correct! *zzzip* How about this one?

File: ComfyUI_00425_.jpg (840 KB, 1920x1920)

840 KB JPG

>>107386488
>seated cross-legged

>>107386655
When was the last time you talked to a real female?

>>107386774
Q6 is still good. Qwen is slopped in the dataset. Also try compare it with base qwen since the edit model isn't mainly for genning.

File: ComfyUI_04821_.png (1.2 MB, 832x1216)

1.2 MB PNG

>>107386770
>increase the seed variance
Doesn't seem like it.
>is it just bloat
The model benefits from longer prompts up to a degree so I wouldn't say that.

>>107386788
She crushed someone in img 2
Also that's indian style not crossed legs

File: Zurbo_00055_.jpg (606 KB, 2688x1536)

606 KB JPG

>>107386774
I usually run Qwen at fp16 and the concept bleed is very real, still. I noticed that Z barely has any, if ever. It's kinda amazing.

File: Z Image Turbo.jpg (3.82 MB, 3277x4096)

3.82 MB JPG

https://civitai.com/models/2093591?modelVersionId=2456392

Z Image Turbo DALL-E 3-like Girls LoRA

File: Capture.jpg (608 KB, 2048x1216)

608 KB JPG

>>107385604
You can read out this tutorial if you don't want to be a promptlet anymore I guess.
https://www.reddit.com/r/StableDiffusion/comments/1pawepr/advanced_prompt_techniques_for_stronger_artistic/

>>107386850
What the fuck why would anyone make this?

File: 3321543235.png (1.09 MB, 896x1152)

1.09 MB PNG

>>107386828
It really doesn't want to do crossed legs.

>>107386868
I like the style

>>107386856
>spamming miku 24/7
>calling others promplets

>>107386802
It benefits from adding a lot of details, not word fluff.

>>107386850
not trying to shit on your work but, why? these look like shit, the art style I mean.
the lora seems to do what you wanted from it so good work I guess.

>>107386868
You'll understand when you're older.

>>107386901
The effect is the same as just using surface blur in photoshop.

>>107386868
word the dataset is too old

File: 1739455005951285.jpg (382 KB, 1920x1080)

382 KB JPG

>>107386297

I've got games to play but my gpu is busy generating asian girls I hate this

File: ComfyUI_temp_qnyyp_00081_.png (1.8 MB, 832x1248)

1.8 MB PNG

I guess I can't just paste the system prompt into the clip window...The chink runes are bleeding into the pic.

>>107386791
I tested it with t2i qwen.

>>107386836
If it's an issue even with fp16, then I guess zit is really better.

>>107386922
that's why you get another cheap gpu ($200-300) to play games.

>>107386933
What do you call that dress my man?

>>107386789
>female

pot and kettle here

>>107386955
post some more inspo

File: StoreYunyun2.jpg (3.71 MB, 3072x3456)

3.71 MB JPG

please RECOMMEND good workflow to animate my wife.
i have a 24GB gpu.

File: z-image_00211_.png (2.28 MB, 1920x1080)

2.28 MB PNG

>>107386942
A striking image of the planet Saturn dominating the star-speckled black sky, its majestic rings clearly visible in the vast expanse of outer space. In the foreground stands a lone figure—a young woman with pale porcelain skin and piercing blue eyes. Her hair is cut into a sharp, luminous white bob, contrasting dramatically with her elegant, flowing black robes that drape around her defined hourglass shape. Bold black lipstick accents her lips, and a delicate silver necklace rests against her collarbones. The scene captures the nostalgic grainy texture of a 90s film photo taken with an instant flash, illuminating her face against the infinite darkness.


im running old 1.5 prompts through glm 4.6

My favorite hobby is imagining zimage base getting released and all the stuff I'd do with it

>>107387012
Can you just go and rub one out to calm down on your own instead of larping here?

>>107387030
this is what a 30 year old asian woman looks like

>>107386970
Just grab a recentish wan 2.2 workflow and you are good.

>>107386933
>>107386998
Kino images
NTA but thanks for the prompt.

uh a lot of these girls look underage

>>107386991
where did you get that pic of me

How do you use the prompt enhancer?

>>107387100
it's just a prompt you put into an llm

>>107386209
https://huggingface.co/spaces/victor/Z-Image-Turbo-MCP/blob/main/pe.py

File: HugYunyun.jpg (2.71 MB, 3072x3456)

2.71 MB JPG

>>107387047
thanks i'll check, also good workflow or how to generate good quality backgrounds, in this gen the backgrounds sucks... a lot.
>Yunyun: "come here to my arms my tired little pog champ!"

File: 1746847365874988.jpg (41 KB, 562x575)

41 KB JPG

>>107386991
I'm preoccupied with imagining what it must be about prepubescent children and underdeveloped women that compels you to share your sexual desires publicly on the internet in such a fashion.

>>107387124
Model?
Also how are you genning that resolution?
Maybe just post a catbox to show us the metadata.

So can I look up the images posted above or is it going to put me on a list? You can never tell with chinks.

File: 50_3.jpg (84 KB, 1024x1024)

84 KB JPG

>>107386206

Anyone else get the sinking feeling that we will not be getting that base model?

>>107387171
There is no reason we won't.

File: ComfyUI_00008_.png (3.46 MB, 1920x1088)

3.46 MB PNG

Let Qwen3-Next write you a story to be used as a prompt for Zit

>>107387171
The chinks saw all the degenerate coomslop everyone is posting and now know why Americans censor their models. It's going to be retrained with safety in mind.

>>107386856
explain how this weighting works

>>107387160
perfect illustrious
https://files.catbox.moe/kb5apd.png
https://files.catbox.moe/jd0ga7.png
https://files.catbox.moe/t7uqfl.png
https://files.catbox.moe/po715e.png
https://files.catbox.moe/19mziq.png

>>107387195
china doesnt care about safety, chinese hate safety and like to do things risky and unsafe.

>>107387202
just read the tutorial?
https://www.reddit.com/r/StableDiffusion/comments/1pawt6o/advanced_prompt_techniques_for_stronger_artistic/

File: 1761578356512369.png (2.01 MB, 1024x1024)

2.01 MB PNG

>>107387190

>>107386758
hey, i used to play this game with my niece!

>>107387210
It's always funny seeing other people's schizo negs.

File: ComfyUI_00011_.png (3.15 MB, 1920x1088)

3.15 MB PNG

>>107387187

>>107387171
Alibaba corporate saw how unexpectedly popular the distilled model is and now the undistilled one will be API-only, ala Flux and Seedream.

>>107387195
they already saw that with WAN 2.1, and WAN 2.2. i know this is a shitpost, but in case anyone actually thinks china cares, they don't.

>>107387185
>There is no reason
They can come up with any reason they want at any moment.

File: ComfyUI_00013_.png (3.11 MB, 1920x1088)

3.11 MB PNG

>>107387236

File: 1745784277127543.png (2.13 MB, 1504x1504)

2.13 MB PNG

a japanese woman wearing a white bikini in a hot spring in Japan during the day.

ty for non plastic women zimage

File: 1738616537500404.gif (1.4 MB, 360x238)

1.4 MB GIF

i used z-image to make a loli lick a lollypop while lying on a bed with her ass showing then used twerk lora to animate it

good lord my millennial brain was not built for this level of pleasure

Is every single person in this thread a victim of yellow fever?

>>107387248
Yes my point exactly.

>>107387279
Just the pedos

>>107387279
No I'm a 2D only fag patiently waiting for the BASE to drop and for booru forks to get forked

File: 1751737099109157.jpg (166 KB, 640x898)

166 KB JPG

>>107387190

>>107387275
Imagine being 14 and having this technology available.

>>107387275
make some lolis using this https://civitai.com/models/2181576/

>>107387295
Most 14yo have no idea how to run this shit.

>>107387279
Z Image Turbo devs are geniuses, they knew that the key for a successful model is how good it can generate pretty asian woman.

File: z-image_00220_.png (2.32 MB, 1920x1080)

2.32 MB PNG

https://civitai.com/models/2172944

zit fp8 6GB for the vramlets
pic related

>>107387298
when will we get sliders instead of this

>>107387306
It's amazing how many companies genuinely do not understand that it is that simple. If your realism model cannot make an inhuman KPOP plastic surgery big tiddy asian woman, it is not worth using. If it's an art model, okay, replace asian with anime but 1girl is the basis of all AI art.

>>107387306
>Z Image Turbo devs are geniuses
they are, even releasing the turbo model first is a genious move, vramlets tasted that forbiden fruit with high speed and are now addicted to it they'll be able to accept the slower base model, if they released the base model first, the vramlets (the vast majority of the population) wouldn't have given the model a chance

>>107387298
bimbo loli..interesting concept

man, some people are really destroying the z model with their loras, some of the samples are really rough lol

>>107387318
Is there a reason to use this over GGUF for 4gb vramlets? Not hating, just a moron who only kind of understands how this all works.

why didn't they use qwen 14b as the text encoder

>>107387335
shieeet if ur gpu got 4gb then it got 4gb

File: Zurbo_00011_.jpg (576 KB, 2688x1536)

576 KB JPG

>>107387279
I have all kinds of fever.

>>107387333
i tried some trump gens with it but it wasn't as funny as i thought it was going to be but bimbo lolis does sound better

So far i've trained 5 different Z loras and they were all WAY worse than the Chroma / regular Flux / Qwen versions I've done on the same dataset. Think I'm gonna wait for the base model to try any more.

>>107387354
how about bimbo trump

>>107387348
I voted for this!

>>107387335
Q8 GGUF is better quality than FP8.

>>107387210
Ok first of all SDXL (and therefore all of its derivatives like illustrious) just kinda sucks when it comes to making detailed backgrounds that are also coherent. Not too much you can do here.
Maybe inpaint better backgrounds with a better model, inpaint the girl with illustrious over a better base image or use an edit model to change background without touching the girl.
Besides that.
1024x1152 is not terrible but not a great resolution. Stick to 1M pixels.
Stick to actual booru tags until you know what you are doing. Shit like "feminine girl" needs to go.
Vae decode to Gan upscale to low denoise upscale is a schizo workflow.
Needless vae decode/encode just degrades quality. Just Switch to a 4x GAN and ditch the last sampler or just use the upscale latent by node (You probably want bicubic) to upscale with the sampler without the GAN
CFG 7 can be a bit high for some illust models.

>>107387235
lol, you saw my negative, is hard to find the sweet spot

>>107387195
Why would they care when they even care about the tons of chinese porn sites out there?

>>107387334
No surprise, distilled isn't the best to build concepts on.
I hope all that energy will be on the base model when it'll be fully trained and released.

>>107387357
You have to try it in comfyui directly and try at strength 0.7

>>107387360
Troons are still here and Israel gets more funding. Congrats

>>107387335
None, stick to gguf.
>>107387348
He is only based like that in our sloppa.

>>107387279
I don't care about race. I usually make 1girls of every race.

Are you going to apologize to me when they inform us the base will be API only?

>>107387373
IF it gets released, you just know there are already people who don't want it to be released and will try to stop them

>>107384783
nice!

File: file.png (273 KB, 518x428)

273 KB PNG

>>107387379
>and Israel gets more funding.

File: ZIT.jpg (122 KB, 768x1280)

122 KB JPG

>>107387385
>I don't care about race. I usually make 1girls of every race.
Based

>>107387395
It will get released, stop dooming with zero proof outside of paranoid brain.

>>107387288
>Just the pedos
this is false
>t. pedo without yellow fever

>>107387359
i tried earlier but i didn't get anything good so moved on after a few gens, too many loras to catch up on.

>lumina won

File: ComfyUI_temp_qnyyp_00094_.png (1.77 MB, 832x1248)

1.77 MB PNG

WHY DOES THE MODEL DESPERATELY TRY TO COVER HER BODY HOLY FUCK. It's fucking conjuring some opaque shirt out of thin air how do I stop this??
>>107387298
Good one.

>>107387334
too many people who've never been told their outputs are shit, especially after flux

>>107387364
>1M pixels
what is that? i use resolutions composed by numbers of the power by two sequence.
also i have 24GB vram, what res do you recommend for landscape and portrait generation?

>>107387448
add uncensored in positive and censored in negative.
(not expert in workflows but those tags work for me)

Yeah Z-image is cool but can the chinks hurry up and make a visually lossless scaler already

Reminder now is the perfect time to earn a shitload of buzz when Z-Image-Base releases. Make essential NSFW loras and you will never have to worry about buzz again.

>>107387470
negatives don't work with zit. it's a distill model. it doesn't use cfg.

>>107386297
king

>>107387486
>you will never have to worry about buzz again
You don't have a gpu?

>>107387064
record and post the video once its december

>>107387470
The image has been trained on images that have 1 million pixels (1024x1024) so anything above or below that the coherency starts going down. One of your examples has 1024X1152 which is not ideal. (For the base image.)
>landscape and portrait generation?
Start with 1216x832 832x1216

>>107387530
*The model has been trained.

>>107387295
i was 14yo when SD1.4 released

>>107387502
You need yellow buzz to download early access loras. nochekaiser881 in particular makes good loras for XL, but he makes the early access ones cost 5k buzz.

>>107387544
>Admitting being underage.

>>107387558
i am 18 now???

>>107387530
but 1152 is 1024+128 why is 1216 or 832 better?

>>107387492
You can increase the CFG and use negatives with ZIT to some extent. I've tried it and it works just fine.

>>107387555
>saw a cool Z-lora
>early access: two more weeks
absurd

>>107387490
but it's worth it anon, destroying anatomy understanding is the best thing ever if it means safety is reached!

>>107387571
which one? most sane creators only make them cost 100 or so buzz, which anyone should be able to have.

>>107387574
yeah! it's so safe no one wants to use it.

File: 1759980771847212.jpg (1.56 MB, 4096x2068)

1.56 MB JPG

https://civitai.com/models/2181569
Imagine finetuning Z-base so that it has MORE bokeh, jesus

File: 2025-11-30-21-06-49_00001_.png (1.37 MB, 832x1216)

1.37 MB PNG

>>107387590
>I love using AI models but the lack of real lens control frustrates me.
He has a point.

>>107387590
bokeh is good for artistic shots, but some people go over board and use it literally everywhere.

File: 1764525894039346.png (1.37 MB, 1024x1360)

1.37 MB PNG

whoever posted this, can i have the prompt? this is cool as fuck

Wait, if you vae encode a base image instead of an empty latent image and feed that to Z-Image with a denoise of 1, you can still see the base image in the final output in composition and color. How the hell does that work? Shouldn't a denoise of 1 make it almost entirely unrecognizable?

File: 1764526383242205.png (1.22 MB, 1584x672)

1.22 MB PNG

>ctrl f
>comfy
>10 results
Ehemmmmm you didn't learn your lesson you little sheeps, every thread until everyone unanimously stops using that garbage UI.

File: 1745392080920856.png (1.02 MB, 1024x1024)

1.02 MB PNG

game show:

>>107387565
>14 in 2022
>18 in 2025
The math doesn't add up
>>107387568
Multiply 1152 and 1024.
Now multiply 1216 and 832.

>>107387647
1. Can neo load up workflows from comfy right away?
2. Can you generate almost the exact same images as comfy if the seed and the rest of the params are the same?

>>107387659
it does because it's been more than 3 years and he could've been 14 years old and 11 months or something

>>107387659
brother..
Yes, it is possible to be 14 years old in August 2022 and 18 years old in November 2025. This would happen if your birthday is in September, October, or November.

Here's why:

From August 2022 to November 2025 is 3 years and 3 months
If you were 14 in August 2022, you would typically turn 15 in the months that follow
By November 2025, you would have had your birthday 3 times (in 2022, 2023, and 2024), making you 17
But if your birthday is in September, October, or November of 2025, you would turn 18 before November 2025
For example, if your birthday is in September:

August 2022: You're 14 (haven't had your birthday yet)
September 2022: You turn 15
August 2023: You're still 15
September 2023: You turn 16
August 2024: You're still 16
September 2024: You turn 17
August 2025: You're still 17
September 2025: You turn 18
November 2025: You're 18
The same pattern works if your birthday is in October or November.

Zamn these retards are pumping out Z Loras FAST
All trash though

>>107387700
why dont you do one yourself?

>>107387700
we have good pussy loras. just need a good penis lora that works from most angles. I have no idea why people don't do penis loras first.

>>107387659

About Dimensions
A minimum of 262k pixels and a maximum of 1.04m pixels are recommended when generating images with 512px models, and a minimum of 589k pixels and a maximum of 1.04m pixels for 768px models. The true pixel limit is 1048576.

To avoid the dreaded 6N IndexError it is advised to use 64px increments when choosing an aspect ratio. Popular ratio combinations for 512px models include 1536 x 512 and 1536 x 384, while 1536 x 640 and 1024 x 576 are recommended for 768px models.

For 512px models, the minimum useful sizes are 192-256 in one dimension. For 768px models the minimum useful size is 384 in one dimension.

Generating images under the recommended dimensions may result in undesirable artifacts.

i got it why. https://platform.stability.ai/docs/legacy/grpc-api/features/api-parameters

File: ComfyUI_155543_.jpg (369 KB, 1920x1440)

369 KB JPG

>>107387700
a lot of them look p fun tho

File: ComfyUI_01759_.png (3.59 MB, 2048x1248)

3.59 MB PNG

At this point I am sitting a 50/50 chance of the base model not coming, 90% chance of the edit model being distilled with no base.

There's too much hype.
No clear unambiguous communication Tongyi.
Previous history of rug pulling within Alibaba.
The general fickleness of Chinese companies.

A screen shot on their hugging face of the models reading "To be released" means nothing.

>>107387700
>All trash though
that one is good though
https://civitai.com/models/2174416/technically-color-z?modelVersionId=2448632

>>107387647
No other UI provides as much freedom as Noodles.Besides, it's super easy to write custom nodes and include custom postprocessing.

>>107387723
i for one - dont care if they release base, because im a vramlet
lora training works on turbo
ill be very happy if they release base because nunchaku will eventually support it but still

>>107387740
>dont care if they release base, because im a vramlet
base will be the same size though

>>107387742
*sound of glass breaking*

File: zimage.png (2.3 MB, 2048x1024)

2.3 MB PNG

>>107387603
based newsposter

>>107387736
This. Making extensions in Forge sucked.

>have a full dataset of stuff I want to train
>it's all tagged for Noob/Illustrious
Am I going to need to run that chinese LLM prompt on all of them to change it to normal descriptions?

>>107387742
source?

>>107387705
icba
i used kohya but now i gotta install the other one and deal with the troubleshooting of installing it no thanks i went through the headache of setting up sage attention today
i tried to train a lora through civit but it canceled on me and the fucks haven't refunded my buzz yet

i think its best to wait it out and see what the best settings are for training. some anon said he made several of the same loras with different parameters and he said the rank 4 is better than the 32

>>107387784
read the paper

>>107387782
I mean give it a shot? I am curious how well it would work or not.
But I would probably just caption from scratch with joycaption.

>>107387700
it is not good to train on turbo due to reasons
to test training? yes it is ok.

these loras will only (and big if) work on their own, cant combine more than two and even two in most cases.

civit stuff is waste of downloads.

>>107387796
civit complained to me when i tried to send the training data with booru tags

>>107387817
loool if he posted this i would start learning chinese that day

>>107387792
Can you point where in the paper they broke their own NDA and said Z-Image-Base is going to be 6B, directly?

>>107387833
>they broke their own NDA
what do you mean?

>>107387833
nta but by default distilled model is smaller than the base.. that's the logic.
i am speculating the base model is 16-32b which is big ouchie for rectum

>>107387847
A user asked them what size base would be, and they said "can't, NDA".

>>107387848
>nta but by default distilled model is smaller than the base.. that's the logic.
it's absolutely not, there's 3 ways to distill a model:
- guidance distilled
- steps distilled
- size distilled

so far the only model we know is size distilled is Flux Klein, we don't know for the rest

>>107387848
nunchaku will save us
threadly reminder qwen image uses only 3gb vram with nunchaku

>>107387861
he also said he doesn't work for them, how does that work?
https://xcancel.com/bdsqlsz/status/1994804699657392222#m

>>107387863
My rectum is ready.

>>107387848
Of course. But anon insists it is 6B because "the paper said they are targeting Z-Image for vramlets!"

>>107387882
I mean, they said that what BFL did is overkill with their 32b model so it's obvious it won't be that size

>>107387873
>but there is some cooperation.
If he can't tell others that information, then the team could have told him(@bdsqlsz
) it was due to an NDA.

>>107387889
No one is expecting 32B. I'm not even expecting 24B. Ideally 8~12B, but 6B for base would be ridiculous.

File: ComfyUI_00192_.png (2.21 MB, 2048x2048)

2.21 MB PNG

comfyui is the best just on capability i understand people dont like the interface but the advanced wan vace and animate workflows really need something like comfy

Fools getting excited for a model that will never come. You get to enjoy the demo version. For the real stuff you need to fork out.

in comfyui with dynamic prompts, can you 50 50 a single word to be included or not? like maybe {girl|}? or is it {girl}

>>107387907
b-but they said it was coming soon™..

>>107387905
XY grid.

>>107387910
after testing yes, it seems {word|} works, leaving an empty option

>>107387901
>Ideally 8~12B, but 6B for base would be ridiculous.
I would settle for 12b, the vramlets can use nunchaku if they're too desperate

>>107387714
Most people aren't gay?

Can you convert a tag based dataset to a natural language one?

>>107387950
desu, if you dont have 16gb gpu minimum you really have no business doing ai stuff.

>>107387953
but most people look at straight porn, and there's a penis in there anonie

>>107387953
you can't do porn without a penis retard, unless you're only into lesbian shit.

>>107387963
Sounds gay.

File: 1733470759460981.png (69 KB, 1571x829)

69 KB PNG

shit like this is very easy to test with comfy.

>>107387714
>good pussy loras
last I looked it was roasted beef like detailed, I'd rather have just a slit

File: ComfyUI_00219_.png (1.25 MB, 832x1216)

1.25 MB PNG

>>107387932
what do you mean?

>>107387969
meant for >>107387942

>>107387972
https://civitai.com/models/2179465/zimage-innie-pussy?modelVersionId=2454215

>>107387901
My guess is at most 12 to target consumer cards.

>>107387972
king

File: HornyYunyunWantsToBreed.jpg (1.69 MB, 2688x3456)

1.69 MB JPG

>my wife Yunyun when she wants a baby

>>107387983
oh, that's a promising start

>>107387983
I can't pinpoint what the issue is but all the example images look off.

>>107388003
It works great and doesn't change the appearance of the person at all. You know what that means

File: z image.png (1.11 MB, 731x728)

1.11 MB PNG

>low seed vari-ACK
yes, just use a lora

>>107388016
broly culo?

>>107388009
holy sameface

>>107387872
reminds me of the old dune film lol

>>107387983
OHHHHHHHHHHH

>>107388005
it's too large

File: z-image_00241_.png (3.77 MB, 2048x1168)

3.77 MB PNG

>ywn do mushrooms with a qt redhead and experience the mystical whimsy of getting lost in narnia together
yeah nah ima kms

>>107387983
nice

>>107388009
>Same pose

nude loras all fucking suck. you're gonna have to wait for a finetune trained on nudes if you want to gen anything good looking.

>>107388068
just prompted "girl" with a low 1024x1024 res with the dalle girl lora that focuses on girls

>>107387955
Training seems to work fine with just tags. You can just joycaption to convert.

>ksampler seed set to fixed
>hit start
>changes the seed

WHAT THE FUCK COOOOOOOOOMFFFFYYYYYYYYYY

>>107387975
automatic1111 had as one of its frontpage features a grid generator that let you specify variables for any part of the generation. for example specify 5 models along the x axis and 5 seeds along the y axis and output a single file with a grid of all of these images.

when people switched from auto1111 to uncomfyui, this was and continues to be the #1 complaint after noodle soup is understood, because there's no straightforward way to replicate this function which was trivial in auto1111.

File: 1764329147458994.jpg (793 KB, 1344x1728)

793 KB JPG

>>107387526
Pay for my flight to China and I'll upload several videos of me fucking those chinks. I'll even go as low as 8 years old girls if that's what you want to see.

>>107388109
highresfix was also a checkbox
adetailer was a checkbox

>>107388113
i meant anon cooming video

Weights don't work on z-image?

File: 2013001-A disposable phot(...).png (1.59 MB, 1024x1024)

1.59 MB PNG

This shit is pretty crazy for how fast it is!

File: z-image_00243_.png (2.68 MB, 2048x1168)

2.68 MB PNG

>>107388168
yeah they popping the ai bubble with this one

File: 1563912521393.jpg (18 KB, 451x451)

18 KB JPG

If they want me to use a prompt enhancer, why didn't they make the TE an LLM and use the llm adapter node like Rouwei-Gemma and some other sdxl clones where you set the system prompt in the node.

File: WanVideo2_2_I2V_00546.webm (764 KB, 1248x704)

764 KB WEBM

Wow it's pretty nuts what you can do with Z image and wan, gonna blow my dick off here.

>>107388109
I still end up switching to forge to test loras with the xy grid. It's way too useful.

Do we know when we're getting z-image-edit?

>>107387411
As long as you include gassing the nonwhites in the montage we good.

>maximum file size allowed is 4mb
which resolutions is that cba to figure it out.

Uhh guys, what's the discord @ of zimage devs? I want to tell them theres no need to release the base version if turbo works this well.

>>107387448
try the negative?

>>107387478
flux2 has a new vae thingy. is it compatible with z-image?

>>107388213
>>107387492
oops, but isn't there some kind of alternative negative thingy?

>>107387848
they should have worded it better but from the git page
>Z-Image is a powerful and highly efficient image generation model with 6B parameters.
>Currently there are three variants:
but again, they are giving the model for "free" (we pay by melting our brains to ai porn)

>>107388175
bubble, irrelevant: ai is here to stay.

>>107388234
>but again, they are giving the model for "free" (we pay by melting our brains to ai porn)
They do it to undercut competition. Make no mistake there's an AI war going on with China and they're winning.

File: file.png (1.31 MB, 832x1216)

1.31 MB PNG

A base model just flew over my house!

>>107388305
Z-Image being released a day after Flux2 was basically a fucking nuke. Absolutely killed BFL. I wouldn't say it's Deepseek R1 tier, but it's pretty close.

>>107388305
welp
they shouldn't have spent all that time and effort lobotomizing their models
not my problem!

>>107388210
Thank you I have forwarded this to the team :)

>>107388324
>Oh no! The people who provide us with no income by using our free model have been taken away by another free model. What ever will we do with the revenue they don't generate?

>>107388338
stakeholders ganna melty!

>>107388332
they don't have a choice. law makers and western policies force that on them. the whole 'protect muh children's' agenda is unironically going to end the USA as a super power.

>>107388332
Although they posted that wall of text about safety, in practice Flux 2 isn't meaningfully worse in that area than Z.

>>107388305
I know, I'm still waiting for a new CHADseek moment to buy more nvidia and maybe google this time too.
Maybe after the new iteration of Chinese models, the ones that will have 2kk token lenght the stocks will tank by a lot.

>>107388338
>So tell me again why do I need to buy your model if I have this better one for free?
Still, maybe boomers will choose to pay kek

>>107388345
>You lost HOW MUCH nonexistent revenue?!

>>107388362
>I need everything spelled out for me because im retarded!

>>107388338
I mean what is BFL even for? If you want to use a free model, Z-image is the obvious choice all around, If you want to do Pro stuff commercially you would obviously use Nano banana. What audience were those morons even going for lol

>>107388338
Retard. How much money China gets from users doesn't matter right now. Once they have undermined western AI companies, they can then do whatever the fuck they want. This can change geopolitics and make every other country dependent on China.

is there a good way to implement tits, vagina, anus, and camel toe (and what? too much ass?) DETECT so you can automatically move the gen to a "sus" folder?

>>107388374
If I recall correctly. Many of the BFL employees are former SAI employees. So it stands to reason that their retardation was always part of their company structure, it just wasn't ready to show itself yet.

>>107388374
>What audience were those morons even going for lol
Safer than nano banana.

File: 1762314966803766.png (577 KB, 1021x746)

577 KB PNG

What do you even think the base model is going to bring? We can already train excellent loras so what is the point? it's not like you're gonna finetune it.

>>107388346
>they don't have a choice
No they do have one, they just enthusiastically chose to go all the way in safetyism, they don't even have the excuse of being forced to do it, not when they even want to police what people do with derivatives of their models. That's insane.

>>107388374
Flux 2 is better than Z if you're not doing pure 1girl material, and can still be customized by people with resources (ie professionals), unlike the banana family.

>>107388399
the loras will have more flexibility since it was trained from a bigger model.
you're also forgetting the most important aspect, finetunes.

>>107388399
- better loras
- people finetuning it
- community focusing on it and finally abandoning sdxl

>>107388410
Dev looks like crap in pretty much all aspects, you are thinking pro which is a good model but again doesn't hold a candle to Nano

>>107387352
>https://civitai.com/models/2172944
fukin hell

>>107388399
>excellent loras
most of loras published on civitai are overtrained trash

>>107387647
neoforge doesn't support linux, eat a fat fucking dick

>>107388463
because only a few people know how to actually make quality loras. 99% just throw a bunch of images in with ai captions with 1 gorillion steps and hope for the best

>>107388410
flux 2 being better in some things at the cost of insane bloat and no community support is only gonna matter for the craziest of niche "pros" who have thousands of dollars to burn to train their superniche furry lora and nothing else, you will either use zimage or nano

even for some insanely complex prompts just wait for proper thinking in zimage and thats it

>>107388399
I want to run it at full cfg for uncucked results.

You can't add variety or do a large scale finetune to ZIT Turbo without breaking its distillation. It's over.

File: 1754318274239284.png (2.52 MB, 1224x1840)

2.52 MB PNG

>>107388463
civitai loras have always been trash. its just a bunch of indians throwing random images they found on google images at the civitai trainer.

>>107388499
What site has better loras?

I'm trying the basic low vram z-image workflow in comfy with 9 steps euler + simple and I'm getting very noisy images.

Am I just out of luck because I'm using the fp8-e4m3fn.safetensors model + qwen3-4b-q6_k.gguf? Are my images noisy because of the model or can I change any settings to help with this?

Many thanks.

>>107388496
the distillation was already buckbroken with the dedistill adapter, and it will be more so with v2 thats in training

>>107388510
you train loras yourself

>>107388511
post wf

>>107388482
there are cfg alternative methods. I will have too look those back up. People also know how to add negative prompting.

>>107388517
based

>>107388499
aha. Some of them are randomly kind of like... good but don't do what they're supposed to.

>>107388499
oh shiiii

File: 2051004-A black and white(...).png (1.39 MB, 1024x1024)

1.39 MB PNG

Huh, I feel like I'm running out of ideas on how to push z-image.

>>107388511
>low vram z-image workflow
what's that? just set use --novram when launching comfy, surely.

File: lack of skill.jpg (106 KB, 759x1152)

106 KB JPG

>>107388109
and all those extensions made for it make it top tier even if it has not been updated for so long
shame, such a shame.
forge is ok but some extensions do not work on it well enough.
>>107388383
good point
>>107388424
>community
what a fag
>finally abandoning sdxl
not possibru
1.5 is still in use
so shall sdxl be

maybe the janny's asleep

>>107388499
>civitai loras have always been trash
not at all.
guy who made training script used to post his loras there.
there were some like him as well.
maybe 20 authors at most.

and all of them left (some appear on rare occasions) - due to those you mention and many more retards gathering on that site.