another good collage
we're on a roll here

this thread has been claimed by jen con

Garbage collage

>>109327658
explain

blessed thread of frenship

is the 78 year old unc from last thread still kicking

>that collage

At this point I dont know what to say anymore....

>>109327664
hes not in it

>>109327664
If my gen isn't on it, it's a bad collage :(

>>109327680
its so good your speechless

My collage education is paying off

^look at him... just.... look at him....

>nigga didnt have the balls to include >>109327362

Listen, I'm only here to discuss local diffusion and look at images of beautiful girls. What's the issue?

collage would be upgraded 1000% if 1girl gens were banned. at this point it's not 2023 anymore and they are not impressive in the slightest.

>>109327706
thats a very tame gen compared to whats in OP

the safety bypass lora for krea just fucks with the image quality so much that it's not worth using.

Not to mention, they aren't even "hot" unless you're underageb&d in which case I think thats this entire thread.

>>109327624
what kind of gooner shit is this pic

>>109327723
>artist name, masterpiece, best quality

>>109327624
I wish I was a video chad so I could animate some of those JenGens

>>109327735
Use the krea rebalance node.

>implying mona lisa doesnt exist
up your 1girl game chud

At this point I'm beginning to think there are bots auto deployed to every moderate trafficed thread to consensus crack and make sure no one can enjoy en mass.

You know whats sad is that I find out Krea knows so many styles and characters natively and yet all I ever see is fucking tifa and realism 1girl slop.

Thats how you know - human intelligence peaked in the 90s and its all down hill from here

>my medium effort gen makes it into the collage again
ez

>implying there arent 10 slop generals with anime slop

>>109327788
Blame the blind retard who bakes these troll threads

>>109327799
but only blue board ones let gookmoot collect money for sexchatbot ads :)

>anon thinks his gens are special

>32 posts
>no images

>anon is again buttmad because his gens aren't collage worthy

File: 1933.jpg (96 KB, 683x1024)

96 KB JPG

>>109327831

we shall protest by not posting our gens ITT

we r legionz

File: ComfyUI_Chroma_Krea_00462_.jpg (3.56 MB, 2048x2048)

3.56 MB JPG

Chroma has been solved after like 200 different versions of it bros. I've finally done it. Finally we can lay to rest the argument of which Chroma version is le better. And no, this is not a LoRA cope or finetune like SPARK, a simple merge brings out the best of it.

>>109327922
chroma krea merged checkpoint? wtf, share it bruh. i'm a chroma fan too

File: ComfyUI_Chroma_Krea_00468_rh.jpg (2.9 MB, 2048x2048)

2.9 MB JPG

>>109328063
Not exactly, several exciting things are happening in my WF. I merged the HD delta weight with 1 HD to get its difference, this brings into the HD Flash weight enhanced prompt following and a bunch of styles missing from HD Flash (essentially Chroma 1 HD is now just 8 steps). This works best a 2K res. Krea comes in as a refinement of this improved Chroma HD Flash model in a 2 pass workflow. In some gens like these last two I'm doing a color transfer. The results are stunning. The WF itself does the work of fixing Chroma HD Flash's gen, but HD Flash delta can be mixed with for instance v29 (I haven't tried this myself), so essentially Chroma has been solved. Initial Chroma gens are not too bad as you can see here
>>109327580
I'll share the wf when I upload the model, it's quite big (17 GB).

>>109327953
Masterpiece!!

Vermeer “Girl with Pearl Earring” version with JCon next?

File: ComfyUI_temp_bpymv_00216_.png (2.28 MB, 1024x1024)

2.28 MB PNG

Chinaman during battle of Nanking

File: ANIMA_bface_bad_00001_.png (872 KB, 832x1216)

872 KB PNG

The reason why krea 2:
>>109328123

Looks so horrible is because it's bad at anatomy.

Anima was trained on NVIDIA Cosmos. an architecture which ironically runs nicely on AMD gpu.

Anima realism workflows have achieved massive gains that result in realistic genuine photos, that don't have anatomical inconsistencies of poor cgi.

Compare this image. It's far from perfect, in fact, I don't like it. However, it incorporates a lot of real-world knowledge. The fact is, in all the world, anima is the best at feet.

>>109327624
<first image

bad spine. Instantly recognizable as krea 2.

>second image
inconsistent humeri. In the left image her humerus doesn't reach her belly button, but in the right of the montage it does.

This kind of problem comes from chasing the "numbers" instead of trying to get artists (such as myself) and other subject matter experts to rank output.

>pinky up
The anatomy is basically incorrect. You can find people with strange hands or whatever, but that's not correct for how hands work.

>catgirl
broken anatomy that is in krea 2 fashion pasted over by elastic skin.

What bones are under krea 2 models' skin? It must be some horror freakshow.

^Look at him.

Just....

look at him

*spams the generate button*

>mfw Resource news

07/20/2026

>Kandinsky5 Lite I2V
https://huggingface.co/kandinskylab/Kandinsky-5.0-I2V-Lite-5s

>Cosmos 3: Omnimodal World Models for Physical AI
https://huggingface.co/nvidia/Cosmos3-Edge

>PE-Field 4D: Video Generation Models as Canvas
https://github.com/MTLab/PE-Field

>StructGen: Disambiguating Multi-Reference Image Generation via Structured Context Modeling
https://jianingpeng0382.github.io/StructGen

>SlotMem: Character-Addressable Internal Memory for Narrative Long Video Generation
https://github.com/YilaiLiu-HKU/SlotMem

>Crayotter: Traceable Multi-Agent Workflows for Long-Form Video Editing
https://github.com/idwts/Crayotter

>JLC Flux2 ControlNet for ComfyUI
https://github.com/Damkohler/JLC-Flux2-ControlNet

>JoyAI-Image-Edit-Plus-ComfyUI
https://huggingface.co/Comfy-Org/JoyAI-Image-Edit

07/19/2026

>Kura: Experiment workspace for LoRA training AI agents and image-comparison runs
https://github.com/nomadoor/Kura

>China bans AI “boyfriends” and “girlfriends” over addiction and birth rate concerns
https://www.dexerto.com/entertainment/china-bans-ai-boyfriends-and-girlfriends-over-addiction-and-birth-rate-concerns-3388737

>Local AI Toolkit: Curated Collection of Locally-Runnable AI Models
https://huggingface.co/m15dg/local-ai-toolkit

>Krea 2 ReID Reference
https://huggingface.co/yijunwang2/krea2-reid

07/17/2026

>Rare Concept Generation via Counterfactual Inference in Diffusion Models
https://github.com/200204jzy/CI-Diff

>Uni-AdaVD: Universal Concept Erasure for Visual Generation via Orthogonal Value Decomposition
https://github.com/QifanZhou/Uni-AdaVD

>MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators
https://harahan.github.io/meanflownft-project-page

>TanGO: Training-Free 3D Editing via Tangent-Space Guidance and Optimization
https://github.com/siw00-lim/TanGO

>GlobalForge: Towards Robust AI-Generated Image Detection
https://anonymous.4open.science/r/GlobalForge-BE0F

>mfw Research news

07/20/2026

>ArtChart: A Benchmark for Faithful Artistic Chart Generation with Integrated Text Rendering
https://arxiv.org/abs/2607.16060

>Efficient Difficulty-Aware Dynamic Routing for Diffusion-Based Real-World Image Super-Resolution
https://arxiv.org/abs/2607.15711

>Debiasing Text-to-Image Evaluation via Implicit Cultural Alignment Reward Modeling
https://arxiv.org/abs/2607.15740

>Test-Time Noise Guided Adaptation for Realistic Autoregressive Video Generation
https://mever-team.github.io/tango

>DiTango: Cost-Effective Parallel Diffusion Generation with Selective Attention State Reuse
https://arxiv.org/abs/2607.15650

>FVAttn: Adaptive Sparse Attention with Runtime Load Balancing for Video Generation
https://arxiv.org/abs/2607.16190

>Personalized Image Aesthetic Assessment via Preference-rich Sample Mining and Cohort Merging
https://arxiv.org/abs/2607.15752

>Attention-Guided Saliency Maps for Interpreting Visualization Literacy in VLMs
https://arxiv.org/abs/2607.16105

File: Screenshot 2026-07-20 230459.jpg (128 KB, 1759x592)

128 KB JPG

going to do a wakfu amalia lora for krea2. should i strictly use screenshots from the French cartoon show of her or mix it with fanart that somewhat resembles her season 1 and 2 appearance.

>>109328252
>it is not correct for how hands work to have one pinky up when drinking
Huh?

>>109328299
Yeah, use a 3D hand model, you'll see that there's a problem.

Again, I reeeeally don't want to memorize my bones just to critique krea 2, but shall I be forced? um. like. the ones before the metacarpels ahhhh I don't want to know all the bones plox

>>109328304
Do you happen to take daily medication?

File: ANIMA_bface_bad_00002_.png (906 KB, 832x1216)

906 KB PNG

>>109328252
anima understanding anatomy:

>Her left little toe is bent outwards.

>>109328318
yes but that is a pointless thing to axe

>>109328315
I don't think any pay or open models are really on point when it comes to flash. I mean, I can tell. I appreciate that those aren't gnome logo feet.

>>109328298
variety is always good, as long as it's not too off model. better if you can find an official artbook or character reference sheet though.

>>109327624
>all that gpu power
>wasted on generating roasties

>>109328374
this is good. normies will get bored of that shit after a month and then the space can return back to the real enthusiasts. just need to wait for the trash to filter itself out

File: ANIMA_bface_bad_00003_.png (897 KB, 832x1216)

897 KB PNG

>>109328329
Now the right one.

>>109328252
>>109328329
>use example with heavy dof to prove a point that your model isnt shit
>look closer
>the dof is misaligned and not based on depth; its applied randomly by gut instinct
>fades out at the edges of corners
>scarf is clearly visible while the shrub on the same depth level is fully blurred

I just got a 9070XT and I'm running cachyOS
What do I have to do to set up AI generation on this thing? What exactly do I use?

>>109327624
So which of these am I downloading and learning if I just want to nudify girls I know? Images are fine.

>>109328397
please sar its the bones to pay attention to

>>109328399
the easiest way is using swarmui since the installer support amd ootb

>>109328409
flux 2 klein 9b + https://civitai.red/models/2592290/sexgod-klein-9b-image-edit-female-nudity-helper

>>109328397
You can fix that in post. Can't fix bones in post.

>>109328409
ignore poster above he's a pubetranny
use this https://huggingface.co/wikeeyang/Flux2-Klein-9B-True-V1/blob/main/Flux2-Klein-9B-True-bf16.safetensors

>>109328423
why are you helping india?

>>109328399
>cachyOS
why would you do that to yourself
>>109328423
probably the worst one ive seen

>>109328445
what the fuck is wrong with cachyos

File: Screenshot from 2026-07-2(...).png (241 KB, 656x1212)

241 KB PNG

On some models picrel doesn't work as well, such as z-image, because it handles the prompt rather differently. Still, I would say it's useful.

What I am doing here is putting together info that Comfyui *should* pull together, but doesn't. Basically, it *should* store a flipbook of the gen's vae preview. anyway, you want to watch each step, see, the issue is that different sigmas determine layout of differently sized features. Here, I use the Accessibility zoom tool of Ubuntu to turn on the crosshair. It lets me see what step I am on, and the corresponding sigma. I know I am aligned by checking the command prompt window which displays the numeric step, so eg 50 I want the 50th step to align.

tan2 lets you force extra energy on one frequency. imo could use a "tan3" perhaps, but this is fine since I'm doing it for fun.

>>109328430
Sure I'll take a look.
Is there anything like that older program DreamTime? That was around when deepnudes were just getting popular.

>>109328470
just stick to comfy
you just drop a picture, write whatever the hell you want to do with the image in the prompt, and press run

File: 1760937061905379.png (624 KB, 896x1152)

624 KB PNG

>>109328046

>.png
Your VRAM is being exploited right now

>>109328399
avoid docker. it's very confusing. It's anti-home-user. It's not actually a vm. You wind up having to pass your gpu in, and pass ports to get access, and then, you can't really use snapshots because they aren't real vm snapshots. It's intended for servers.

So, you need to install Python on your system, and pip.
Then you need to setup a venv
then every time you do stuff with your comfyui install, you "activate" your venv.
actually, you can use git clone without activating it, so yeah. but when you install the requirements, you need it activated.

yeah, the requirements, you install them, but like uh... you need to install the amd pytorch and a couple others, so what I find is I install the requirements, then unistall the ones that are nvidia and install the amd ones.

so that's the thing to get your head around. You have a separate folder for your comfyui venv where basically a bunch of python stuff gets installed by pip.

And you have your comfyui folder, and realize inside of your comfyui folder you have your models folder, and there are specific folders for all your model downloads, including the vae, the clip/qwen type thingies, loras, and of course the models themselves.

the other big folder is the comfyui custom nodes one. usually you use the manager to install custom nodes.

YOUR LAUNCH STRING FOR COMFYUI MAY NEED SOME PARAMETERS, for amd. I know I do.

>>109328414
does comfyUI not support AMD?
what do you have to do to make it work on that?

File: 25445-Princesse-Amalia-Sh(...).jpg (280 KB, 500x800)

280 KB JPG

>>109328350
trying to look for those. too much low res screenshots and amateur art in the boorus and google search. might have to use gpt image 2 with references of her to generate 2mp resolution images of her to add to the dataset.

>>109328534
>does comfyUI not support AMD?
it does, but i think you have to manually install pytorch with rocm support. swarmui is just a comfyui wrapper btw

>>109328539
you can download the show and screenshot

>>109328508
Did you know that the vram storage of the various layers of your os display ie say your windows and stuff can be graphically somewhat compatible with your model's latent space, if the wrong page is pulled by the vmm?

>>109328399
oh! also

First, before you attempt ComfyUI, get Blender acceleration working. That's a really good test of whether or not you have your user setup right.

You have to add users to groups, forget the ones that need which ones. basically, it tells amdgpu which apps are allowed to use it.

>>109328542
the swarmui github just has a .sh installer which I'm a little wary of because I don't know exactly what it's doing
I want to be able to completely uninstall everything if I felt the need to

we need way more diversity in terms of our language. more tags, not fewer.

An example, why can't we richly describe noses? What about big toes, those can be really different. Fingernails are a huge one, some people have a lot more flesh around the nail, while others have nails that spill over the edges of their fingers before the tips. There are distinctive palm lines as well.

File: ComfyUI_Chroma_Krea_00470_.jpg (2.73 MB, 2048x2048)

2.73 MB JPG

>>109328063
>>109328123
>I'll share the wf when I upload the model, it's quite big (17 GB).

Actually, it's the int8 version kek, but everything is blazing fast.

Here is the WF
https://files.catbox.moe/sop5gj.png

Model
Chroma HD Flash and Delta Weight merge int8 convrot
https://gofile.io/d/1Lq0K5

T5 XXL int8
https://gofile.io/d/n0Pijm

Not VRAMlet friendly unfortunately. But fast enough if you have at least a 3090. Feel free to play around with the distill weight (0.9-1) and denoise amount (0.5-0.55 works best)

>>109328252
Nah, Anima is not fully correct in >>109328329
And on top of that its realism world knowledge is bad despite being trained on a world model, though Anima is neat for anime styles/NSFW and I still use to in Krea refinement workflows.

File: 1762657560196195.png (3.45 MB, 1920x1080)

3.45 MB PNG

>tfw think I'm finally burnt out of generating random Kiara images to procrastinate working on my VN of her
feels bad.

>>109328560
if you want to uninstall everything, just delete the swarmui directory and uninstall dotnet

File: Screenshot 2026-07-21 000213.jpg (189 KB, 1911x1069)

189 KB JPG

>>109328544
I'm rewatching the episodes on youtube and taking screenshots of her. i see why many anons fell in love her and Evangelyne character designs.

>>109328607
You're not an artist, so your opinion doesn't matter.

>>109328592
they all have the same feet. Extremely unsettling.

>>109327624
>beleted
kek
>>109328613
could be worse at least youre not a schizo or suffer from some other affliction of the mind like that other anon

>>109328613
krea 2 is the model that one shotted the mentally handicapped (autists) who can't discern art or anatomy, and belong in a cage.

krea 2 is capable of detecting autism. add it to the dsm-7.

>>109328626
Please tell me you downloaded the video with yt-dlp and are systematically splicing each into X number of frames instead of taking screenshots of the webpage

>>109328626
What an ugly art style.

>>109328616
I asked chatgpt and it's recommending against the swarmui installer
It's saying to just point swarmui at an already instead comfyui backend

>>109328673
then you need to set up a working comfyui backend lol, at that point just use comfyui

I'm retarded and don't understand anything about AI
Why does comfyUI run in a virtual environment? Why does it run a local server?

File: Screenshot 2026-07-21 000757.jpg (162 KB, 1825x1072)

162 KB JPG

>>109328661
hell no I'm not downloading whole episodes of this show just for a lora. its easier to pause the video, use the snipping tool and select to spots to capture.

>>109328673
ask gemini free, it's better at (basic) Linux tips.

>>109328692
>snipping tool
right click on the video twice

>>109328691
>Why does comfyUI run in a virtual environment
it's a python thing
>Why does it run a local server?
welcome to modern software design

>>109328692
>its easier to pause the video, use the snipping tool and select to spots to capture.
not when you have yt-dlp + ffmpeg and know how to use them kek ask ai to help you fren
>hell no I'm not
then clearly you do not care about how good your lora will look

>>109328691
to filter the normies imo

>>109328692
if you download the episodes you can use this if you have linux or WSL github.com/negaga53/neme-anima

File: videoframe_238762.png (2.39 MB, 1920x1080)

2.39 MB PNG

>>109328705
that works, nice but now i have to manually crop her from the rest of the group. It's way easier to pause the video and use the snapping tool. This save the video frame approach would make sense of for art style lora of the cartoon or general multiple character lora pack.

File: 1783575594926313.png (2.9 MB, 2022x1080)

2.9 MB PNG

oh nevermind guess I'm not

>>109328770
broken wrist.

anima wins yet again.

>>109328778
Incorrect

I literally did the pose in real life to make sure before I posted.

>>109328748
I think it would be more sensational to run with the "he was making nude images of his underage coworker" rather than "he had ai pizza" desu.

>>109328794
the left hand is incorrect. where's the pinky?

File: 1764714996593374.jpg (31 KB, 600x584)

31 KB JPG

>>109327624

Why AI cant just OPTIMIZE ?

- Imagine using RAM instead of VRAM to load your models. Or better, using SSD/HDD to load your models
- Imagine model compression from 100gb of models to just 1gb of models with no quality loss
- Imagine using both of your CPU + GPU to generate your models faster

Why AI cant just OPTIMIZE ?

stoopid frong poaster

>>109328825
Answer my question you faggot. Why AI cant utilize my Ryzen 9950 ?

>>109328691
Everything runs on local servers now because you need all of your jeet employees to access it as well

>>109328838
It can if you use a real AI inference program like AniStudio.

>>109328847
there's no risk of anything leaking with it running this way is there?

>>109328872
Unless you're a dumbass who enabled --listen you should be fine

>>109328838
>Therefore, in terms of independent task scheduling and control logic, 5,000 stream processors are roughly equivalent to 80–100 CPU cores, not 5,000.

This assumes clockspeeds are locked.

So, you can defo do the cpu route, in theory, like with idk. threadripper? but for such a price, you'd just get a big gpu, because there also is a memory issue, gpu usually have faster ram

cuz

like dawg

computer is CRAZY.

File: 1776724377272421.png (2.36 MB, 1248x1664)

2.36 MB PNG

Dunno what the fuck that guy is talking about - but okay.

>>109328838
Why can't you utilize my cock in your ass hole

The faggot hours are from 4 AM to 5 AM.

Shut the fuck up.

>>109328886
I am a dumbass computer illiterate but I daily drove a steam deck as my only computer for years and now I have a proper Linux desktop so I'm trying to actually learn things
and I learn by asking stupid questions that I get made fun of for asking because they're really basic and I have huge holes in my knowledge since I never did any kind of studying

>>109328907
why ask on 4channel instead of chatgpt

>>109328907
do u have a good card tho cause thats important

>>109328898
>Dunno
yep

>>109327851
Can someone make a moot lora for krea?

>>109328960
finally a real use for the abominable krea 2, which does not offend God.

>>109328967
#badanatomy #krea_2

>>109328890
>gpu usually have faster ram
Not a issue if you just read instead of Copying like AI always do.
Optimize nigga ! Optimize !!

>>109328913
I don't trust AI to not just make shit up
>>109328916
9070XT

did the slop factory explode

>>109329005
How do we solve the equations, with the cpu being an advantage?

Maybe it can be done, can it?

>>109329033
Ask the guy who made Roller Coaster Tycoon 2

File: debo_sa_k2_00001_.png (3.63 MB, 1792x1195)

3.63 MB PNG

>>109329033
just solve for x bro

>>109329039
idk man

>>109329016
vomit inducing errors

>>109329016
where is jen con

>>109329047
no more jen con today

>>109329052
fuck off fag

an example of autism

>>109329052
>>109329047

>>109329056
you got two whole threads of jen con fuck you

>>109329060
we need two more

Finally make it into the faggollage and it gets deleted.

>jen con faggolage gets deleted
and that's a good thing

can krea2 do black mario poppin a goomba while hittin a fat doobie

>>109329047
i had a dumb idea but the short humeri really put me off
https://files.catbox.moe/nj4ypy.jpg

I definitely have some sort of addiction.

>>109329136
ah, too bad

I just tried ollama for the first time and gemma4 is saying it can't do anything sexual
what is even the point of this gay shit
what model do I use to write smut and roleplay with

File: file.png (65 KB, 912x763)

65 KB PNG

>>109329277
try a system prompt (picrel is in open webui) or append something like this one to what you're trying to do if you can't figure out where to set a system prompt.

>>109329277
Ask in /lmg/

I made a character lora for krea 2 and it worked pretty well after only 500 steps and at a scale of like 0.15 WTH? Anyone else try to see how low you can get the LORA weights to go?

>>109329292
Do I have to use webui or can I just keep using the command line?
I can set system parameters through that but wasn't explicit about it ignoring safety bullshit

>>109329305
i'm a dummy that's happier with a gui, dunno about command line usage. there are uncensored and abliterated models of gemma4 out there but this system prompt seems to work ok for me on regular gemma4.

>>109329277
Look for an uncensored or abliterated model. That said, if the model wasn't trained on what you want it do, there will always be limits to what it will output.

File: 1755152722793360.png (34 KB, 528x606)

34 KB PNG

>>109329277
Gemma-chan is really good but she requires a bit of brainwashing to get going.

Look up uncensored ggufs on HF.

File: 1771416718794936.jpg (302 KB, 2048x1816)

302 KB JPG

>>109327624
Before i wasted another 20gb....

Can Krea2 Edit existing Images ??????????????????????????

>>109329323
I am retarded
what are these exactly

>>109329277
You can turn him into a sex pest by using those abliterated models but his maths will take a hit

>>109329323
>>109329331
nevermind I found these, how do I use them through ollama?

Im just so fucking used with SDXL prompting every other AIslop generator feels like shit to me..... except doing Text i guess

File: 1756242426486775.png (94 KB, 904x1006)

94 KB PNG

>>109329331
Just ask Gemma-chan she will answer all your questions.

I'm using LM Studio here.

>>109329355
>No example
Looks like a scam

>>109329292
I'm trying this shit and it's still refusing to do anything
what the fuck is the point of a local model when grok will ERP with me whenever I want it to
I fucking hate computers

File: ComfyUI_Chroma_Krea_00489_.jpg (2.42 MB, 2048x2048)

2.42 MB JPG

While playing around with Chroma Krea I found something interesting that pertains to Krea gens, lowering denoise on regular Krea produces the same exact style that I asked for, in fact the model becomes much more interesting with styles with a denoise of around 0.65 as opposed to 1. But this generation itself is prone to limb errors and even moreso if the textfusion is enabled, so it would need to be 2-passed and fixed. It's also much less censored at that lower denoise level. I will test this out more after this.

All sane anons have left ldg
Most of you faggots are no skill low IQ schizos.

Where do you guys post your nsfw gens? I need to expand, not just my dick, but my traffic.

>>109329430
civitai discord

File: 00014-45.png (2.44 MB, 1728x1344)

2.44 MB PNG

Any music anons around? What is the best tools for AI cover songs right now?

>>>/wsg/6199738

>>109329551
I'd like to know as well. Stock ACE-Step seems kind of ass.

>>109329379
>it would need to be 2-passed and fixed
Interesting results with Krea with 2 Ksamplers at low denoise, but it's too slow and prone to body horror (bad feet). I'll stick to Chroma Krea
https://files.catbox.moe/v3ly2b.png
https://files.catbox.moe/iepmrt.png

>>109329551
>>109329577
>Cover songs
That would be ACEStep with a cover. Use ACEStep.cpp and select no-fsq feature at a strength of about 0.5 on the or less depending on which model you're using (I still recommend Turbo-Base merge). But if it's just plain cover, as in you don't need to change the contents of the song and just the voice, the use a RVC model instead (ACEStep devs specifically left out voice swapping out of their model for their API).

Qwen Music will have cover mode too
https://huggingface.co/papers/2607.11699

But we still don't know whether that model or its cover mode will be open-sourced, but since the paper is out hopefully not long before we know.

about to infect myself with dolphin porn
https://www.reddit.com/r/comfyui/comments/1uw9tts/how_to_prevent_unnecessary_reads_from_the_disk/oycro51/?context=3
/blogpost/

File: Wan21_SCAIL2_webm__00002_.webm (2.08 MB, 624x352)

2.08 MB WEBM

>>109329619

https://github.com/ServeurpersoCom/acestep.cpp

I assume this is acestep cpp. Thanks, I'll mess around with acestep to see what it can do.

>>109329619
Thanks. I tried ACE-Step in ComfyUI and I didn’t like the sound of it but I’ll try it again with those settings. I have a couple of songs I’d like to cover/remaster/remix.

Why is op image deleted

>>109329669
Np anon. ACEStep.cpp is faster and much more memory efficient than Comfy, even on good GPUs despite its name. Keep in mind the cover feature itself is it hit or miss, if syllable count is too misaligned it will sound off, and it will still skip lyrics if you change all of them, none of the gens I got from the nofsq cover mode were perfect since it was more of a workaround than an official mode. Here are two covers
https://vocaroo.com/11E5J1UtRHkV
https://vocaroo.com/12EgDtyMeM0z

Those two both contained many changes, this one didn't contain as many
https://vocaroo.com/19RJdI0fBJ45

The voices are spot on but the instruments are not good, or at least as far I tested it (there could be better settings). The output could be placed on top of the actual song if you have the instrumental version of it and find a way to isolate the vocals.

okay I just got ollama working on my pc with an uncensored model
how do I set it up so it remembers past conversations and I can run it outside of the terminal?

So I've started using Anima and holy hell the hand problem is back like I'm stuck in 2023. They look like garbage. Is there a way to fix this?

>>109329803
https://chatgpt.com/

>>109329790
I don’t want to sound like an asshole, but those examples are kind of what I mean. I need it to sound more dynamic and less like MIDI.

File: uvc and mangio rvc.jpg (394 KB, 2051x843)

394 KB JPG

>>109329790
This is how I isolate vocals. If anons have a better method, do share. UVC does the brute work. Mangio-RVC separate the vocals and background choir accompaniment if it gets stuck with the main vocal.

>>109329794
That's a man.

File: 1782900768346391.png (211 KB, 396x542)

211 KB PNG

I miss Near so much....

>>109329662
alright this is some tranny shit. it extended the first load by 13% !AVOID! /blogpost/

>>109329833
>I need it to sound more dynamic and less like MIDI

All I can say is look into Matchering 2 remastering of the tracks (available via UVR 5), it does bring out vocals and instruments. I do not have much experience with ACEStep cover though, it sounds like shit since the devs never properly supported the feature, but if you must use cover for some reason, I'd recommend training a LoRA of 10-20 songs at rank 256 so it memorizes the precise songs, then playing around with the cover feature and strength, you never know how good the results will be.

File: 918302163606100.png (1.33 MB, 960x1280)

1.33 MB PNG

>>109329842
nyo~

>>109329900
>never properly supported the feature
(And I mean for us plebs). Their paper had a key adapter that was never released and is instead available on their proprietary API https://acemusic.ai/

On top of that they have a better stem splitter that removes voices and instruments on that same API.

File: 1763988894173557.png (2.16 MB, 1474x1580)

2.16 MB PNG

Remember when Alibaba didn't release Qwen Image 2.0? Yeah that was funny, well, now they made Qwen Image 3.0 kek.
https://xcancel.com/bdsqlsz/status/2079470829621129233#m

>>109329900
Luckily, they’re instrumentals. Training a LoRA is probably the only way to make them sound any good.

>>109329939
Sad that when Alibaba got good even with small models, their first instinct was just to betray us. Z was the exception and not the rule.

>>109330000
>Z was the exception and not the rule.
yeah, it was made by a small team so they didn't get much scrutiny, Alibaba thought they were gonna release another garbage model and they didn't

>>109330017
The model was good, but not NBP good. Qwen Image 2 was NBP good despite only being 7B. They know exactly what they're doing.

>>109329904
>>109329967
>>109329985
>>109330002
>>109330025
>slop spam
I thought I was on /sdg/ for a second

>>109330028
>Qwen Image 2 was NBP good despite only being 7B.
not even close, it still looks very slopped, the Qwen Team is a souless team, training their model on synthetic slop is their motto, same shit for their LLMs they're synthetic trained benchmaxxed pieces of shit

>>109330028
And the model that came out now appears to be GPT Image 2 good (likely still slightly worse in overall knowledge though, as was Qwen Image compared to NBP).

>>109330037
Have you prompted for graphic designs or images with lots of text? It was neck to neck with NBP on those, just didn't know all the precise fonts like NBP, but didn't produce garbage like all the models we had until recent models that are good at text (Ernie, then Ideogram 4). The only reason it still feels like we don't have NBP is because Ideogram 4 doesn't have Edit (its slop aside).

>>109330061
>Have you prompted for graphic designs or images with lots of text?
We already have a local model specialized for that no? That SenseNova thing

>>109329346
you install ollama and pull the model you want to use from the cmd line.
>install ollama, run it
>go to ollama.com/huihui_ai/gemma-4-abliterated
>open a cmd line
>copy and paste " ollama run huihui_ai/gemma-4-abliterated " into your cmd line and hit enter
>if you are using ollama in comfyui just refresh your shit with R and select the new model.

File: 00174-1965200038_cw.png (3.19 MB, 1087x1920)

3.19 MB PNG

Chunky stew.

>>109330169
can you make her hold the spoon in the same position with her foot that has a bracelet on it?

Nonsensical early diffusion images? Totally soul and not slop.

>https://qwen.ai/blog?id=qwen-image-3.0
open the page source and check the meta tags

>qwens corning oakridge roof shingles, qwens corning roofing warranty,
>qwens corning weather lock, rick jones obituary richmond va qwen,
>rqvin starfire and qwen sfm, rule 34 qwen stacy, rule 34 qwen tenison xxx,
>rule 34 xxx qwen tenyson, rule 64 spider qwen, sexy busty qwen stacy,
>sexy qwen ben ten, sexy qwen stacy, shadbase qwen, shemale eskort qwens,
>shy-girls.com qwen, skudbutt qwen, skudbutt qwen por,
>spider qwen diaper, spider qwen diapered, spider qwen
>doujin, spider qwen dove cameron, spider qwen dp hentai,
>spider qwen footjob, spider qwen funko, spider qwen futa, spider qwen
>gets fucked, spider qwen harry, spider qwen hentai, spider qwen hentai
>comic, spider qwen hentai comics, spider qwen hentai manga, spider
>porn comics, spider qwen porn game, spider qwen porn games,
>spider qwen poses, spider qwen pussy, spider qwen r34 comic, spider
>qwen rule34, spider qwen rule34 xx, spider qwen series, spider qwen
>series resding order, spider qwen sex comic, spider qwen sex comics,
>spider qwen sexy comics, spider qwen shiroqq, spider qwen shirt,
>spider qwen shoes, spider qwen spiderman tv cartoon series, spider
>qwen stakeout, spider qwen svg, spider qwen symbiote, spider qwen
>symbiote hentai, spider qwen thicc, spider qwen tied up, spider qwen
>toddler costume, spider qwen trans, spider qwen wallpaper, spider qwen
>teamwork, spiderman and qwen stacy, spiderman into the spider verse
>did qwen meet peter, spiderman qwen stacy, spiderverse how did qwen
>get there first, spiderverse qwen, spiderverse qwen porn, spiderwoman
>qwen, spidet-qwen, spidey qwen hair, spidey qwen svg, spyder qwen,
>stefani qwen, sto qwen phil procotor, stop qwen, stripper harley qwen,

Were you looking to gen 'spider qween futa in a toddler costume'? Look no further, qwen got your back

>>109330306
>rule 34 qwen tenison xxx,
wtf?? lmao

File: ddc.jpg (25 KB, 320x320)

25 KB JPG

>qwen 3 4b ablieteirnd, qwen 3 4b alibteateratoanga
>anal spider qwen, all of ben 10s monsters fuck qwen, ben 10 four arms porn qwen
>aunt qwen tranny of alberta, auntie qwen tranny, auntie qwen tranny photos, ben 10 porn boobs female qwen, casting couch-x paige qwens

>>109329842
stop forcing your faggot fantasy onto everything tranny gay cross dresser. stupid faggots cant just be normal they have to demand and larp everything is faggot and trans. nope its not you're 3% of the world population flaming queer groomer paedoe. stfu

File: 1046177522646710.png (340 KB, 480x640)

340 KB PNG

>>109330385

>>109330385
>ablieteirnd
that's so 2024

>>109330306
>ben 10 using all 10 monster to fuck qwen
>qwen tennyson giantess pussy
>color picture of harily qwen
>is qwen stephanie for hillary
>jojo's menacing adventures roblox how to git qiller qwen
>madison park qwen accent chairs hardwood birch faux linen
>qwen stavy porm

I love the idea of people searching for this shit somehow ending up reading a Qwen press release.

>>109330306
Wonder how good qwen image 3.0 is at generating qwentai, seems like they're confident about this area

File: file.png (317 KB, 444x440)

317 KB PNG

>>109330306
>>109329939
So now that they have a version 3.0, they don't need the 2.0 anymore and they'll give it to us right?

File: 55031198684791.png (329 KB, 640x480)

329 KB PNG

>is qwen stephanie for hillary
Well?

File: 1773296875914546.jpg (177 KB, 880x1168)

177 KB JPG

>>109330414

File: 1040696402239311.png (393 KB, 480x640)

393 KB PNG

>>109330495
kek

i just found out u can save WAN's video latent then use it to continue the genning the next video segment. but it's troublesome and a lot of manual work.

File: Blue Skyscraper, drawn by(...).jpg (570 KB, 960x1356)

570 KB JPG

What's the best model to do landscape artwork like pic related?

File: jiaxing-shang-111.jpg (1.21 MB, 3000x1421)

1.21 MB JPG

>>109330589
Another example. Not just hyperrealism but also painterly/drawn look.

>>109327624
>deleted
why?

File: 148381139216584.jpg (800 KB, 3072x1280)

800 KB JPG

>>109330589
>>109330605
Probably Krea 2?

File: 1772807612692008.png (158 KB, 288x292)

158 KB PNG

5 years ago i wouldn't imagine my favorite anime girl getting fucked in any degenerate pose i want and look at we are now.

AI is really amazing. Artists need to accept reality.

Yes yes i am pajeet indian yadda yadda yadda

Can someone gen some blonde Japanese girls, with big sagging breasts if possible.

>>109330002
her feet are ugly.

Collage got deleted, Karma for it being goonerslop of 1girl, 80% of the collage literally being softcore porn of the same washed celeb

>>109330759
collage got deleted for being too based and mind raping zoomers like you

Should i monetize my slop for few mcdonalds bucks ?

>>109330759
Women scare den Janitor

>>109330796
You don't ever gen anything
Fantasizing and using buzzwords won't make a difference

What scheduler do you use for Krea2? I hate scheduler variety the most because on some models there really are multiple viable good picks and some that are outright better, but the community average IQ is usually too low to converge on the actual good ones because half the retards are imagining things and the other half are using schedulers where each step is double length and think they're getting "free" quality improvement. My usual solution is just to stick with Euler and not waste my time trying to be the hero to XYZ grid my way to the most reliable scheduler, but I'm sure I'm leaving quality on the table.

>>109330811
euler/beta

File: 1107013326074713.png (352 KB, 480x640)

352 KB PNG

>>109330749
that's good, in a sense

>>109330848
damn, animu misato mogs irl misato hard

File: 447679048902315.png (366 KB, 640x480)

366 KB PNG

>>109330811
er_sde/simple

File: 1102354954644751.png (414 KB, 480x640)

414 KB PNG

>>109330863
as she should

>>109330811
i tested every sampler in the ksampler and most of the clownshark samplers, for single pass euler is the best imo.

>>109330811
With a distilled model you want the schedule of the sigmas to be as close to the ones it was trained on, that should lead to the closest to intended image. I've read that comfy's simple scheduler is slightly different from the reference, so you may experiment with that, but other than you are theoretically best off sticking to the euler sampler.

File: ComfyUI_temp_ljnjm_00006_.png (2.3 MB, 1024x1024)

2.3 MB PNG

>>109330811
>>109330921
>>109330865
res_multistep is good if you want to spice things up a lil. It sharpens detail without significantly degrading image quality. Though it's not really an improvement, it adds random spice which can make the image seem less real. I would say Euler is the best. All other combos are certified placebo.

>>109330636
Can't post my next gen on a Christian website
https://files.catbox.moe/js38tz.png

>>109327624
Kik Epp23g
Tele Bgftg33
Tg BGGTNYDDG
Discord Grn9363

Can anyone make a Lora of my asian gf?

>>109331027
doesn’t work. catbox has been fucked lately.

>>109330811
>>109330924
https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler This is the reference scheduler for comfy.

>>109331027
>>109331041
I see, https://litter.catbox.moe/69u9znnwhyeajaen.png
will do

>>109328480
any chance you have a nudify workflow laying around i could use? i am genuinely stupid when it comes to comfyui and all this stuff and i have no idea how to build a workflow

>>109331133
Just use AniStudio. It has everything you need, and you never have to worry about building your own workflows

>>109331033
She's your gf. Just ask her for nudes

>>109331158
can i use anima

>>109329838
genuinely i haven't found anything better than lalal for vocal separation. nuostems is decent and local though.

>>109331166
Not yet

>>109331060
From diffusers docs:
"""
scheduler (FlowMatchEulerDiscreteScheduler) — Euler flow-matching scheduler. The Krea 2 sigma schedule is the resolution-aware exponential time shift, so the scheduler config is expected to set use_dynamic_shifting=True together with the Krea 2 shift parameters (base_shift=0.5, max_shift=1.15, base_image_seq_len=256, max_image_seq_len=6400).

Whether the transformer is the few-step distilled (TDM/turbo) checkpoint. When True a fixed timestep shift mu=1.15 is used;
"""

File: anima_fudder.png (32 KB, 414x197)

32 KB PNG

>>109331166
Half a year later and you still can't. And you know why? Because the creator views it as competition.

File: image.png (149 KB, 845x516)

149 KB PNG

Does cumfartui not support Krea2 in int8 mode?
What does your log say when trying to load the model after restart?

[INFO] Native ops: int8_tensorwise, convrot_w4a4 , emulated ops: mxfp8, float8_e4m3fn, float8_e5m2, nvfp4
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[INFO] Requested to load Krea2TEModel_

>>109329939
fucking jackasses never even released any version of the previous qwen image 2.0. They're truly done with releasing local open source models for low end hardware. Seriously alibaba can go fuck themselves if they think they're SAAS worthy in competing with bytedance, kling and Z.ai.

File: ComfyUI_K2_slop__00326_.png (2.29 MB, 1240x1864)

2.29 MB PNG

1. finepornV2TURBO

File: ComfyUI_K2_slop__00327_.png (3.21 MB, 1240x1864)

3.21 MB PNG

2. kreamania_variant5