Blessed thread of frenship

why isn't my gen in the collage

File: highlights_g_110275467_17(...).jpg (1.33 MB, 2375x2073)

1.33 MB JPG

Ty for bake,
Alt collage

File: MOSS-er_sde-1.0-12-2026-1(...).png (3.06 MB, 1344x2368)

3.06 MB PNG

>>110282455
we'll get 'em next time lad.

File: 2026-11-09-105850_1265x49(...).png (28 KB, 1265x499)

28 KB PNG

lora trainer gods,
Should I care about loss?

>>110282490
no

>>110282463
Much better collage, thank you.

>>110282490
IS this loss?

>>110282510
cumbrain

>>110282490
Short answer, no.
Unless there is a sudden massive jump or fall (very unlikely to happen.)

>>>>>>AniStudio: https://github.com/FizzleDorf/AniStudio
yeah no this dogshit thread needs to be abandoned and purged ASAP
>>110282463
use to bake a proper thread

>>110282530
Oh shit... baking

So trani was also the base > sft troll
Makes sense

>>110282521
The weird sheepishness of this OP about all things sensuous and lusty (even things well within the bounds of the SFW board rules) tells me that he was mindraped into submission by schoolmarms as a young man. The resulting cramped and fearful asexuality, lashing out in moralfaggot outbursts against anons who gen pretty girls etc, shows a sickness in the soul. Watch Ari Aster's great masterpiece about this, Beau is Afraid. Pay attention.

>>110282552
There's no point telling a general audience that base > sft. The kind of anon who gets it will figure it out quickly on his own; the kind who doesn't never will and can't be convinced.

>>110282575
not him but in my experience the mods are a little more strict about op images. regularly i've seen the exact same pictures deleted from an op which aren't otherwise.
either that or they're just more likely to see a risky op image.

>>110282590
>>110282590
>>110282590
Real bake

>no rentry
*yawn*

The rentries are more important than a pathetic anistudio insert that will be removed next bake.

ETA days/weeks until jannies rangeban all trani proxies?

Some anons really try everything they can to take down ldg yet they always fail.

>>110282338
bottom right of thumbnail looked like ryza's ass up face down

>>110282338
why did you bake early again

>>110282725
bro's a fuckin dunce haha lol

>>110282695
see >>110282530
we still need a proper thread btw

cozy breas

>>110282740
great, wanna make a 6th thread and get banned for spamming? go ahead

Is kind of bothersome than certain words have such a heavy anime weight, if you use only one of them your realistic gen converts into 2.5. For example "smug" due to the 69666420 smug anime girls on the internet.

this is the thread.

Imagine spending all your time baking troll threads and then posting this >>110282729

>>110282748
>>110282759
setting yourself on fire would be the most productive use of your organic matter

>>110282756
use photography and other related terms in the genre/medium meta json value and youll never have that happen again

>>110282575
>sheepishness
No anon, Ani & friends are quite open about their sexuality, it's just that it centers around things like sodomizing little boys, things of this nature. They have no feelings at all about photos of women with large breasts/etc, and their own tastes would get them banned, so their ops are 'wholesome chungus' quirkshit instead, which also is a performance of their "I am an innocent little boy" autopedophilia

>>110282338
Cumfy broke Minimax with the last update big time, the default workflow from cumfy itslef do not work. all custom node are broken.
1 Is there a fix yet?
2 Can I rollback ?
If I manage to rollback, what's the last working version of cumfy for minimax H3
Thanks.

>>110282800
>niggerjak projection: the post

>download kromoss
>it doesn't know all the ponies
What's the fucking point of this finetune?

>>110282822
That author is a diaperfag not a brony, you have your finetrooners mixed up

>>110282507
>>110282522
ty. So is there a way to tell if training is going well besides just testing the lora?

>>110282810
I guess you weren't there the day he posted the catbox. Nobody who was there would disagree with what I said.

cozy thread compared to the other where they're lamenting the general itself

>>110282773
What do you mean with: "in the genre/medium meta json value" I already use realism pointers like:
>Professional photograph of two naked girls. Shot on a full frame camera.
and
>The camera is positioned at a slightly high three-quarter angle from behind the goth girl. Simple dark background. Cinematic blue lighting. Soft shadows. Dim lighting." It doesn't work if the word is too anime heavy.

Or do you mean:

Medium: Photograph of two naked girls shot on a full frame camera.

Subject: Two girls naked.

It doesn't work either, using similar words like "arrogant" instead of "smug" works.

>>110282862
Yeah this one is fine, I'm staying here.

Am I a tinker trainy if I spend more time training loras than actually using them?

>>110282831
many trainers let you sample intermediate epochs during training.

>>110282880
That's what I mean by "testing the lora"

>>110282887
Oh. There is validation dataset and validation loss thing, but I never managed to get anything worthwhile out of it.

>>110282867
Search the previous threads for "photography" it was shared very recently. Some anon shared a json string that includes a prompt but also this special json format moss was trained on. In the special format, you can tag specific mediums and genre tags (not in the prompt) that guide your output to whatever you want. I've been using it and it works great for when you keep getting animu with realistic gens.

Here it is I found it >>110269693

>mfw Resource news

11/09/2026

>Open-H3-Regenerate-2K-IC: Consistent High-Resolution Video Refinement via In-Context Learning
https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K/IC
https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K-IC

>T2I Models Rarely Invert Negation but Often Fail to Suppress Negated Concepts: Benchmarking Negation Suppression in Text-to-Image Generation
https://github.com/IAB-IITJ/T2I-NSI

>MESAM: Enhancing SAM2 for Few-Shot Segmentation through Semantic-Visual Memory Generation and Refinement
https://github.com/tianyu-zou/MESAM

>Fizgig v9.3.0: Training maps, faster compile warm-up, and fine-tunes with the optimizer you pick
https://github.com/shootthesound/Fizgig/releases/tag/v9.3.0

>LTX x Chroma Awards Season 2
https://submit.chromaawards.com/pack/ltx

>HexAnima: Anime image generation that runs entirely on your phone's NPU
https://github.com/ayanoby/hexanima

>slopfab: Use agents (Claude Code, Codex, Antigravity, ...) to make videos, images and music locally
https://github.com/bitti-ai/slopfab

11/08/2026

>Viggle Image Swap: Built with Qwen
https://huggingface.co/Viggle/Viggle-Image-Swap

>GrabTag: Cut a reference image into named regions
https://github.com/blackll7/MOSS-image-edit-grabtag

>ComfyUI-SAM3DBody-openpose-editor
https://github.com/tori29umai0123/ComfyUI-SAM3DBody-openpose-editor

>Anima Prompt Generator for ComfyUI
https://github.com/GUntilB/Anima-Prompt-Generator

11/07/2026

>P-H3-Image pDMD 2-Step LoRA
https://huggingface.co/akhaliq/p-h3-image-pdmd-2step

>HyperFlow H3 Flash (14B)
https://huggingface.co/VideoRebirthAI/hyperflow-h3-flash

>gputaskman: An htop for what's inside your GPU
https://github.com/shanselman/gputaskman

11/06/2026

>HyperFlow H3 Flash
https://github.com/Video-Rebirth/hyperflow-h3-flash

>Pai v1.0.0: painting app with AI generation
https://github.com/g45t345rt/paiapp

>mfw Research news

11/09/2026

>InstaGen: Group-Shared Importance Sampling for Fast Video Generation
https://arxiv.org/abs/2611.08610

>Trajectory Forcing: Exploiting Diffusion Trajectories for Autoregressive Long Video Generation
https://arxiv.org/abs/2611.09096

>UniRGBA: Unifying RGB and Alpha in Pretrained Video Diffusion Models
https://arxiv.org/abs/2611.09069

>Scale as a Modality: Multi-Scale Pixel-Space DiTs
https://arxiv.org/abs/2611.08802

>MotionCore: A Latent Neural Simulator for Video Generation with Rigid-Body Dynamics
https://arxiv.org/abs/2611.08904

>OmniLatent: Unified Latent for 3D Part Modeling
https://arxiv.org/abs/2611.09055

>Emo-LiveSync: Real-Time Universal Lip Synchronization with Explicit Emotion Control
https://arxiv.org/abs/2611.09074

>A Bridge from Paste to Place: Physics-Guided One-Step Image Composition via Object-Scene Interaction
https://arxiv.org/abs/2611.09387

>Segmenting Geometric Textures of 3D Meshes
https://arxiv.org/abs/2611.08897

>Align3D: Image-Structure Co-Aligned 3D Scene Generation with Observability-Aware Fidelity-Preserving Guidance
https://arxiv.org/abs/2611.08861

>FANO: Anchored Forward-Noise Optimization for Diffusion Inverse Problems
https://arxiv.org/abs/2611.08923

>On the Few-Shot Adaptation of VLMs: A Few Gradient Signs Are All You Need
https://arxiv.org/abs/2611.08800

>SIFR: Stable Signed Flow Reinforcement
https://arxiv.org/abs/2611.09175

>Training-Free Active Test-Time Adaptation for VLMs
https://arxiv.org/abs/2611.09084

>Seeing What Matters, Verifying What Pays: Load-Adaptive Speculative Decoding for VLMs
https://arxiv.org/abs/2611.01617

>UniMore: Fine-Grained Supervision for Unified Understanding and Generation
https://arxiv.org/abs/2611.08996

>Panoptic Scene Program Diffusion Transformer
https://arxiv.org/abs/2609.31780

>DiRotQ: Rotation-Aware Quantization for 4-bit DiTs
https://arxiv.org/abs/2605.16732

>Beyond MMSE: Enhancing PnP Restoration with ProxiMAP
https://arxiv.org/abs/2605.16396

I have been here for month and I still don't know who exactly is against who and why, but is pretty fucking gay. What I know is that there is at least one authentically bitter and rageful anon, who samefags hard.

Anyway, I'm going to use this thread just to piss off the anti ani side.

What's with all these "hey I've only been here a week and hate this person, but I actually don't know who is who anyway" posts?

File: ComfyUI_temp_nvjeg_00001_.png (2.56 MB, 1664x1248)

2.56 MB PNG

So, 2000 steps (39 epochs), This is what I get but I have to crank the lora strength to 1.5. so it's kinda working.

  --network_dim=32 \
  --network_alpha=1 \
  --network_train_unet_only \
  --learning_rate=2e-4 \
  --optimizer_type=AdamW8bit \
  --lr_scheduler=constant \
  --timestep_sampling=truncated_logit_normal \
  --logit_mean=-1.09 \
  --logit_std=0.95 \
  --discrete_flow_shift=1.0 \

I don't really know what I'm doing desu will do another run and set alpha to "2"

>>110283027
Every single one of them is catjak. Or Debo. Or Julien. Or whoever you hate the most—you get to decide.

>>110283049
> --learning_rate=2e-4 Little high there fren

>>110283006
I have been here for four years and yeah it's pretty gay, but if you can believe it there have been 'hot' phases in these conflicts which were 10x more gay

>>110283061
What do you recommend?

>>110283064
>'hot' phases in these conflicts which were 10x more gay
when

>>110283078
1-e4

>>110283049
Up it to 70 epoch and keep alpha at 1. If you increase it it'll just fry your shit.

>>110283116
>>110283095
so I upped to 4200steps (~70epoch), 1e-4 learning rate, alpha 2.
>If you increase it it'll just fry your shit.
The lora didn't do anything at strength 1.0 so I kinda want a run that will fry my shit. Just experimenting. But it could very well just be that it didn't do anything because it needed more steps.

>>110283049
>but I have to crank the lora strength to 1.5
You are undershooting
Set alpha to 32
Use --lr_scheduler warmup_stable_decay --lr_warmup_steps 0.1 --lr_decay_steps 0.2
Batch size? Optionally lower lr to 1e-4 if it is 1. Didn't you have 200, 300 images or something? You may need to train like 1.5x or double steps.

>>110283160
>The lora didn't do anything
Give an example test prompt and example caption.

>>110283160
I won't stop you from just increasing the alpha, and increasing it to only two might work, I'm just saying I keep it at 1, 70 epochs, 1-e4, and it works great for me.

File: mosss.jpg (522 KB, 1536x2048)

522 KB JPG

moss lady

>>110283205
>base face

>>110283027
newfags love operating under false assumptions with full confidence for some reason. really quite bothersome.

File: mosss2.jpg (386 KB, 1525x1970)

386 KB JPG

>>110283207
the frickin heck is a base face

>>110283223
yeah it's dumb because base produces a plethora of unique faces without even trying, but there's a good shot he knows it's dumb and is saying it as a troll

then again it's 2026, we're not being neurotic about "am I being trolled or not" anymore, it was too stifling, made it impossible to listen to outsider perspectives ("am I being trolled rn?")

>>110283176
>Didn't you have 200, 300 images or something?
not the same anon, I have a 60 images
Thanks for the settings tho, I will try them next. Batch size 1, but that makes me realize I need to crank dat shit up because I've got plenty of room on my 3090.
>>110283188
didn't save anything and I'm running training right now but at 1.0 it wouldn't really change the image much besides making it "brighter" it's only at 1.5 that you could see it applied the style, but it makes more complex scenes incoherent.
>>110283202
I'm guessing you're probably just letting it cook for longer. The reason I want to mess with the alpha is because at 39 epochs, the lora didn't feel like it would be where I want it if I just let it run for longer.

>>110283264
I don't understand how you can't instantly tell that the face just looks "off" which is an immediate tell that they are using moss BASE instead of SFT.

>train lora on my highschool crush
>generate images of her sucking my cock under my desk at work
kino

>>110283304
>because at 39 epochs, the lora didn't feel like it would be where I want it if I just let it run for longer.
usually mine start looking like the dataset after 20 or 30 epochs so it might be your captions or something else. gl anon

File: 2026-11-08-232746_1566x10(...).png (226 KB, 1566x1017)

226 KB PNG

>>110283367
captions look like this. From what I gathered about making loras is that you shouldn't caption the thing you want to capture. So basically in this context, not mentioning the style. Trigger word is added by sd-scripts.

>>110283356
Is she what I consider beautiful? No. Is it an interesting face which is different from all the others I see in these threads? Yes.

>train lora on my highschool crush
>generate pov images of me sucking her cock under her desk at work
kino

>>110282801
>claude, fix comfy not working with minimax
simpleas

>>110283380
Die of HIV you piece of shit poofter :)
With all the hate in the world ;).

>>110283338
Lol you're not winning the face prize for SFT no matter how hard you try to meme this

The reason her face looks "off" is because it is unusually flat with no midface protrusion and a sharp, pointed chin, plus her eyes are open quite wide, showing a little white at the bottom. Plus piercings, and acne obscured under makeup. In short, she looks exotic in a maybe-slightly-dysgenic way, and a little insane.

But that's not "off" in the sense of being a failing of the model. That's "off" in the way that real people can look "off".

>>110283538
you lost :)
and in 2 years you will lose even harder. prepare your anus ;)

>>110283567
It is unimportant, as long as you die of AIDS.
I mean it ;)

>>110283554
Tell me you've never been with a woman without telling me you've never been with a woman.

File: 1792515777919287.gif (838 KB, 300x300)

838 KB GIF

>baddie

>>110283589
You performed the same stupid routine last thread, so you're obviously not responding to the details of any particular image in front of you. "Here's the idea. We call it 'base face' any time a face looks a little unusual. If we get pushback, we call anon a virgin. Guaranteed replies"

"Trolling" like this in 2026 is so boring

So that the theme this evening; faggots, useless fucks and Meta faggotry only retards give a shit about.
Enjoy your stay you waste of oxygen!

>>110283633
>If we get pushback, we call anon a virgin.
Is he wrong?

>crying about ldg
Just close the tab lilbro

>>110283671
he'll adjust he's new

File: debo_pc_km_00004_.png (3.82 MB, 1792x1076)

3.82 MB PNG

some of you guys are very high stress and weak framed

>>110283411
Are you training multiple artists at the same time or something?

>>110283654
Transparent. Of course anyone sounds defensive if they respond to that, which is the point of using it.

This is not my first rodeo. 4chan is over 20 years old. "Anyone who disagrees is a virgin" is very old tech.

>>110283695
no? What makes you think that?

Anon yearns for my kino sovl

>look in thread
>still seething
>>110283693
You always seem to be the most stressed. Are you still salty about getting kicked out?
You begged like a dog for the mods to help you and they just laughed at you and demoted you. You have no idea how many people still make fun of you.

>>110283790
I have never seen debo actually post anything that would make him seem stressed. He sometimes replies to shitposts and accusations made against him but that's it. It's you who seems to be seething all day every day about him for some reason.
How could he possibly have slighted you to warrant this sort of obsession? I have zero idea.

>>110283380
The funny thing about gay pics is that no matter how hot you make the guy, he ceases to be hot when he's being kissed by a guy. This is why the best gay porn involves straight guys whose masculine relationship (bros, rivals, etc.) *escalates* into sodomy domination, never getting kissy smoochy lovey-dovey wholesome chungus

To be clear, I am not gay, however if you listen to testimony of guys who went gay it's usually sissy hypno and Dominated By Doug type stuff that did it (maybe also psychedelic drugs or being molested by an uncle), the idea being that one's place in the bully-and-bullied alpha/beta hierarchies becomes the site of one's sexuality, the nerd fantasizing about being bullied by the jock so hard that the jock actually fucks him in the ass and turns him into a girl

Obviously this is just as natural and valid as liking women and these people should be able to get married and adopt small boys. Love is love.

File: debo_pc_km_00006_.png (3.77 MB, 1792x1076)

3.77 MB PNG

>>110283810
inb4 he accuses you of being debo

File: 1664886200471889.jpg (47 KB, 555x563)

47 KB JPG

>>110283837

>>110283864
Seeing you post here and mald makes me happy. You're so defeated she you don't know how to cope with /sdg/ being dead. I love that anons call you disabled too.

>>110283810
That's because Debo's manner of posting is highly affected, always. I'm not saying he's red in the face with anger whenever some argument about him happens, but if someone is *always* nonchalant, never even allowing themselves to *appear* like they're getting heated in a discussion, then they are probably deliberately cultivating an appearance of superiority and high-mindedness. To mistake that for the real quality of his mind, rather than a kind of narcissism, is naive.

With that said, I'll concede that you are right that the other anon has no grounds to accuse Debo of seeming stressed. Even if one thinks he is, he is very cautious to never *seem* so.

>>110283837
>maybe also psychedelic drugs
lol. since you're willing to admit that you're not sure on that, ill forgive you and clarify.
it doesn't make you gay lmao
if anything it demystifies porn, unless you're one of those hopeless gooner types, i can't speak for them.

>>110283810
This, he posts his gens and some "anon" always attack him. That "anon" has a palpable and resentful anger in his words. I would also like to know what the fuck happened to make him so bitter.

>>110283884
>but if someone is *always* nonchalant, never even allowing themselves to *appear* like they're getting heated in a discussion, then they are probably deliberately cultivating an appearance of superiority and high-mindedness. To mistake that for the real quality of his mind, rather than a kind of narcissism, is naive.
not debo but lmao I feel called out

>>110283837
you sound closeted anon

File: a.png (1.55 MB, 782x992)

1.55 MB PNG

still need some practice but i'm getting closer to what i want using bounding boxes. first time i've managed to force it to give me this slightly upwards camera angle i've tried many words/phrases to get

>>110283904
I follow a guy on Twitter who is quite vocal about how he became gay after a trip on either acid or mushrooms, I forget which. I remember when it happened, it was like a switch flipped and suddenly this guy that never gave any indication he was gay was now openly gay. Of course you can argue this was latent in his psyche and he just didn't recognize or express it, and the psychedelics simply opened him up. Who knows. But there are real testimonies of people who didn't "know they were gay" until they took psychedelics.

>>110283948
I think someone closeted would have a hard time being that honest about their ability to grasp the appeal of gay porn. But who knows, maybe I'm just one acid trip away from becoming the bottom I was born to be

>>110283938
Don't worry man we're all narcissists. It's the age of social media.

>debo fishing
We still remember you crying

>>110283948
no, he's just gay. he literally called gay porn hot, describes his favourite kinds of gay porn and gives a specific example by name.
he's far enough into it that he doesn't realise none of that is at all normal for straight men to say

File: debo_pc_km_00007_.png (3.82 MB, 1792x1076)

3.82 MB PNG

>>110283884
>the fact that you're not bothered shows how bothered you are
interesting ruleset
>eat bait: lose
>don't eat bait: lose
so whats my wincon then?

the honest truth is that its impossible for be to be bothered by the anti-debo schizo at this point. he's been on the same grind for years. how could I be affected by something so pathetic? at most, I feel bad for him sometimes. I hope things get better for him someday

this will be my only comment on the topic tho

>>110284014
>the fact that you're not bothered shows how bothered you are
Oh wow, this is some Debo-tier comprehension. I wonder who's replying to me.

>>110283980
if you've never done psyches then i would take what anyone says about it with a grain of salt
>>110284004
yeah lol
acid would clear up his self-unawareness

File: Moss-0102.jpg (673 KB, 1664x1248)

673 KB JPG

>>110283837
>the best gay porn
>To be clear, I am not gay
Can rate gay porn and offer nuanced qualitative comparisons on something he has zero reason to ever watch.

>>110284014
he's started accusing me of being you as well now because i didn't get mad at his bait, lol
i wonder how many people he thinks are debo

It's really sad to see the state our disabled little gobblin is in

>>110284038
It's called having interest in media outside your comfort zone

>>110284004
>gives a specific example by name
The specific example was a big meme on Twitter, and other examples in my mind were famous tv shows like Heated Rivalry (which I haven't seen so maybe it's more lovey-dovey I don't know), or other memes like the show you who's the boss in this gym thing. Which, to be fair, I didn't find that last one erotic at all, and I don't think anyone did, but it seemed nonetheless like an example of gay porn genre tropes, which exist for a reason.

As for the specific fantasy I described, I was very into MTF gender-bending porn for a while which fell into distinct types. I was more interested in the idea of sexuality as a foreign compulsion, and the body as a material object; that this could be the site of involuntary changes was therefore an erotic idea, in the way all other kinds of body horror transformation stuff is erotic. But when you are in those threads on /d/ you notice that you are sharing a thread with many autogynephiles, people who post dogshit aspirational comics of a man putting on a dress and becoming a beautiful lady, and respond to these comics with "this makes me so sad :(", also some gay ones, whose idea of hot was "the schlubby nerd becomes a babe and then has sex with the guy who used to bully him". I had a lot of exposure to the bad porn made by and for both of these types of people because I was in the same threads looking for a different sort of high.

I don't want this to become excessive navel-gazing but I think it's important to understand the sexual side of the internet is a very big place and being any kind of pervert at all puts you in contact with a lot of different kinds of perversion, you do learn to understand them.

>his two troll threads fail so he instead he spends his time sliding the real thread
Pathetic

>>110284109
glad we can all finally agree that anistudio belongs to the op

>>110284085
And most importantly I want the gay smut in this thread to improve. I am offering advice. I think no matter what sort of content people are going to be posting (and I have no power to change that) they should be doing it at the highest level and posting content that is compelling, even if by compelling we mean more potently homoerotic. Better content attracts smarter posters, and smarter posters mean better discussions, etc.

>>110284085
sure, i'm into some weird stuff as well, stuff that is on more niche sites that host lots of other weird stuff, so i've also learned about many things i'm not into.
the difference here is that you'd said you're into some of it [gay stuff]. straight men are not into gay stuff, not just neutrally either, it's actually repulsive.
i'm not calling you repulsive btw, you do you, all i'm saying is that you aren't straight.

>>110284126
dont worry anon will fix on next bake like always :]

File: ComfyUI_temp_qhoij_00004_.png (1.27 MB, 832x1248)

1.27 MB PNG

>>110283176
>Use --lr_scheduler warmup_stable_decay --lr_warmup_steps 0.1 --lr_decay_steps 0.2
Trying this now.
>4200steps (~70epoch), 1e-4 learning rate, alpha 2.
is picrel at epoch 42. strength 1.5

>>110284085
>>110284127
ok ok i understand
you're not gay
you're a gay jew
now fuck off you meandering weirdo

>>110284126
Indeed... he he.

>>110284163
You sure about that faggy-boy?

So why does he spend all his time shitting up the thread when the rentrys are present in OP? If he doesn't like it, why doesn't he just go to one of the two threads up without the rentrys?

>>110284216
>why doesn't he just go to one of the two threads up without the rentrys?
Oh jk mods just deleted those two :] bwahahah

who is he talking to

>>110284224
somehow he doesn't realise he is the troll

>tries for years to get a two links at the end of OP removed for good
>always fails
You would call this insanity, no?

>nigbo spam
not a great look

>>110284268
nice

>>110283810
Are you still not old enough to know namefags or those who are "known" for posting recognizable images often post as anon and in the third person in attempt to make themselves look better? It's a very common occurance on 4chan.

>the guy who spam threads without the links that always die and get deleted by mods isn't the troll guys the troll is actually the regular threads with links in them that everyone uses
kek

>>110284289
>Are you still not old enough to know namefags or those who are "known" for posting recognizable images often post as anon and in the third person in attempt to make themselves look better? It's a very common occurance on 4chan.
and? how do you know which posts are them if they are anon, retard?

>>110284328
I'm not the one saying some namefag has a clean record because I don't understand anonymous websites, anonie. No need to start getting upset.

ffaze is always so hostile when anon seething. Too stupid to use variation in his post.

>>110284286
sounds like it's you who has a problem with being called gay, anon. i haven't said anything about being ashamed by it.
"gay" is a pretty broad term, it can effectively refer to anything not completely straight. these days that's kind of split off into "queer", but i'm a bit oldschool and still consider gay and queer to be the same thing.
being even a little bit into gay porn is enough for me to not call you straight, even if you haven't practised anything. the interest alone is what makes you not straight. you don't realise how averse straight men are to seeing any kind of gay stuff because you've simply never been a straight man.

>>110284286
exactly how a secular subversive jew speaks

Local Diffusion

File: 1621575761655.png (42 KB, 472x472)

42 KB PNG

>>110284403

I do apologize as I am sure that this has been asked many times before. I'm looking to get into video generation, however I am unfortunately a Vramlet.(5070 12gb card). I also have a 128GB Unified Memory PC. However, the generation speed on that is unbearably slow. I remember when I was using Wan 2.2 14b at ten steps, I gave up after 71 minutes trying to generate a 5 second clip.

Do I just need to stop being poor? or is there any potential for me to generate decent videos at reasonable speeds?

File: image (42).png (714 KB, 642x594)

714 KB PNG

>>110284403
you can thank the men in lingerie poster

>>110284417
>Do I just need to stop being poor?
basically, yea

>>110284417
Use minimax h3 its the best shot you have

>>110284376
>consider gay and queer to be the same thing
Off-topic, just lurking, but reminded me of this
https://www.youtube.com/watch?v=i6wu81BEVBI

Go back to talking about moss training instead of gayshit tho

File: debo_pc_km_00013_.jpg (1.27 MB, 1792x1076)

1.27 MB JPG

>>110284443
but how can anons know if they're gay or not without a rigorous and spirited debate?

>>110284421
they are locally diffused thougheverbeit

>>110284368
who?

Anon is investing in his vacation fund.

>>110284464
Do you post here because no one talks about image generation in sdg anymore

I need to ask the dumb question
a 16gb vram GPU and like 48 gb of ram
is there a minimax h3 model that I can run?
I think I gotta use one of those pruned ones but not sure which exactly
anyone have an idea?

>>110284554
comfy should automagically use the int8 version which runs on your system

I wish I was paid for training models and generating kino. I hate wagecuck life.

>>110284443
>training with moss
Isnt moss base already trained on shit-tier anime slop by a non-anime team, poisoning any further training?

>>110284565
is this with the portable or desktop version?
i'm just using the portable
my outputs have been somewhat noisy so I'm not sure if it's the config or lack of ram or something

>>110284430
Isn't that generally considered the best local video model? I figured it wouldn't run at any level of appreciable speed with only 12gb of VRAM (assuming I am also using LoRAS to reduce steps and the like).

>>110284586
That's not how training works anon. You think SD1.5 SDXL didn't have shit tier anime slop in their dataset?

>>110284598
>You think SD1.5 SDXL didn't have shit tier anime slop in their dataset?
Uhhh, no I'd say no? Moss was trained with synthetic GPT images from what I heard, SD1.5 in that era would have zero of those.

>>110284609
>Moss was trained with synthetic GPT images from what I heard
?????????

>>110284589
It can run on even 8gb, it's a matter of hoe patient you are at that level though. From my experience it runs better than WAN at least.

>>110284586
>>110284609
You should at least place at the technical report before saying things that are inaccurate.

>>110284609
>Moss was trained with synthetic GPT images from what I heard,
The model was designed to beat GPT at layout and text generation, you don't beat a model by training on its outputs. Read the technical report desu.

>>110284611
Didnt they state they excluded ai generated images from their danbooru dataset, but that might've missed some? Also just scraping random images off danbooru is not how a competent team would go about it I believe.

In order, how much semen i've spilled per model:

#1 - MiniMax H3
(1 billion gallons)


#2 - WAN 2.2
(500k gallons)

#3 - illustrious
(200k gallons)

#4 - Moss/Krea
(50k gallons)

>>110284628
#1
SD 1.4/1.5
(Oceans)

#2
Everything else.
(Dry, a waste land, nothing left)

File: images (3).jpg (14 KB, 360x360)

14 KB JPG

i made the model make a woman's vagina

>>110284622
I didn't read everything in the report so I could've missed that, but it's clear they looked at what the Z team did in regards to synthetic data considering the models ability to "look so real" imo.

>>110284666
cheeky.

>>110284554
i ran minimax on 8gb vram 16gb ram so yeah you can. use Pinokio if you don't want to bother with technical stuff

>>110284628
i cannot stop generating pics of my highschoop crushes after training loras on them. with moss ive even been using multiple of my crushes loras together and instructing it to combine all the faces into one and ive completely fallen in love and will never have a real girlfriend again

>>110284646
yes, i too remember when bstaber for sd 1.5 was released and being quite happy, but that stuff is nothing compared to what's possible now.

>>110284681
training settings?

>>110284687
Just ask codex or claude to set it up for you (after it does some research on its own of course)

>>110284700
I'm asking you anon, not claude.

>>110284677
Are you using the same prompt as your XL gens that were like this? Kek I really liked those for some reason.

File: moss-155179961054003_00001_.png (932 KB, 1024x1024)

932 KB PNG

>>110284714
It was easier to just write up some new prompts on the spot, but yes this is one of my favourite of all my old gen ideas too lol

What can Moss do that Krea can't?

>>110284622
>Also just scraping random images off danbooru is not how a competent team would go about it I believe.
methinks a large part of the animu pics in moss come from one of or multiple researchers personal stash. i have nothing to back that up but it feels right.
i try to gen no prompt pics and it just gives me loli kek

>>110284744
>NSFW without lora
>NSFL without lora
>bbox prompting
>looks like real life without a lora
>makes me cum easily
>more receptive to training
did you need any more

>>110284744
nb4 sasori replies with "it can't look exactly like how my krea loras look when i use my same captions and params so it is bad" kek

It's alright if you dont care about anime moss is a win for you, but don't think its a good base model if you're waiting for the next illustrious

okay schizo

>he was just larping

Reply to me with the best gen you've ever seen posted by anon

>>110284800
I like this one

Why would I be waiting for the next illustrious when Anima is already here?

>>110284803
I have 14k images in "I should save this slop" folder, how would I pick one? (To be fair, a decent chunk of them are just coom subject to much lower standards.)

>>110284824
Am I supposed to take that bait?

are there any sites where you can get character loras for krea 2? civitai dont seem to have any celeb ones. I tried prompting before anyone asks.

>>110284831
>how would I pick one?
Whichever has the most kino and the most sovl, it should stick out like a sore thumb.

>>110284831
>I have 14k images
i wonder how many are mine...

>>110284833
I can train whatever I want so I don't need illust

>>110284886
10 loras and its still worse than NAI

>>110284744
filter those with issues of skill

>>110284835
Why use krea with two loras one for the person and one for NSFW and layer coping when you could just use one lora for likeness with moss and ejaculate with ease?

File: ComfyUI_temp_dphqy_00002_.png (1.35 MB, 832x1248)

1.35 MB PNG

>>110284168
>Use --lr_scheduler warmup_stable_decay --lr_warmup_steps 0.1 --lr_decay_steps 0.2
>32/32
epoch 21, still had to crank it to 1.5, I'll let it cook to 42 and test again.

>>110284906
Tell that to >>110284786

>>110284835
Huggingface but have fun finding shit.
A few trainers who used to post them on civit still post them in other places. Maybe try search user names you see posting such loras for older models on civarchive or whatever. I am not going to handhold you more than that.
Dig 4chan archives and see if someone catboxed lora for your celebrity.
In any cases, the average lora is still as shit as they used to be before they got booted from civit, so training your own is still your best bet.

>>110284464
Thank you for your valuable contribution thread schizo.

>>110283411
>Trigger word is added by sd-scripts.
not sure if this is the "proper" way or any different than what you do but i make the prefix (not trigger word) "Art by [name]" or "An image of [person]" (for characters). Then, I turn the captioning models thinking off and have it continue the caption from my prefix.
im really not sure why you're having to bump up the strength to 1.5 to see the style apply i hope you figure it out fren i wish i knew the answer

now that anistudio is officially endorsed in the OP, does anyone have recommended settings for low VRAM?

try again but this time dont make it so obvious

>>110284750
How can safetymaxxed models even compete with cunnymaxxed models? Is not even a competition.

I still prefer Krea2 due to his superior size. BIG Moss, when!? We need a BIG Moss Model. We need to give our 1girls the BMM.

File: 1792153130330709.jpg (558 KB, 1024x1504)

558 KB JPG

>>110284803
Not sure if the best but I really like this one.

>>110285140
>2hu shit in 2026
at least post Reimu's pits

>>110285231
Touhou is forever.

>>110285072
kek, now I want to see him in a Bollywood dance number

>>110284800
>>110285030
>>110285047
can moss generate these in the style of 90s crpg portraits

>>110285276
Just try it debo.

>>110285276
No idea, maybe I will try.

But as all genners know, it's impossible to totally extricate the subject matter from the style. They are linked.

We will see.

>>110285140
Moss

>>110285279
im not that guy, just a random dude who often makes it into the collage (kind of a big deal)

File: moss-854467350834942_00001_.png (576 KB, 528x800)

576 KB PNG

>>110285276
Working on it, trying to figure out how to squeeze that 1970s movie star extravagance into a very different visual style.

File: moss-249441700665347_00001_.png (734 KB, 528x800)

734 KB PNG

>>110285450
But as for "can moss do 90s crpg portraits" yes, trivially. And the more conventional your subject matter the easier it is.

File: moss-664749289897987_00001_.png (716 KB, 528x800)

716 KB PNG

>>110285450
(while writing the text of this post, I discovered I could simply describe the character as having 'sumptuous dark hair like a 1970s bollywood movie star', resulting in the image that I attached to that post. Sometimes trying to describe your problem points you to your solution.)

>>110285085
Desu bigger size =/= better the chinks already proved that scale at all costs is a dead end.

>the big announcement was just their dev channel
https://x.com/ComfyUI/status/2119862775971566041

>Launching
>@ComfyDevs

>Our dev-focused account.
>First 50 people to quote, engage with the post, and fill out the form linked below the post will get $50 in comfy credits!

seriously? does "dev" even have any meaning anymore when ai agents can do everything?

Any custom node that can torch compile int8 moss currently?
Everything I have errors out.

>>110285561
You'd be surprised at how retarded some are that even the do-everything-for-me machine cannot help them

>>110285561
lol

>>110285599
This. Artificial intelligence doesn't work without someone intelligent operating it.

File: ComfyUI_temp_qpebe_00015_.jpg (94 KB, 352x544)

94 KB JPG

>>110285450
>>110285469
>>110285497
i tried myself and am having fun with the "sumptuous dark hair" prompt

>>110285561
>>First 50 people to quote
why did they say people when its only jeets whod do that

>>110285625
Here's the full prompt I ended up at, although it still needed work. Some of the description was provided by ChatGPT, which I had prompted with,
>Describe, plainly and not poetically, the visual appearance of a typical CRPG portrait, e.g. from Baldur's Gate II: Shadows of Amn; as if you were looking at a picture of it, out of its context. Let's say for this example that the character in the portrait is Mel Gibson's Mad Max.
>Be as concise as you can. This should not be an exhaustive description, instead it should be a dense impression.
It answered:
>A tightly cropped, semi-realistic painted portrait of a rugged man in worn black leather, shown from the chest up at a slight angle. His face is stern, weathered and dramatically shadowed, with tousled dark hair and rough stubble. Muted, earthy colours, warm skin tones, soft brushwork, strong contrast, and a dark, indistinct background. Detailed but visibly hand-painted, with slightly idealised features.

And my prompt ended up as:
>1998 crpg character portrait of Abhimanyu, young warrior-prince of Rajasthan. From Black Isle Studios, published for Windows, based on a D&D campaign setting. His ceremonial breastplate glimmers, adorned with bright woven damask. A tightly cropped, semi-realistic painted portrait of an Indian prince of Rajasthan, shown from the chest up at a slight angle. His face is androgynous and beautiful, dramatically shadowed, with sumptuous dark hair like a 1970s bollywood movie star. Rich bluish colours, warm skin tones, soft brushwork, strong contrast, and a dark, indistinct background. Detailed but visibly hand-painted, with slightly idealised features that are distinctly Indian.

Although in the examples I was posting, it said "manly and beautiful", not androgynous.

>>110284750
You're theory is probably true. I noticed that MOSS responds much better to lora training with anime/manga than Krea2.

if only i could post the stuff im making
you just dont understand

thank god moss didnt filter me id probably kms if it did

>spend entire day off constructing the perfect goon vid
>succeed
>goon
worth every spark of electricity

>>110285743
i doubt it's any good anyway

>danbooru is still down
So uh.... wtf am I supposed to do?

File: moss_image_t2i_00759_.png (1.83 MB, 880x1184)

1.83 MB PNG

>>110285900
Gather your dataset from gelbooru, rule34, sankaku etc. instead.

>>110285922
But the danbooru wiki pages :( I haven't found anything like it on the alternate sites

>>110285885
>it was some slop strip video

File: 1758136630090480.jpg (340 KB, 800x807)

340 KB JPG

came here to post this
>>110283618

>>110285951
what is that?

>>110286070
It tells you a bunch of information about whatever tag, related tags, tags that should not be used together, etc. It's really handy when prompting a model trained on booru. Most of the time I'm prompting 1girl I have it open for this reason.

e621 has something similar but whenever I've looked at it in the past, all the normal tags are void of any information while all the diaper, scat, and fetish stuff has plenty of info unsurprising.

>>110282338
Man H3 i2v is good, is this the best model I can reasonably run locally with a 5080? I feel like I live in the future being able to gen this at home, have been out of the loop for a bit

>>110286146
>It tells you a bunch of information about whatever tag, related tags,
Autocomplete plus is a partial substitute.

>>110286158
It's the best local vid model yes.

>>110285885
>won't even catbox the vid to show anons without getting the hammer
for shame

>>110286158
h3 ref2va is the best local video model and nothing else is even worth talking about. i dont see it being topped for a long time.

>>110286146
i see, but do you have an example of incompatible tags? i can't think of something besides trying to mix things like big butt and small butt i guess

>>110286208
I would if the site wasn't down kek. There are some mutually exclusive tags, like what you mention. The only example I can think of are the monochrome, greyscale, restricted palette, spot color, etc tags. Someone unfamiliar with these (me before I got into 1girl genning) could easily mix them up or not know which ones would work well together or which are mutually exclusive.

There are other more nuanced examples but I can't remember any off the top of my head.

>>110286255
cute

>>110286242
>monochrome, greyscale
are these incompatible?

File: kromoss.jpg (3.22 MB, 2512x3344)

3.22 MB JPG

Kromoss gives significantly better results when it comes to loras

>>110286305
did you train it on moss or kromoss?

>>110286315
kromoss

>>110286282
Monochrome is when the image is any single hue and its shades and tints, so not only black. Greyscale is only black and white and shades in between.

Every greyscale image is monochrome but not every monochrome image is greyscale.

>>110286305
are you using the kromoss approved workflow?

>>110286339
no

>>110286305
>kromoss
What benefits does it bring tag-wise? I seem to remember his previous tunes not including artist tags but maybe I'm misremembering. I'm just wondering what sorts of things you can get on kromoss that you can't with the base model. Also which version are you using?

>>110286344
it hasn't lost anything, I found it to be pointless until I start4ed training loras off of it

>>110286344
>I'm just wondering what sorts of things you can get on kromoss that you can't with the base model.
if you liked chroma, then it has the same training data, which means (mostly) better furry nsfw content and nsfw in general. ie, it can do really good anal sex and stuff

also he released another update 21 hours ago
https://huggingface.co/lodestones/Kromoss/tree/main/diffusion_models

kromoss_int8_convrot_OPD-priming

>>110286331
i know, but it would be strange if the prompt following quality was reduced if you mixed those together since they should be technically compatible

>>110286352
sometimes moss gives me anal even when i prompt for vagene sexo kek i hate anal

>>110286370
what is opd-priming and why should i care

>>110286389
according to lodes discord, it has better support for bboxes and more training. of course, the damn furry rarely explains anything

File: ComfyUI_temp_kxjzx_00001_.png (1.13 MB, 832x1248)

1.13 MB PNG

>>110284953
So I let it run all the way to epoch 70.
Looks like I just had to be patient, this is strength 1.
Line art seems a bit fucky, I might try a run at 32/16.

File: screenshot.1794267668.jpg (193 KB, 781x765)

193 KB JPG

moss maxing

>>110286373
It has happened previously for me when I was a young buck, and I'm sure others have encountered this, where before you learn the difference, you add both any time you want a greyscale / black and white manga page. You'll add monochrome and end up occasionally getting red monochrome or whatever because those are included in the tag.

>>110286398
>better support for bboxes
Intriguing but I'd be curious to see some comparisons

>>110286403
what res is your dataset or rather what buckets are you using

>>110286440

INFO     bucket 0: resolution (832, 1152), count: 1                                                                                   dataset.py:808
                    INFO     bucket 1: resolution (1024, 896), count: 1                                                                                   dataset.py:808
                    INFO     bucket 2: resolution (1024, 1024), count: 55                                                                                 dataset.py:808
                    INFO     bucket 3: resolution (1088, 832), count: 2                                                                                   dataset.py:808
                    INFO     bucket 4: resolution (1152, 832), count: 1                                                                                   dataset.py:808

>>110286352
Ah, thanks. Maybe the next time I want something more lewd than post marriage missionary while kissing, then. Shame it doesn't support more artists.

1girl

1girl
c word
cctv footage
straddling on couch
creaking couch sounds
soft moaning

File: anima_aestheticV11_00941_.png (1.98 MB, 1672x1144)

1.98 MB PNG

>>110286461
you are like baby
watch this

832x1152: 27 files
896x1152: 13 files
832x1216: 11 files
1152x896: 4 files
1024x960: 6 files
1024x1024: 13 files
768x1280: 5 files
704x1408: 1 files
1152x832: 5 files
960x1088: 3 files
768x1344: 6 files
1216x832: 2 files
1088x960: 2 files
704x1344: 1 files
896x1088: 2 files
960x1024: 1 files
1088x896: 1 files