Blessed thread of frenship

i ejaculated 5 times today

>>110116280
>>110116311
You're part of the problem.

File: MossHR_27144_.jpg (2.15 MB, 4096x2304)

2.15 MB JPG

bbox prompting is great

File: 1790909890933.jpg (325 KB, 1408x2128)

325 KB JPG

moss is so good bros

>>110116331
>>110116341
That's what ketamine must be like.

File: kek.png (1.22 MB, 925x673)

1.22 MB PNG

progress

>>110116322
Relax anon
It's a blessed thread
Of frenship even

File: moss-image-sft-dit-int8-c(...).png (1010 KB, 768x1152)

1010 KB PNG

>Prompt executed in 16.76 seconds
moss is saved

File: moss-image-sft-dit-int8-c(...).png (2.39 MB, 1728x1152)

2.39 MB PNG

6 steps

File: moss-image-sft-dit-int8-c(...).png (2.39 MB, 1728x1152)

2.39 MB PNG

>>110116489
8

>>110116323
Nice legs mosskek

File: MossHR_27158_.jpg (2.07 MB, 4096x2304)

2.07 MB JPG

>>110116323
High res pass mistake

>>110116544
Damn, we truly never had a model capable of such a gen before!

genning is literally Plato’s allegory of the cave

>>110116323
>>110116544
Why does everything besides Miku and the bird look like it's made out of mashed potatoes?

>>110116661
{
"label": "object",
"caption": "A dramatic vertical sunset sky filling the upper half of the frame, transitioning from deep violet at the top to fiery orange and soft peach at the horizon.",
"bbox": [0, 0, 2304, 2100]
},

File: 2140001-Style Academic Or(...).png (3.85 MB, 1536x1536)

3.85 MB PNG

Luv moss, simple as.

>>110116489
>>110116496
looks like my older sister when she was younger

>>110116790
that's what i was going for

File: download (1).jpg (1.09 MB, 1600x2560)

1.09 MB JPG

>>110116199
Kroma pretty good

File: 2151001-Style Academic Or(...).jpg (1.91 MB, 1824x1248)

1.91 MB JPG

>mfw looking through hundreds of wildcard moss gens I had queued up the night before

>>110116790
I doubt she was that pretty but good for your genes anon

how much garbage do you get per 100 moss gens?

>>110116874
101 give or take

File: 2026-10-19-16-30-18_00001_.png (2.47 MB, 1920x1088)

2.47 MB PNG

slag time

File: 2212001-Style Academic Or(...).jpg (1.84 MB, 1536x1536)

1.84 MB JPG

>>110116874
On my wildcard setup it's like 90% garbage which is actually really good compared to other models. The good ones are a league above other models. It's obviously not a flawless model by any means, but it does the style I like much better than any other model by a lot.

File: ComfyUI_Moss_Kroma_00024_rh.jpg (1.68 MB, 2048x2048)

1.68 MB JPG

Moss will be here to stay because it's uncensored and understands most stuff better. Even if Krea 3 is better, it will serve as img2img for Moss, similar to Krea 2.

>>110116869
i said my sister, not me. i got hit by every ugly branch on the tree.

>>110116489
>>110116496
dem i miss late 90s early 00s

File: F3K_Distilled_Edit_151535.png (2.73 MB, 1728x1152)

2.73 MB PNG

>>110116496

>>110116967
this makes me feel bad for him but then i remember he's an autistic psychopath

>>110116680
>>110116675
>>110116341
>>110116331

is this the basis for a system of government?

>>110116995
yes the watery tart handing out swords is famous for certain a monarchist governmental system

File: image (74).png (1.68 MB, 880x1184)

1.68 MB PNG

https://huggingface.co/lodestones/Kromoss
he released new checkpoint

>>110117137
Fire placement seems a little dangerous.

who are you talking to?

>>110116930
that sounds promising, thanks anon

File: 3400b862-8280-4d79-928a-5(...).jpg (1.21 MB, 2560x1440)

1.21 MB JPG

>>110117158
Ah, it'll be fine, it's a cold climate

>>110117017
I'll try it out. There's weird plastic texture added by the DMD LoRA, hopefully this fixes it

>>110117320
jfc that looks just like my sister wtf

>>110117320
>>110117390
i bet if i sent her this she'd ask where the fuck this was taken and how i got it

https://desuarchive.org/g/thread/110111565
LOL! :)

File: kromoss_bf16_00013_.png (2.14 MB, 1152x1728)

2.14 MB PNG

>>110117397
kromoss won

File: F3K_Distilled_Edit_51515.jpg (3.09 MB, 2128x1408)

3.09 MB JPG

>>110116675

File: kromoss_bf16_00014_.png (1.94 MB, 1152x1728)

1.94 MB PNG

i like how theres the detail of someone having taken a photo of a photo

>>110117431
>2023
why would it look like that

>>110117425
heh

>>110117445
Old camera.

>>110117445
Because the only thing the Kromoss finetune can generate is extremely blurry images

They took Moss base which already has a blur bias, and then 'hey, let's make this ten times blurrier'

File: kromoss_bf16_00021_.png (1.88 MB, 1728x1152)

1.88 MB PNG

>>110117445
probably never heard of scanners
silly zoom zooms

>>110117509
lel

>>110116984
Where'd you get that picture of my house

>>110116199
man those are some hot futa girls, OP

>>110117654
Thank you.

Moss is highly trainable, just have to tweak the default training params, so far we cookin, but we still do get that spaghetti limb now and then, just have to shake up the prompt, definitely can sit alongside Krea, but clearly I am still gonna wait for the next model that does everything without plasitc censorship.

File: iter_0.png (189 KB, 512x512)

189 KB PNG

>>110116199
"a cute cartoon walrus"
seed=1
zit.

>>110117847
kek

File: F3K_Distilled_Output_5351.png (390 KB, 512x512)

390 KB PNG

>>110117847
idk if you actually generated it at 512x512 but I tried same seed, actual 512x512 on F3K

>>110117978
um... ok?

>>110117987
i answered a random post about zit seed 1 at 512x512 with an equally arbitrary anecdote, what's wrong with that

>>110117818
lol noice gen.

>>110118020
why were u skeptical about him generating a 512x512 image?

File: iter_1.png (163 KB, 512x512)

163 KB PNG

>>110117847
>>110117847
Now in perfected form after lots of computing (based on some paper I vibecoded)

File: mpv-shot0252.jpg (159 KB, 768x1024)

159 KB JPG

>>110116937
I'm just hoping krea3 gets a local release soon

>>110069910
I tried to upload it but the link gives a blank page? am I banned? uploading the .json file gives a 0 byte link too.

File: krea 2 style pose (back).png (758 KB, 991x1059)

758 KB PNG

>>110118075
Did they fix the humeri?

>>110118078
i look like this

zitjeet is still here seething over krea? if he just bought a better gpu back then it would be worth double now. what's with all the poorfag sabotaging local?

>>110117996
You stole my Knightformers idea I had in my head but never got around to because I'm cooming instead, I hate you

What do you use in your prompt to make a person into its anime version withoiut the ai seemingly just tracing stuff?

Is the Moss shill campaign over? It has been really tiresome those last few days

>>110118386
Krea 2 is the overall best image model I've used, but Moss is the most fun model I've used, it has so many styles and you can get so much cool looking variation out of a single promt.

I think most people are kind of waiting for best training practices to be fully figured out so there's good quality loras released.

is there such thing as training a lora based on gens that you liked better than others for the same prompt? so that you can promote those neurons. i don't think it's the same thing as just re-running the generated result into a regular training regime

moss is awesome if you like training stuff and see how it looks on it. If you are just a tourist then don't bother with moss.

>>110118542
definitely a lot better than anima trash out of the box. there's plenty of comfy shills seething here already. cumfart's not gonna get good roi on turdrussel's turd

>>110118482
Sure, people have been doing that for ages, particularly for visual styles. Usually the problem is that generated faces are very similar (some models worse than others, Flux models probably the worst) so if you train on people, the lora will learn the facial / body details and bleed those into all images, even if you combine with loras of real people.

Also since the output is synthetic, a lot of the real world 'noise' that makes training harder is gone from the images, meaning it will learn much faster and also overtrain faster, so keep that in mind.

>>110118589
>Also since the output is synthetic, a lot of the real world 'noise' that makes training harder is gone from the images, meaning it will learn much faster and also overtrain faster, so keep that in mind.
Isn't it more like it will learn that all those little mistakes that AI makes are actually correct and it will consequently produce them even more? Hence why you shouldn't train on synthetic images.

>>110118589
are you sure you are talking about the same thing i am thinking of? i think the closest thing would be reinforcement training. basically, i would run a forward pass to see what kind of output i get and then if i like it, the weights that produced that are strengthened. maybe it's like a forward pass training method. however, i should only be focusing on a small subset of the weights for it to work like a lora. otherwise, i am just making an entire finetune of the model

now that the dust has settled, is MOSS the new local SOTA?

>>110118075
what's her name?

>>110118581
Yea, i'm glad that it exists. I stopped using anima and hopped over to krea and was happy about it but got a bit tired of the inflexibility and turbo artifacts. Moss is an awesome replacement for anima and if the devs keep their word and do a full anime finetune then its game over.

>>110118672
>>110118581
but keep in mind that the first anima preview version was literally the best anima version there was. Hope the moss team manages to keep a healthy balance between moss pretrained knowledge and the anime knowledge added. Because whenever you go full retarded danbooru overfit a model automatically becomes dogshit.

>>110118600
If the exact same mistake is repeated over and over in the images, like weird teeth, anatomy, sure. But then it's a bad model to use output from to begin with.

But seldom repeated stuff, not really, there's a lot of 'weird' things for the ai in real world images as well, which it will misinterpret and which will cause problems if repeated often enough, hands and poses with overlapping limbs will still cause problems at times even though most models have been overtrained on smaller subsets of easily readable poses and hand gestures to mitigate this, leading to the 'stock pose syndrome' of ai image models.

Every large scale image model out there has lots of synthetic images in its data set. When you train loras, if you have real world images of whatever you want to train, that's always the best, but sometimes synthetic is your only choice if you want a decent size data set.

File: moss_image_sft_int8_convr(...).jpg (689 KB, 1504x2048)

689 KB JPG

>>110118657
After finetunes

>>110118745
pedophile.

>>110118766
technically...

>>110118766
You're an idiot anon.

File: moss-image_02420_.png (997 KB, 1216x832)

997 KB PNG

just rawdogging moss base

File: moss-image_02421_.png (1.01 MB, 1216x832)

1.01 MB PNG

>b...baka hentai!

>>110118817
why is there a butt on her crotch?

>>110118832
you mean the FLUX CHIN on her crotch?!

Is moss better than krea 2?

>>110118852
yes IMO it is, it's a smaller model so sometimes the output is finicky and requires better prompting, but it just looks better. The advantage of Krea2 is that you can stack LoRAs practically indefinitely to compensate but that can also be tiresome at times.

>>110118832
girls have an extra butt instead of a penis, that's what makes them twice as sexy

>dmd lora at 0.5 weight, 15 steps, cfg 3-5
makes moss base pretty stable, still fast and output is variable and dynamic.

>>110118896
why did no one tell me this before?

>>110118909
Don't you lose the 'turbo' speedup if you go above 1.0 for CFG ?

>>110118928
Yea but you really want to have negs with base and 15 steps is not that slow in int8. At CFG=1 the dmd output is also extremely desaturated. You go higher and immediately colors and textures pop again. I use the dmd lora more like a stabilizer and steps-distill.

>>110118852
For local training yes

>>110118935
Agreed, the extra control you get with negatives and prompt strengthening through (prompt:value) is really worth the extra time

>>110118959
Still waiting onetrainer support

you know that moss is instinctually recognized by the community as "sdxl 2" by the way all the amateur photo style loras are popping up immediately on civitai.

>>110118997
SDXL finetunes was king for anime, Flux 1 was king for photo, then Z-Image Turbo killed Flux 1, Krea 2 killed ZiT.

Now it seems Moss might be killing Krea 2, will it finally kill the SDXL finetunes as well, they seem indestructible ?

>>110118997
>by the way all the amateur photo style loras are popping up immediately on civitai.
Which is weird since really none of them look any better than what you get out-of-the-box with just turbo-lora and base.

This is a pro moss board, make sure that's your final negative comment on this board.

>>110118823
>>110118817
Dude... is that.. is that a 1girl.... kneeling??!! Holy shit novelai is kill we have endgame at home!!!!!!!!!

I tried comfy's moss template and the results were terrible compared to the recommended workflow by MOSS' official
This is why you shouldn't trust him and pre release workflow yourself.
if only BFL had known better

File: ComfyUI_06928.png (1.21 MB, 1424x800)

1.21 MB PNG

filtering prompts through an llm is really useful. for this one i kept the character description entirely unchanged, even had a second character description left in, only changing the scene description. the llm sorted out things like the fact most of the items are on the bed instead of being worn and the other character isn't in the scene so excluded those words.
just wish my gemma was more consistent, maybe i need to make a lora instead

>>110119139
I was just joking ok, it was just a joke.

File: 1762592439548878.jpg (283 KB, 778x1172)

283 KB JPG

lately when I generate krea 2 images in comfyui I get fucked up pictures like this but only like once in 100 pics, why is that? most of them are still recognizable with the noise and I can redo them with the same settings to make then right, but this one just ended up really fucked up and redoing it results in the same ugly pic

>>110119188
>filtering prompts through an llm
how?

>>110119204
added this to pi;
https://github.com/OpenMOSS/MOSS-Image/tree/main/skills/moss-image-prompt-enhancer

then i tell it to make a moss-image prompt using that skill using the text i give it, and have it output a json which i then use as the prompt to make the image

>>110119204
>claude, make me an llama.cpp bridge node in comfyUI that support master prompt.
it's that simple

>>110119204
?

Pretty much every Comfy template image workflow the past 3-4 months have a llm they filter your prompt through

File: passporte.png (312 KB, 256x1024)

312 KB PNG

>moss
>mossad

>>110119228
which template do you use?

>>110119246
For example, load the Moss template, right click on the Moss Text to Image box, choose unpack subgraph, now you will see how it passes your prompt to a llm and 'massages' it before generating it.

You can modify the 'massaging' instruction as per your need.

>>110119267
not him but i didn't know they made a comfyui template, guess i've been using it wrong

File: 1773073007174587.jpg (139 KB, 1131x726)

139 KB JPG

>>110119280
for some reason there's 4

>>110119313
what am i looking at?

>>110119321
bumfart templates, I've never used them either

>>110119321
these are the templates you find in comfyui

>>110119328
oh, since when was there a template finding thing in comfyui?

>>110119334
for several years now...

>>110119334
button on the left sidebar ('new' UI enabled)

Are you using the ER SDE 25 step like Comfort suggests or something else?

>>110119062
Moss is like 5x slower than SDXL for me if both use 1024x1024 and 25 steps, so I'm skeptical it'll break SDXL's deathgrip on sloppers.

>>110119363
I almost forgot the Moss devs use euler and default. (I was testing out 50 steps and cfg 3 while still using er_sde and simple. 25 steps does mess up limbs more, but 50 steps is darn long.)

Btw, there's no "default" in the scheduler options. Did the template instructions mean to say "normal"?

File: moss_image_sft_int8_convr(...).jpg (1.22 MB, 1504x2048)

1.22 MB JPG

cube hatsune

>>110119449
try simple and sgm_uniform first, they are the recommended.

File: moss_image_sft_int8_convr(...).jpg (1.06 MB, 1504x2048)

1.06 MB JPG

>>110119464
er_sde / bong_tangent works fine too

>>110119449
>Btw, there's no "default" in the scheduler options. Did the template instructions mean to say "normal"?
The Moss team has published standard values for the scheduler, which differs slightly from the "normal" scheduler. I think that's what's refereed to as as "default"

>>110119473
good gen, absolutely disgusting

>>110119453
>>110119473
If I knew it wasn't gonna bite me, I could still fuck this Miku's mouth to completion.

someone prompt Dr Octopus holding /ldg/ and saying "the power of the coom in the palm of my hand"

Kromoss updated
https://huggingface.co/lodestones/Kromoss
>"extended the poopy-diaper training -lode"

>>110119082
well i have to agree. Base is really fucking good ootb. But there are also specifically loras for different camera lenses which is pretty cool. The one guy that trained lots of those had bascially the best performance with flux 2 dev and klein because of the flux2vae. Models with qwen vae and esepcially krea2 due to being aesthetically polished couldn't really deliver the raw noise that is early digital photography. With moss we have something more capable than f2klein and smaller than fl2 dev.

File: moss_image_sft_int8_convr(...).jpg (1.04 MB, 1504x2048)

1.04 MB JPG

>>110119480
>>110119484

>>110119540
moss is boss

File: kromoss_int8_poopy-diaper(...).jpg (818 KB, 1504x2048)

818 KB JPG

>>110119511

>model trained on CSAM
>its called MOSS(ad)
nice try epstein-class

>>110119606
>CSAM
chain saw anus man?

>>110119606
I never downloaded moss.

>>110119611
you're goddamn right

File: Chansawman(1).jpg (2.22 MB, 6072x9000)

2.22 MB JPG

>>110119611
>chain saw anus man?
Why yes that's my retarded brother... we uh....we don't talk about him

>>110119511
is this thing even viable yet?
Kroma was really underwhelming because he trained it on low resolution shit

I think generative AI sucks

File: you.jpg (58 KB, 640x480)

58 KB JPG

>>110119703

File: _.png (1.15 MB, 1024x1088)

1.15 MB PNG

flashbang allert

>>110119703
but llms are incredible tools though

>Once this is available in comfy-kitchen, use it for ~30% faster MOSS-image inference with int8 models:

This looks promising from kijai, so ehh, when will 'situ' be merged into comfy-kitchen ? inb4 never

I would really enjoy a 30% Moss speedup

>>110119796
ehh, forgot: https://github.com/Comfy-Org/ComfyUI/pull/17099

>>110119796
>inb4 never
>Add SiTU input activation to the INT8 ConvRot quantizer
>comfyanonymous approved these changes 3 days ago
???

>>110119813
Nice, has that japanese 'House (1977)' movie feel

still no announcement for flux 3. just as well. i can keep generating things with h3 and moss lol

I just wanted to say Moss is the best model we've ever gotten. And the fags who hate it, can't master it. Its fantastic.

>>110119837
Yes, this is merged, but 'situ' needs to be merged into comfy-kitchen as well

At least I haven't noticed any performance increase after this was merged than before it was merged

>>110119852
don't do drugs

File: 388528872084917.png (704 KB, 832x1216)

704 KB PNG

MOSS brings me back to simpler times of fucked up hands and diverse art styles.

Are the ZITFags and Kreafags finally gone? Can us MossBros finally gen in peace?

So Moss can't do img2img or editing? Why should I care?

File: moss_image_sft_int8_convr(...).jpg (1.22 MB, 1504x2048)

1.22 MB JPG

>>110119842
The Blob (1988) was my inspiration

How are you still pressed for speed when the turbo model already provides better images with base but is also faster
>>110119883
Always could you drama baiting sperg. There is no model war

>>110119867
This was merged into ComfyUI
https://github.com/Comfy-Org/ComfyUI

Comfy-Kitchen is a separate project
https://github.com/Comfy-Org/comfy-kitchen

It says:
>Once this is available in comfy-kitchen, use it for ~30% faster MOSS-image inference with int8 models:

>Once this is available in comfy-kitchen
...

File: 106004530529549.png (1.81 MB, 1728x1024)

1.81 MB PNG

>>110119890
It can do img2img, but it's not an edit model.

>>110119892
Spoken like a true jeet.
There isn't a model war, Because MOSS WON!

>>110119903
https://github.com/Comfy-Org/comfy-kitchen/pull/251
idk how this PR didn't add this "situ" into comfy-kitchen, but ok, keep waiting

>>110119892
Why lie?

>>110119915
Lie about what?

>>110119914
Damn, so there's no more speed coming

My day is ruined

>>110119915
He's a butthurt Krea 2 fag. That's why he's lying. He can't handle that Moss has won the war. He'll sperg out and cry about it, but we'll know why he's doing that. Krea 2 fags keep losing.
Moss CHADS keep WINNING

>>110119924
Schizo I think Moss is a better model, I'm not mentally ill and need to go all or nothing with free shit that can have it's niche

waifu wars on /a/ are less gay than this shit here

>>110119929
Krea2 fag in denial

>>110119931
It's one retarded faggot making a mountain out of a molehill. We have mentally ill faggots here that would make a moe thread blush

>>110119944
>just an ordinary guy...with a chainsaw

loved to train and use krea and i'd say i pretty much mastered the model (anime, don't care for realism) but the model creating community was awful. Basically just 2 retards on civitai spamming 20 movie loras a day and some nsfw here and there. Totally ridiculous.

Hope its funnier with Moss.

>>110119972
Krea 2 always sucked and we don't want your kind using Moss. Stick to Krea 2, faggot.

>>110119997
I liked that OP image too. :)

>>110119997
prompt

>>110120012
>>110120020
A young woman with long, straight blonde hair and fair skin sits in a large, plush green armchair with an orange cushioned seat. She wears a blue dress with white lace trim and a white apron over it, paired with white thigh-high stockings that extend to her upper thighs. Her legs are spread slightly apart, and she is seated with her arms extended outward, resting on the armrests of the chair. She has a pair of large, blue rabbit ears with white inner linings attached to her head, positioned symmetrically above her ears. In her right hand, she holds a black can of Monster Energy drink, which is prominently displayed with its iconic green logo and branding.

Her facial expression is neutral to slightly bored, with her eyes looking directly at the viewer. Her left wrist bears a simple white band or bracelet. The chair has a rounded, cushioned design with thick, padded arms and a low backrest. Behind her, the scene is a fantastical forest setting, filled with oversized, colorful mushrooms—some red with gray and blue spots, others with multicolored caps—and dense vegetation including moss-covered rocks and scattered flowers. In the background, there are tall, spire-like structures resembling castles or towers, partially obscured by mist or fog, giving the impression of a surreal, dreamlike landscape. The overall lighting is soft and diffused, with a cool, bluish-green tint dominating the environment, enhancing the otherworldly atmosphere.

>>110120020
If you have to ask, you're a ZITFag. GTFO

>>110120019
natty neck

What nodes are people using in Comfy to draw bboxes? Ideally, a wf would be good. I'm apparently retarded.

>>110119972
That's because Krea2 is fucking massive and only a handful people can train loras for it, what did you expect. Krea3 will be bigger so even less people able to train shit

File: 370629029572017.png (1.01 MB, 1152x896)

1.01 MB PNG

So are fucked up hands the price you have to pay for a model that can actually be fairly creative and do a bunch of different styles? This is at 50 steps CFG 3.0 btw.

Can a 5060ti 16gb generate videos okay?

Anyone here use this GPU? How is it?

>>110120035
just use dmd lora at lower weight as a stabilizer, do 15 steps, cfg3, its not rocket science.

>>110120035
Try the turbo lora?
>>110120055
You will struggle

Find out yourself, Kreafag

File: 26979782.png (1.56 MB, 1024x1232)

1.56 MB PNG

>>110120027
average moss bodyhorror unfortunately

>>110120031
you need to activate node 2.0 in the options. There is a bbox workflow for moss on civitai

File: 277873043317416.png (1.16 MB, 1152x896)

1.16 MB PNG

>>110120055
It's possible, yes. If you do 0.6 MP and 8 step turbo lora in Minimax H3 you can gen a 10 seconds videos in <10 minutes. Something like that.

>>110120024
yap yap yap. okay in other words whats the lora/character/style called? Seems so familiar.

>>110119931
that's because we have a couple of literal fags who don't like anime here

>>110120066
is minimax that much slower than good old wan turbo? I thought its super fast. Or was that LTX

>>110120059
>>110120066
Would you rather pick a 5070 12g or 5060ti 16g?

Can MOSS remember faces now? I have various celeb datasets waiting, Gillian Anderson needs to suck grey alien dick

File: 334817529763398.png (914 KB, 896x1152)

914 KB PNG

>>110120073
It's from a movie "Maquia: When the Promised Flower Blooms"

https://civitai.red/models/2944589/maquia-style?modelVersionId=3417844

File: 2026-10-19-16-46-33_00001_.png (2.73 MB, 1664x1248)

2.73 MB PNG

>>110116920
I love this model's face variance.

>>110120087
it has catastrophic forgetting

>>110120079
LTX is the fast one, H3 is actually faster than WAN though.
>>110120083
If you can stretch for the 5070 TI 16GB that would be worth, otherwise just the 5070 probably not worth it.

>>110120095
>this meme again
Where do you fags come from

File: nyc-skyline-1995.png (2.67 MB, 1920x1072)

2.67 MB PNG

>>110120087
malcom's already got a lora for moss so i'd say yes

>>110120102
turd russel please

>>110120093
ohh right right right, I watched it recently actually. Nice lora.

File: 243083097390341.png (998 KB, 1152x896)

998 KB PNG

>>110120058
That does seem to work pretty well, thanks.

>>110120102
It's the same retard repeating annoying phrases, it's clear this sperg is saying nonsense to push for bump limit

File: 51882.jpg (40 KB, 944x1248)

40 KB JPG

I catastrophically remember that meme
god I love this general

>>110120055
Yes, if you have at least 64gb of system ram, then it's no problem.

>>110120101
hmm well with my 2-stage turbo workflow (2 steps high noise, 3 steps low noise) on a 4070ti its a bit over a minute for 512x512 6s video.

>>110120102
People simply don't know how to train modern models. They use the exact same settings they've been using for the last 2 years with SDXL. They train a lora at batch size 1 with 2e-4 LR (or even worse, use something like Prodigy) and then act surprised when it overfits, fries the model, forgets base knowledge, etc.

>>110120143
you can just use ai toolkit that uses the settings the model makers suggested. This is solved

File: radio.png (569 KB, 512x512)

569 KB PNG

>>110119606

>>110119606
>bstaber insanity all over again
kek

>>110120094
they all have the same face in that image anon

>>110120142
A 6s video on H3 would be pretty fast too, it gets much slower when you increase the duration though.

>>110119606
I doubt it's been trained on that. In its common use it's a near-meaningless abbreviation anyway, since it can mean anything from loli anime illustrations to hardcore child porn, depending on who you ask. The model doesn't even seem to be accurately portraying ages as well as Krea2 (which is only slightly better than the average for that, at least with anime illustrations).

>>110120143
>(or even worse, use something like Prodigy)
Prodigy is the worst snakeoil ever peddled in AI

It adjusts the LR automatically, except now you need to twist and turn every other training knob else it will pick the wrong LR and overshoot or undershoot, and it's all dataset sensitive, so be prepared to mess with d-coefficient, decoupling, bias correction etc settings for every data set

No wonder AdamW is still THE most used optimizer

https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd/blob/main/lora_dmd2.safetensors
8-step v2

File: 200802436841580.png (960 KB, 1344x768)

960 KB PNG

>>110120130
Did you like the movie? I watched it when it first came out and thought it was pretty good, but it never caught on.

>>110120191
moss just does these uncanny children that gen x pedos would love

>>110120197
I pretty muched liked the ending. But I'm not really a movie guy desu.

File: 173588347678776.png (1.02 MB, 1152x896)

1.02 MB PNG

>>110120209
Fair enough.

>>110120146
Ostris sucks and the defaults are terrible even for XL

>>110120195
The power of autism, gotta respect that

The first version was really good, here's hoping this is even better

>>110120055
Got a 5070 Ti, the 16GB 5060 Ti has like half the compute, half the the PCIe bandwitch (only has 8 pcie lanes), half the memory bandwith, but the same ram size. So it be like a 5070 Ti but double the gen time.
Just tested H3:
1MP 5-second video with 20 steps and kitchen attention takes like 340 seconds.
The same 5-second video with 0.5MP takes 127 seconds and with the 8 step lora it is 61 seconds.
You'd had double of each duration with the 5060 Ti 16gb, if you are okay with that, go for it.
Things in mind with H3:
-you should have 64gb of Ram
-lower MP and the speed loras will negatively affect the quality, it dips noticeably below .6 MP, the combination of both only makes it worse
-the MP and video duration affect the render time not linerally, but worse. for example a 0.5MP video with 10 seconds duration took 144 seconds to render.

>>110120083
The second you run out of Vram you have to offload to your ram and then everything goes to a unbearable crawl. Since the 5060ti is only about 25% slower than the 5070 when it comes to compute, i would go for the 5060 Ti. Vram is king when it comes to AI.

>>110120250
me when someone knocks on my bedroom door

>>110120260
>>110120250
>wait mom just a second-

File: mommy walking.webm (1.93 MB, 768x1344)

1.93 MB WEBM

there seems to be some native ass restraints in minimax, even when giving it references of monster bubble butts it limits it. need to figure out a way to guide it.

>>110120297
There's like a million loras for this

>>110120195
Moss Base is ass with or without this lora, it can't into normal high-res-fix style upscaling. SFT does it fine with or without the SFT version of that lora though.

>>110120247
What can you run with 32gb ddr5?

>>110120332
sd1.4

File: MossHR_27221oldturbo_.jpg (2.58 MB, 2656x3552)

2.58 MB JPG

>>110120321
You sure about that bud?

>>110120343
I mean videos?

>>110120360
what's dynamic thresholding?