Blessed thread of frenship

why didn't you add my gen to the college

>>110286809
What gen?

>>110286791
TANKS 4 BAKE ANON

>train lora on my highschool crush
>generate images of her behind the counter at subway making my sammich
kino

>>110286809
>>110286380

>>110286835
benchod

What happens now that the chinese have solved image gen?

>anthropic is now pressuring the gov to install firmware locks on GPUs to prevent anon from running dangerous Chinese models.
Your days are numbered, anon.

So, what now? am I supposed to just bake my lora for another 70 epochs?

>>110286849
Oh they did? What's the solution?

File: file.png (2.04 MB, 1402x1122)

2.04 MB PNG

>>110286832
oh yeah anon? do tell.

>>110286877
Basically no synthetic data which lets you beat everyone else and their blotmaxxed slop SFTs

>>110286874
This cant be not real ....

>>110286875
Refine the captions, higher network rank if applicable, see if you missed any source images to add to the dataset...

>>110286886
You mean like how Stable Diffusion 1.0 did it?

>Blessed bake
based

>>110286919
In a cooler and more chinese way with some other stuff added on

File: kromoss_bf16_OPD-priming_(...).jpg (3.31 MB, 3360x1920)

3.31 MB JPG

hmm..
pretty gud

>>110286741
Whats that? I couldnt find anything on miniconstruct minimax h3.

>>110286832
Can't wait to see you in the news :)

>>110287003
>Can't wait to see you in the news :)
He's just baiting.

>>110287033
Masturbating that is ;)

/long dick general/ won

longest dick general

>>110287003
>>110287033
It's the distributing part that's illegal, not the creation for personal use. Don't bother replying to me I will not read it.

File: 185488.webm (3.62 MB, 448x256)

3.62 MB WEBM

On the seventh day, they got up at daybreak and marched around the city seven times in the same manner, except that on that day they circled the city seven times. The seventh time around, when the priests sounded the trumpet blast, Joshua commanded the army, “Shout! For the Lord has given you the city!". When the trumpets sounded, the army shouted, and at the sound of the trumpet, when the men gave a loud shout, the wall collapsed; so everyone charged straight in, and they took the city.

File: a2.jpg (1.22 MB, 1334x1424)

1.22 MB JPG

>>110286706
>>110286732
A without GRIT Realism chain flow.
B is with GRIT Realism enabled.

File: 2026-11-09-201920_620x267(...).png (152 KB, 620x267)

152 KB PNG

>moss can't spell swamp

>>110287076
Kino

File: kromoss_bf16_OPD-priming_(...).jpg (3.32 MB, 3360x1920)

3.32 MB JPG

>>110287077
ill run it against this

File: kromoss_bf16_OPD-priming_(...).jpg (3.3 MB, 3360x1920)

3.3 MB JPG

interesting.

Been wondering where the kinos were at!

>>110287143
>>110287148
hard to say which is better for this comparison. i think it definitely boosts detail by loses some of the contrast. I should note I did not use any turbo loras either.

File: kromoss_bf16_OPD-priming_(...).png (2.88 MB, 3360x1920)

2.88 MB PNG

>>110287179
im using turbo and hires lora

>>110287077
B is WAY better. you can see the dust particles reflecting off the head beams in the window pane. her vellus hair and hair strands are more pronounced. the texture is more detailed

File: kromoss_bf16_00542_.jpg (3.45 MB, 4032x2304)

3.45 MB JPG

without grit

File: kromoss_bf16_OPD-priming_(...).jpg (3.28 MB, 3360x1920)

3.28 MB JPG

with
and it is a different model so not a fair comparison now that i think about it

>>110287224
yeah, only valid if using same model/seed/sampler/scheduler/etc

>>110286809
Because it sucked ass majorly
Get beasted on

>>110287250
benchod

where do you post your blacked gens?

>>110287256
/hdg/

File: kromoss_bf16_OPD-priming_(...).jpg (3.5 MB, 3360x1920)

3.5 MB JPG

alright
people that dickride krea are retarded

>>110287256
>>>/b/956231395

File: 537863.webm (3.7 MB, 448x256)

3.7 MB WEBM

When Moses approached the camp and saw the calf and the dancing, his anger burned and he threw the tablets out of his hands, breaking them to pieces at the foot of the mountain.

>>110287260
The same people who are still dickriding krea now will be dickriding moss long after it has been deprecated. They will always lag behind.

>>110287260
i like how the rings aren't just smooth lines. they have a wavy pattern.cool touch if you didnt specifically prompt that

>>110287260
moss is garbage, cant even generate an appealing 1girl or anything that involves a living organism without body horror.

>>110287292
Major skill issue

yawn.

>>110287270
those threads are weird. you have some weird hugbox shit going on similar to /sdg/. all that friendly greeting and sucking up to namefags annoys the fuck out of me. i cant browse it

>>110287297
You also failed to mention that they're all pedos.

>>110287273
It's crazy how you're literally the only one who can gen kino with LTX that BTFOs 90% of anons using H3. Based.

>>110287298
I mean yeah it's /b/, that's kind of expected. are you some kind of normie? go back.

>>110287297
>you
i don't go there lol
you wanna post nigga dick post it there or in /h/ or /aco/

>>110287309
/trash/ also has an ai general though i doubt they want you there
idk though i dont go in porn threads

File: 357874.webm (3.66 MB, 448x256)

3.66 MB WEBM

>>110287302
i would be using h3 if it had good t2v quality for realistic videos, but it always made things look like a video game so i gave up on it

>>110287302
>90% of anons using H3
90% anons aren't posting their h3 gens.(because they're porn)

>>110287323
>it always made things look like a video game
It's all the synthetic slop they used. smart but souless model.

>>110287260
Those Kromoss gens are really good, some of them have congruence details that trigger my autism, but in general look pretty nice.

Krea2 is still clearly superior for some cases. Like, can you give a Moss based model a 500 words long prompt that involves two characters in a complex interaction, and get good gens without borked anatomy, or gibberish backgrounds. Example: >>>/hr/5304423 The pose and description of each 1girl is pretty specific too.

>>110287329
you can kind of make it work if you use an image model to generate realistic video frames, but i like pumping out videos quickly without needing to switch models around so i will stick to ltx until something new comes around with good t2v kino abilities

File: 1792714918947973.png (18 KB, 900x806)

18 KB PNG

where kino?

>>110287367
CTRL + F "kromoss"

>>110287367
Where's my turkey?

>>110287323
Interesting. While minimax is a beast its very "stiff". You really need to spend a lot of time with it and schizo prompt to get what you want. Its what I liked about wan and ltx, both have good wobbly loose physics. The same dataset I used to train loras for wan and ltx to then be used on minimax did not work...at all. I ended up having to use generations from wan and ltx as the updated dataset to get what I wanted out of minimax, kek.

So for people who don't know, MOSS was trained on raw barely-filtered web data, relying on filters already applied by the datasets—their additional filtering was only to remove duplicates and AI gens. The model was trained is on a variety of resolutions but mostly at sizes around and below 512x512, progressing up to 2MP+ at the last (smaller) stages of training.

It trained on a mix of autocaptions and original alt-text. Gemma 4 was given the images with their original captions and asked to produce descriptions at different lengths: a one-sentence description, a long full description, and a paragraph condensed from the long description. Defects (e.g. watermarks) were paid attention to and captioned, rather than filtered out of the dataset. JSON layout schema were also generated and used, though I know less about this and it really would be a separate post. Prompting this way is of course optional.

Roughly 20% of training passes used original captions (where available), 80% generated. This is why MOSS lets you prompt a little more like SD. Clit Eastwood, etc.

Caption token caps were 300ish in first stage, expanding up to >4000 by the final pass. It can take a lot of description.

Long captions preserved line breaks in text transcription. So use manual line breaks when you want one displayed in any text in the image, e.g.
>his shirt says, "RUN
>DMC".

Not sure if there's anything else worth saying about it. It kinda just werks.

>>110287396
>Long captions preserved line breaks in text transcription.
this is a good tip that i use all the time. if im being lazy ill sometimes prompt "in three lines, text that reads ..." which works as well.

File: MOSS_02487_.png (1.32 MB, 1184x880)

1.32 MB PNG

>>110287367

File: 1791476864627434.jpg (193 KB, 1170x1822)

193 KB JPG

Is there an LLM that can reliably scan and translate online manga while fitting inside a basic bitch 5060ti? I tried one of those mobile store "AI translators" and their zero context mtl gave me vietnam PTSD...

File: moss-611803443565099_00001_.png (703 KB, 528x800)

703 KB PNG

>>110287396
I still can't believe we have a real SD1.5 successor, guys. It only took four years.

>>110287421
>>>/g/lmg or >>>/g/vcg

File: kromoss_bf16_OPD-priming_(...).jpg (3.59 MB, 3360x1920)

3.59 MB JPG

>>110287341
i mean i can do that with sdxl and controlnet with 1/5 the prompt length and maintain the correct amount of fingies
but i have adhd not autism so ig krea is more for the autistic like sasori kek

>>110287392
one thing about h3 is that is has time stamping built into it, so you might be able to make a prompt generator that mimics the wobbly movement by making a giant series of little timestamps that jitter the perspective around. i had some success with it but it goes back to the same problem of having to switch to different models if you want your generations to look different since having a strict prompt for movements makes there be no more variation

File: moss-1073328522927507_00001_.png (876 KB, 528x800)

876 KB PNG

I bet it would be good fun to play one of those old RPGs with a custom party now. So easy to make portraits

>>110287341
>Krea2 is still clearly superior for some cases. Like, can you give a Moss based model a 500 words long prompt that involves two characters in a complex interaction, and get good gens without borked anatomy, or gibberish backgrounds. Example: >>>/hr/5304423 The pose and description of each 1girl is pretty specific too.
yes actually AND you can bbox it if you want. it also wont look like ai slop like that gen you linked

>>110287436
>this isn't /lmg/
Thanks i'm retarded

>>110287438
My experience with controlnet with more than one character and a complex pose is pure borked anatomy, borked anatomy everywhere. But you are right, Krea2 is the election of the auti... perfectionist AI enthusiast

>>110287446
Can I see an example of that? Asking genially, I plan to experiment more with Moss in the future.

>>110287465
right looks better. whats different?

>>110287529
nvm, i didnt even see that fucking text

>>110287536
Is that a lora? The style is very kino.

>>110287551
ty, my first lora.

>>110287396
>The model was trained is on a variety of resolutions but mostly at sizes around and below 512x512, progressing up to 2MP+ at the last (smaller) stages of training.
retards rip on lode and every other community model baker who do this for some reason

>>110287396
>Clit Eastwood

File: 1763601429072418.png (2.03 MB, 1024x1536)

2.03 MB PNG

can moss generate images like this

>>110287648
God damned, what a based man....

>>110287583
Nice. I'd love to try it but no pressure. GJ anon.

>>110287396
>at sizes around and below 512x512, progressing up to 2MP+
I was able to generate proper gens up to 7MP without much garbage.
Above 8MP it's bad though.
Overall very happy with the model, and so far very disappointed proper finetunes aren't worked on.

>>110287668
too much fragmentation. most image models do everything that people want now, so the finetune autists stick with whatever they have settled with

File: 1779329849733845.jpg (146 KB, 816x1232)

146 KB JPG

perhaps

File: 1780018807760388.jpg (103 KB, 816x1232)

103 KB JPG

>anon, there's something on your face

>runs over glownigger
>rear-view mirror shows a green glare fading in the distance
kino

File: 1788096398026207.jpg (71 KB, 696x700)

71 KB JPG

Want to try out the kromoss update. What's the default workflow? I can't seem to get the base Moss one to work.

>>110287661
ty, I'll upload it soon.

>>110287428
>>110287434
kinoooooooo

>>110287767
my house

>>110287076
goddamn the bible is gayer than 4chan captchas

File: 1781799257363858.jpg (1.32 MB, 1408x2128)

1.32 MB JPG

>>110287537
kek

>>110287862
it only has a few gay sex stories

>>110287920
oh fuck yes

>>110287812
That's a railway, anon

what is good sex lora for h3 2d?
I use mystic xxx, but the animation is so stiff

>>110287920
How do you even caption stuff this abstract?

>>110288003
very carefully

>>110288003
gemma4:e4b because im impatient. its serviceable but im probably going to give in and switch to qwen3.X soon. here's a random image and paired caption from the set https://files.catbox.moe/erix87.jpg https://files.catbox.moe/ttf82u.txt - thats definitely one of the better examples

the next step towards better diffusion models will be better captioning models

File: debo_pf_km_00037_.png (2.69 MB, 1792x1076)

2.69 MB PNG

>>110288004
could be like a junkie xl album or something

nah the next step is autoregressive models

File: ComfyUI_temp_rklyt_00114_.png (1.44 MB, 832x1248)

1.44 MB PNG

>>110288033
It's really impressive how well it's describing the scene. thanks for sharing.

>>110288072
Regress to where? SD 1.5?

did sulphur 3 leak yet and if so where
if not no worries thanks for your time
i want to see how dogshit it is for anime content its very important

>>110288002
One of the vidgen threads probably knows

>>110287256
/v/

>>110287648
I tried for a while and I think the answer is no.

>>110288648
anon did it here
>>110287681
>>110287687

>>110287328
is faceswapping native or do you need a lora?

>>110288736
native

>>110288694
With all due respect I saw his images and, while they're good tries, I don't think they achieved it.

>>110283223
Forgot to reply to this at the time. Not a fan of the gen overall but I like her eyes, how they droop.

File: tmp.9D18NQEDYc.jpg (3.53 MB, 1440x2048)

3.53 MB JPG

bbox's are fun

>>110288911
>nekojak

>>110288911
Where's the Fennec loli?

>>110288938
tried to come up with some way to put known names together that look kinda japanese. couple funny ones if you can read/translate them

>>110288945
(none of the text is invented by the ai/llm, though since i can't into japanese myself i can't guarantee all the characters were rendered correctly)

>>110288911
where THE FUCK is moss anime?!

>>110288282
i don't even know what the fuck is going with the dev team of sulphur 3. everyone is being hush hush about the results of the model in it current state. perhaps some of those donors who have early access are feeling the burnt and want finetune it further rather than prune down the model.

File: sizeslop.jpg (1.03 MB, 3414x5120)

1.03 MB JPG

im trying to build a sizeart pipeline with a dynamic multigen process
any ideas, suggestions or help
all the existing loras and setups suck
the idea is generating the main figure, background, and small figures separately, putting them together, upscaling to 4k after, then adding effects and details to emphasize the size differences and scale

>>110289113
garbage

donate your compute to sam altman

File: ed5.png (162 KB, 680x1112)

162 KB PNG

>garbage

>donate your compute to sam altman

>>110289178
now animate that basedjak doing something funny

Is kroma still the best general all purpose model on the market for non anime stuff? Tried Moss and it feels like Chroma 2.0 in the sense it can be creative but wacky as fuck in terms of composition and terribly slow.

>terribly slow

>>110289322
moss, at least base, works much better in my experience if you use more detailed prompts. probably why the stock workflow preprocesses prompts through an llm to beef them up a bit. i've just been using gemma separately to fill out my prompts if what i write isn't getting me where i want it

File: moss_image_sft_int8_convr(...).jpg (685 KB, 1024x1504)

685 KB JPG

a giant

>>110289328
if you're doing >2MP pictures at the recommended 50 steps it can take some time unless you have really strong hardware. i've found you can go as low as 20 steps even without the turbo lora depending on what you're making, and then you have the turbo lora which can bring it down further

usecase for other image models now that moss exists?

File: moss_image_sft_int8_convr(...).jpg (1.15 MB, 2512x1504)

1.15 MB JPG

https://huggingface.co/lodestones/Kromoss/blob/main/diffusion_models/kromoss_int8_convrot_OPD-priming.safetensors

>>110289445
krea2 for realism everything and anima for illustrated smut

retard here, is there some hard bottleneck that makes modern models so ungodly slow? I find myself going back to XL models again and again because they just don't take 3 minutes to gen pictures.

>>110289463
3 minutes isn't slow

>>110289460
>krea2 for realism everything
moss does that now and isn't cucked

>>110288911
super good, is moss able to write this well?

File: a.png (44 KB, 204x102)

44 KB PNG

>>110289500
that is moss.
the artist credit i put in there similar to real LO covers says "mosu imeji", or moss image

File: tmp.d03kig2NMO.jpg (3.05 MB, 1440x2048)

3.05 MB JPG

>>110288911
fixed a few things, objects are not the intended size and position. even managed to get the 12 dots correct which surprised me.
hard to believe this is straight ai output, no editing or separately-made layers or anything

this is the template i used for layout, just for positioning and what text styles to use, it was not an input itself to the llm or moss
https://litter.catbox.moe/02563m.jpg

>>110289556
are now*

>>110289417
30 steps euler cfg pp. cfg 2.5-3

File: 1755829449544.jpg (886 KB, 1152x1376)

886 KB JPG

>>110289556
what's the actual prompt?

>>110289585
3.7k worth of json

>>110289601
post it nigga

>>110289585
>>110289601
the global caption is just;
>photoscan of an old LO magazine cover, slight damage around the edges and scanner artifacts. spine on the right.
then a whole-image background box with a gemma-enhanced/written prompt detailing basically "three 11 year old girls near a lagoon" to put it simply, though i specified that it's a magazine cover with erotic/candid elements. the other boxes are pretty self-explanatory

>>110289556
B L U R R Y
L
U
R
R
Y

MOSS IS BLURRY SLOP. SHOULD I PUT SOME GLASSES ON? IS IT HUMID? WHY IS EVERYTHING FULL OF FOG? IS THIS SILENT HILL OR SOMETHING?

MOSS IS NOT AN ANIME MODEL
MOSS IS NOT AN ANIME MODEL
MOSS IS NOT AN ANIME MODEL

5B IS NOT LONG ENOUGH FOR HER DEEP NAI PUSSY

File: sizeslop.jpg (988 KB, 3334x2711)

988 KB JPG

>>110289063
some progress

>>110289633
put your 'code back on

>>110289633
been and done ultra-sharp moss pics. this is supposed to look like an old scan.

>>110289649
Send me your whole prompt including bbox, i wanna see how convoluted it is compared to nai

File: 738467.webm (3.7 MB, 448x256)

3.7 MB WEBM

>The CIA niggers glow in the dark, you can see 'em if you're driving. You just run them over, that's what you do.

>>110289682
sure, i would have liked if someone gave me some examples but nobody responded to me.
note that i didn't write this all by hand, much of the finer detailed were filled in for me, like for example the two girls besides gemma were entirely made up by the llm since i didn't really care what they looked like specifically. the llm extrapolated my existing gemma prompt into a swimwear version on its' own.
i used this editor;
https://moss-image-community.github.io/moss-image-layout-editor/?lang=en
no comfyui workflow, i didn't use comfyui.
https://files.catbox.moe/wqrlwb.json
negative prompt:
>low quality, jpeg artifacts, amateur, simple, duplicate objects

>>110289737
Yeah I see, you probably copy that into Forge o algo. I mean its not bad ig, still a lot of effort compared to how its done on nai, but beggars cant be choosers ig

File: MiniMax_H3_00135.mp4 (846 KB, 576x880)

846 KB MP4

>>110289737

>>110289782
i'm just using stable-diffusion-cpp from the command line. haven't heard of forge

>>110289783
>loli
>they're all lanklet high schoolers

>>110289782
>>110289788
also for completeness, here's what i asked from the llm (gemma-4 31b);
>write an image prompt intended for AI image generation for a vertical erotic manga magazine cover that includes 3 cute slim 11 year old anime children in childrens' swimming outfits, who are under shade near a lagoon near midday. describe each girl individually, they're in different poses showing off different parts of their body in an erotic yet casual way. candid photo. the girls aren't intentionally showing off. single line of text. be descriptive with clothing, hairstyle, colours, etc. one of them is Gemma, /unlike/ the others she is looking at the viewer with her mesugaki expression. nice skin texture, non-vibrant pastel colours. barely pubescent with thin thighs, flat ass and developing breasts. they have soft round faces in a modern anime/manga style
this is the extent of what i wrote myself for the background
couldn't say how it compares to novelai as i haven't tried that.

>>110289788
In nai you just create slots and place them manually within a canvas, or maybe just fill up slots with prompts and let it decide by itself, its all very intuitive.

>>110289801
sounds like basically the same thing, isn't it?

>>110289799
Yeah idk in nai you dont need to expand prompts with llms to get good results, you already have (optional) quality tags embedded into prompt output.

>>110289809
Well no, since it doesnt bound you within a box, its a dot you place in the canvas, not a limiting box, so it has more possibilities

>>110289814
you sure it doesn't just do that internally? in moss if you follow the official instructions you will end up with a comfyui workflow that passes what you enter through an llm by default. i'd imagine it'd be similar for cloud services as well since they surely get a lot of really under-developed prompts

>>110289820
Couldnt be unless they have NASA supercomputers, a nai gen takes <5 seconds

>>110289818
sounds more limiting to me. you'd probably struggle to replicate a specific document layout like i did using only suggestive points

>>110289827
if they can generate an image in <5 seconds then they can surely process a prompt in a fraction of that. you don't need a particularly amazing llm just to condition a prompt

>>110289820
you get perfectly reproducible seeds on nai which I think doesnt work if you use llm enhancer pass. pretty sure they're just using a beefy te

>>110289842
haven't looked into it but there's no reason you couldn't reproduce llm output by storing/providing a seed as well. it's random in the same way image generators are, that is, they're pseudo-random and can be made reproducible if you know how it generates random numbers.

>>110289809
ani said that novelai just copypaste shit from local and never develop new tech themselves. they probably stole bbox code and reused it for their dot shit and made a different ui to make the steal less obvious. they are a greedy corpo, nothing more

>>110289842
>>110289869
but-- i don't know what nai does, maybe they do just have a very different text encoder, maybe that can do the same job, i really have no idea.

>>110289837
Dunno, it just sounds like cope to me, since in that case we'd see local be on a similar level. I'd say no just based off the fact that a low res gen in nai is gonna take less than a second, I just dont see a mediocre-fast LLM improving it in any way.

Its just a good model anon.

i dont understand why there are still nai shills in 2026. this isnt 2022 anymore

NAI v5 has a 32 channel VAE and comissioned artists to create art just for them, no, you cannot get remotely near it with local models.

>>110289903
gotta justify how much they've spent on it somehow

File: file.png (55 KB, 750x294)

55 KB PNG

its the same discord clique whos been shitting up the /h/ threads for the past couple of years, ignore and hide their posts

File: out.webm (2.66 MB, 1440x2048)

2.66 MB WEBM

>>110289556
previews to show it was unedited. i only saved previews every 6 steps for this since i didn't intend to make a video and generating vae previews costs time

>>110289535
So you did prompt for every text? (I guess bboxing everything)

>>110289872
this is actually catjak pretending to be ani

WARNING

WARNING

File: Screenshot 2026-11-10 526x752.png (388 KB, 526x752)

388 KB PNG

>>110289887
i'm not doubting it's a good model, i'm sure it is. i've never said anything bad about it. it's pretty nice you can make pics in 5 seconds, my gaymen AMD video card can do that with any semblance of quality for sure.

>>110289997
pretty much, yea. the bbox's can been viewed from what i posted at >>110289737
pic related is what i was working with.

>>110289926
>has a normal well adjusted conversation
>doesn't like the outcome of said conversation
>"yeah its just this group of boogeymen shitting up the thread"

>>110290011
>can do that
can't*
don't you just love typos that flip the intended meaning?

>>110290011
Nice pic, wouldn't mind more bbox analysis pics like that one

>>110290022
i'm certainly going to be playing with it a lot more. this is my first "successful" use of it. that is something good enough to post here in full.

>>110290011
Sorry trying to understand anon, so you started with an existing image as input (the comic LO one), or you just used it as reference and actually started with pure noise?

>>110290046
yes, i used the original LO picture as nothing more than something to trace the box dimensions for and to describe the text styles to use.
it was in no way at all used as input data for any other process. it was used only visually by myself.
the image was generated objectively from a pure noise seed like any other.

>>110290061
Thanks for answering anon, I never played with bbox, but that's really cool how well it made the output.

>>110290100
thank you, i try to only post pictures i'm properly happy with, or at least would be helpful to someone.

>both adt and ldg are dominated by cunnychads
catjak lost

i have to keep name dropping myself to make it seem like i am relevant o algo

>>110289872
>>110290003
>>110290179
dude no one cares about your avatartranny failure

I have some serious doubts about the new Kromoss, yes the colors are better but it's way less coherent it's starting to feel like Kroma now

>>110290593
wait for the new dmd2 lora

>>110290662
It appears he released a new checkpoint and replaced it, so far it appears more stable but it seems like every model from this family shits the bed with loras from base or the previous version.

>>110290687
yes the dmd2 lora was trained against an older version and not optimal for the new version uploaded today, theres gonna be a new dmd2 lora trained for it

File: 20261110_vertigo_finalF.jpg (2.19 MB, 4558x6835)

2.19 MB JPG

>>110289636
this took like 1.5 hours
anyone else a sizefag here

>>110290704
That can be circumvented with it at lower strenghts, the updated model is significantly better than whatever the hell he released 10 hours ago. The biggest gain is the improved colors. Going to train a proper lora for it because even though the model didn't play well with loras from the previous model, the loras trained off the preview version had significantly better color as well.

>>110289463
Yes, your GPU. I gent 2mp images in 30-50 seconds with Krea2 and S(uperior) FT Moss. I have a 5060ti 16Gb they are not cheap but not ultra-expensive either.

>>110290709
>pubes

>>110290795
What settings you using for moss? If I go for high steps/res its like 2 minutes or more per gen on same card. Not good.

>>110290829
pubes are hot as fuck you gaylord

>>110290837
ok homo

File: nytxmk80t3981.jpg (61 KB, 1080x722)

61 KB JPG

>>110290836
16 steps, CFG 3, Euler - Shrimple + Lodestones Dm2 Lora, int8 covrot models.

>>110290914
Why are u using int8 if bf16 or whatever fits comfortably in your card?

>>110290914
Stop giving people fake expectations. The actual non cucked model, at 30 steps or more, is gonna take 3+ minutes on that card.

is char/style reference out for moss yet?

>>110290965
I had problems with Moss (mostly because it isn't capable to generate cunny) (GTX 1060 6 GB), but I found very good checkpoint on Civitai and I'm happy now, because it can generate almost all.

8 steps, cfg 1, euler, beta without any loras, 1024*896, 1-2 minutes for pic

h3 image out:

https://huggingface.co/MiniMaxAI/MiniMax-H3-Image

>>110291009
>68.65 GB

made me check fuck you

>>110291009
69gb
what the fuck

>>110291015
>>110291021
https://huggingface.co/Comfy-Org/MiniMax-H3-Image/tree/main/diffusion_models

>>110291015
FP16
do we have to go over this every time

>>110290931
I experimented with Moss briefly before going deep into Krea2, I usually use bf16 if possible. I loaded and old image on Comfy and those were their settings. Maybe it was an early gen or a speed experiment.

>>110290965
I was using S(uperior) FT Moss original, not a merge or finetune etc. Unless you were referring to base Moss and you are one of those people that confuse blurriness and instability with "soul". In that case, your are wrong. Just like vinyl fans who confuse noise and distortion with "warmness" are wrong. The Dm2 lora at higher CFG and steps gives Moss better anatomy.

krea killer?
uncensored vagoo?

>DiT
yikes. waiting for the next model

File: file.png (2.07 MB, 880x1184)

2.07 MB PNG

thats a pretty good t34

that's it? it's all up, ready to download and merged?
doesn't feel right without a bogus countdown then hours of speculating chinese culture

>>110291051
fr, NAI isnt even DiT why even bother

for the fucking size it is, it better have danbooru artists.

File: file.png (43 KB, 806x297)

43 KB PNG

>>110291058
>that's it? it's all up, ready to download
yes
>merged?
i think the last needed commit should be merging right now
>doesn't feel right without a bogus countdown
comfy had one, im assuming it was about h3 image

>>110291066
They shouldn't be massive noobs and also pull from exhentai, since it will have stuff beyond what the artist would want to hand over for free.

Can someone let them know on discord about this fact?

why is it so bloated?

>>110291083
it's a h3 finetune

>>110291041

Still has weird looking genitals from what I've seen people talking about that used it

Examples have the typical glossy slop faces. not looking good.

NAI is good because of the 32 channel VAE, its the most expensive part, its like a higher internal resolution for the models thoughts (latent space)

maybe it's too early to tell but it looks very deepfried

Is the anime version of moss out already?

>>110291009
How safety-maxxed it is?

>NVFP4
Curious election, isn't int8 covrot superior? Well some will made that quant later, surely.

Also, is going to trigger Moss fags hard, in special Moss base fags.

>>110291009
in a scale between buttchin and ootb nsfw, where is that model

Does AniStudio support H3-Image yet?

>>110291206
same as the regular h3-video, youll get tits and nipples sometimes but no vagoo unless someone finetunes it (unlikely since its a step distilled model with no cfg + huge af with 20B parameters)

>>110291009
>It generates images of about 16 megapixels at every aspect ratio, from 4096x4096 (1:1) to 6272x2688 (21:9), directly and without an upscaler (see Sizes). The checkpoint is distilled to 6 sampling steps (6 NFEs) with no classifier-free guidance.

>>110291221
sad, damn it

>>110291212
Ani surely is working hard on it, he will not disappoint us. I trust Ani.

File: file.jpg (1.22 MB, 3000x911)

1.22 MB JPG

>>110291009
bros it knows vaginas!

>>110291257
kek

back to moss i guess

>>110290914
I tried doing this, or rather attempted to as moss is just as confusing on what model to use and settings, a 1mp image took ~40 seconds, while 2 mp is ~81 at the suggested step and settings: sft 8 int8 convrot, int8 qwen and small vae.
The odd part is, when doing bf16 base non convrot, bf16 qwen and small vae, 1mp takes only 23 seconds while 2mp is 53s, where I see my normal ram shoots up a little bit,

I also have 16gb vram too sadly, do you have a workflow? preset moss has unusual settings like generative text.

Most of you shitters can't even run the prunned model. Stop taking examples for face value, shut the fuck up and sit down while the real pros look at it.
Also they don't have any prompt etiti

well im waiting

etiti

>>110291029
Might unironically be a better anime model than moss

Now there's a sea of half trained models in the wild, great

Miss has such a wide moat because no other lab is going to train their model on hardcore porno

any workflows with that prompt enhancement thing for H3-Image?

>>110291350
This. Waiting on Ani's opinion.

>>110291372
>>110291350
catjak is bored samefagging

>>110291368
yeah, and it's insane how it helped the model look good overall even sfw stuff

>>110291257
69GB btw

where dem workfloes at nigga can do shit widout dem workfloes

>>110291380
You really have that nigger's dick deep in your throat.
>>110291400
Pruned model exists

bros, i have a 4070 super. it's a great card but 12gb VRAM is becoming too much of a bottleneck.
what's the state of ROCm these days? is it much slower than CUDA for image and video generation? i can currently only afford an amd card among the 16gb+ models.

>>110291419
Still going to be huge.