https://files.catbox.moe/fsytg4.mp4
Anyone figured out how to get gens sharp willing to share?

should I upgrade to sageattention 2? seems H3 Mem Eff Sage patch requires it but I'm not sure if it works on my 3070

>>109468105
I too have been trying to upgrade to sage 2.2 but failed two times, followed Claude's instructions
3090
on linux

>>109468105
yes 2.2, it will work

>>109468066
If THAT guy appears. Even if he spams the thread. Ignore it and let someone else make the thread. Just copypaste OP post

any h3 prompting tips?

>>109468130
Which guy? You have a mentally ill local?

>>109468146
He is the mentally ill local

The impact of all these optimisations is huge, but I think I see a trend of worsened physics the more you pile up.

>>109468088
I don't speak your language. You got a guide somewhere? What am I supposed to write in the system prompt?
DS4 is too big for my system. Heavy dram offload.

Something not talked about is the Minimax VAEs are goat and good for training new models.

the reference model is so good.

<Subject 1> is the dog in <Picture 1> and <Subject 2> is the man in <Picture 2>.

the setting is an interrogation chamber at a police station. <Subject 2> is sitting on a steel chair and is handcuffed. <Subject 1> slowly paces around <Subject 2> and fires electric bolts at them with their mouth, and <Subject 2> gets electrocuted and says "Okay stop! I wont shock dogs any more!".

https://files.catbox.moe/6eunni.mp4

>>109468066
if i want to add another ref do i just add another load box? Im using the default workflow for H3 Image to Video?

Thanks to the anon recommending me flux2 klein over kontext a couple of days ago, it's really much better and runs just about as fast with the 9b convrot model, the 4b one sucks tho

i didn't realize wan2gp updated to add some optimizations. gonna check them out later cause i got to go

File: seething chink.jpg (66 KB, 750x962)

66 KB JPG

>NOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO YOU CAN NO JUST MAKE GEN WITH MY OPEN SAUCE MODEL NOOOOOOOOOO

Is there a good node to resize the image & video references to the right size for the dimensions of the output video? It is a real pain to manually do that in photoshop for every reference whenever I want to change the resolution of the output video.

So is Krea-2-Turbo really just a better replacement for Z-Image-Turbo or is that just a meme?

>>109468198
Good someone has to stop you from self-NARCing

Has anyone else noticed that prompt adherence in the H3 ref model is noticeably worse? For me, it seems to struggle with prompts that worked really well in I2V model.

>>109468217
Krea2 simply does not make the errors that Zimage does and let's say you use the Loras, they will work far better than the pretty bad ones for Zimage. you can also stack Loras. That's another thing regarding minimax over LTX, will the Loras be more useable because with LTX you had to lower the strength more an more when stacking them

>>109468105
>>109467728
>>109467730
>>109467642
I built it my self about a month back after using clanker to figure out why it wouldn't build on linux. So search for the actual wheel file name and found KJ has it already built here.
https://huggingface.co/Kijai/PrecompiledWheels/blob/main/sageattention-2.2.0-cp312-cp312-linux_x86_64.whl

>>109468237
>https://files.catbox.moe/6eunni.mp4

It has its own prompting style if you want to look into that.

https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md

I think society was more genuine before AI, it was just people recording real videos on their cellphones.. just living in the moment...

u can use cropped reference video to guide the scene without mentioning <video 1>, acts like a mini-lora. it seem to influence the movement but sometimes the subject merge.

https://files.catbox.moe/kp79u9.mp4
https://files.catbox.moe/1fne8b.mp4
https://files.catbox.moe/mvdbmc.mp4
https://files.catbox.moe/jxogtg.mp4

>>109468269
Yeah I'm trying to follow that, I'm just talking about the action descriptions that I2V understood well REF doesn't seem to.

>>109468261
I just noticed even the convrot version is twice as large, so is it really just an alternative if you aren't a 10GB vramlet like me?

>>109468281
garbage fetish

>>109468286

I agree with you, it seems like the ref model they've made is really good at some things and terrible at others.

>>109468281

I've been trying to use NSFW clips to guide motion and replace people in scenes but it's been terrible, which is unfortunate.

>>109468281
That's SCAIL though, isn't it?

>>109468302
h3 r2v

>>109468313
Alright, replacing people with reference images will never be good enough for me, so I’m not even going to try.

>>109468281
Man I wish some of you fuckers here had some normal fetishes, one fuckers is obsessed with Floyd BBC and another with black ass...

>>109468349
fetishes are definitionally not normal chud

File: 0805222144785-inI5ZSVqhOI.png (193 KB, 1900x1216)

193 KB PNG

>>109468200

hey krea boys
does musubi tuner already have lokr support for it?

>>109468120
https://desuarchive.org/g/thread/109238692/#q109239881

Hope you can figure out from my ramblings.

do you guys rigidly boilerplate your i2v (whether via llm or manually) with all the
>For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
>integrated_multimodal_description: [Shot 1] (general desc). The scene opens exactly on <Picture 1>, (shot desc)
etc or just generally follow the format of declaring [Shot 2] and <d>[English] Fuck you</d> but otherwise freeform it? Or go even more freeform than that? Unfortunately at 5 mins a pop even for fairly quick gens it's a bit grueling to A/B test H3
>>109468200
have claude vibe code you a node with inputs of image,width,height,scale_method (non-input for the last one by default) and output resized image, which resizes the image until one dimension matches the target and the other spills over, then crops the larger dimension down to match the other target, and outputs the image. ensure you roughly match its aspect ratio with your picked input dims.
>>109468367
aren't image dimensions meant to be multiples of 32?

>>109468349
Where do you think you are ?

File: 0805222940579-ixMhuN3XNbb.png (26 KB, 388x512)

26 KB PNG

>>109468367
then just add this in between

doritos pope, street battles

<Subject 1> is the anime girl in <Picture 1> and <Subject 2> is the man in <Picture 2>.

the setting is New York City during the day. <Subject 1> is having a fist fight with <Subject 2>. <Subject 1> punches <Subject 2> very hard several times, and <Subject 2> falls to the ground and says "Okay! I'm sorry about Highguard!"

https://files.catbox.moe/b84p91.mp4

>>109468438
>>109468411

>>109468349
If we had normal fetishes, we wouldn't need AI.

File: Screenshot 2026-08-05 173159.png (35 KB, 297x517)

35 KB PNG

I like this node for cropping images to right aspect ratio.

euler or res_multistep for minmax?

>>109468438
>>109468455
yep, this will work and end up with an output video that's closely matched to whatever aspect ratio the input image had, which is probably easier and nicer to work with than my method unless you specifically want to pick output aspect ratio and fit the image to it. and i might swap to using yours since thinking about it i dont care that much about that, i assume H3 doesn't perform particularly differently at "magic" specific resolutions like SDXL did

better result as I set the shot on geoff first

<Subject 1> is the anime girl in <Picture 1> and <Subject 2> is the man in <Picture 2>.

the setting is New York City during the day. the view is a closeup of <Subject 2>. <Subject 1> is having a fist fight with <Subject 2>. <Subject 1> punches <Subject 2> very hard several times, and <Subject 2> falls to the ground and says "Okay! I'm sorry about Highguard!"

https://files.catbox.moe/mspqgy.mp4

50fps

probably too high but that is some smooth movement

https://files.catbox.moe/yztb4l.mp4

>>109468504
I don't know looks stiff.
Reminds me of janky cartoons in the 2000s.

File: 92104908.mp4 (3.19 MB, 832x1152)

3.19 MB MP4

>>>/wsg/6208574

shieeeeeeeeet Miku

https://files.catbox.moe/7vfkab.mp4

wsg or fuck off
I'm not going to waste my energy by opening your shitbox links with yet another fentboy or anime slop

File: 1772074896632830.png (104 KB, 680x530)

104 KB PNG

>>109468066


Minimax prompting is even harder than LTX huh
You cant just make "1girl have sex with 1boys" and make it work. Your prompt need to be extremely detailed, even more than LTX did.

File: asdasasd4).jpg (24 KB, 281x258)

24 KB JPG

https://huggingface.co/UbuUbu/Comfy_kitchen_wheel
Not sure about quality but kijai added int4 which is another 2x speed up vs int8

>>109468574
Isnt Int4 just like Q4 or even Q2 at that point ?

>>109468571
gibbon dog

>>109468582
apparently its a bit below int8, he kept some weights at higher precision

>>109468301
nice

>>109468574
you need blackwell to take full advatage of that right?

>>109468559
well LTX needs Loras for it to work and you might need a trigger word for the Lora. Minimax is working out the box for a lot of sex actions already with no Lora. I haven't found it particularly difficult to prompt unlike LTX where you had to prompt in a certain way, minimax seems to be far more forgiving

>>109468574
link for int4 ?

>>109468599
no, int4 works on like 99% of gpus

>>109468607
that page has links to his quants

>>109468604
You only need Eros 1.4
That checkpoint is perfect for porn

File: Krea2_SydneySweeney_Outpu(...).jpg (3.61 MB, 1664x2432)

3.61 MB JPG

hello saars
I have needfully broduced most bootiful Sydney Sweeney lora for the Krea 2
Blease to run around 0.8 to 0.9 strength
Trained at the 1440p resolution base
https://limewire.com/d/u7syB#OC2j4p8gnv

>>109468574
I want to try this but right now there's only the t2v model there.
Did he say anything about uploading the ref model version soon?

https://huggingface.co/disguisequence/MiniMax-H3-10Eros-Max-Quants

curious what is this?

no really, you can do anything with the reference model. ANYTHING. I added music to this for fun.
>suddenly, a spinning green ball slams into <Subject 2> on the cheek making a thud sound, and it continues to spin in place.

https://files.catbox.moe/358f8q.mp4

File: 1759017147314034.jpg (413 KB, 1062x531)

413 KB JPG

>>109468066

Honeymoon period is over.
Minimax
+ UNCENSORED
+ Extremely accurate face and artstyle
+ Can do anything what your prompt said......

- SLOW
- Cant generate penis and vagina
- Penetration movement still too basic
- .....Need detailed prompt

Verdict : Wait until better optimization arrives

File: 1212124).jpg (27 KB, 358x309)

27 KB JPG

>>109468712
>slow
its like 2x as fast as LTX now
Go to bandaoco discord.
use sol attention + kijai's cache + int4 convrot

>>109468667
I haven't tried the ref model yet but will get round to it, just using img2video at the moment. It is curious how they made them separate unlike LTX which was just the one model. I always found LTX to be very good with various audio inputs, using a small sample of a persons voice meant that same voice carried over to the rest of the gen very accurately. Minimax tho should be far more flexible and creative with the outputs but still maintain adherence to a prompt rather than LTX which sometimes would do some weird shit - that might be down to the prompt tho as LTX needed specific rules to be followed to work best.

>>109468739
>like 2x as fast as LTX now
Cmon there. LTX can do 720p 20 secs in under 5 minutes
Also the optimization nodes are killing the quality

>>109468739
and that is without turbo lora which is already being teased in the discord with examples. Its gonna be like 10 secs a gen on a 5090

>>109468754
no it cant. LTX downscales by 2x, formost of the gen which is why it was shit

I can do 0.5MP 124 frames 20 steps in 1 minute with H3

>>109468712
Well if you want porn your source image should already have the parts in place. No Loras are needed if your image already has everything. Wan and LTX can't do anything like that without Loras and it actually does know most porn stuff anyway. But sure, text2video probably will be a struggle, hence why you should simply use img2video or ref video.

File: Untitled.png (17 KB, 577x269)

17 KB PNG

what does this do

>>109468712
unironically such a skill issue + low iq issue it's crazy.
we don't even need loras whatsoever for NSFW because you can use references for everything. Poses, motion/movements, genital detail. Hell i fed it a specific image of my character butt ass naked rendered in a specific style from SDXL and it almost 1:1 copied over her genitalia. Only thing i felt like it fucked up was the shape of her anus but given the resolution i'm genning at + still learning, i think this model could very well perfectly do nsfw.
now hopefully that wasn't too many words for your brown skin and tiny brain to handle. 2ldr git gud.

>>109468644
nsfw h3 checkpoint?

>>109468782
if you have to ask you are a beta male and don't need this node

>>109468754
Are we still talking about LoserTX?

>Only thing i felt like it fucked up was the shape of her anus
Now this is high level gooning

1 image reference model test. remember, if you define the concepts early and reference them with the defined name later, it works very well. Otherwise the model has to guess. this would work better with a full body image of frieren btw.

ie: <Subject 1> is the anime elf girl in <Picture 1>.

0.3mp, 10s, 154 seconds with the optimization node stuff (sageattn kj, spectrum, etc). you CAN get good quality (and fast gens) at low mp.

https://files.catbox.moe/az9ddx.mp4

>>109468775
I did img2vid. Animations are really lame

See
Minimax
>>>/gif/30993805
LTX
>>>/gif/30993834

I tried everything.

>>109468774
Im pretty sure you dropped LTX too early but nowdays LTX is extremely fast

>>109468802
Prompting is worse than LTX. This is not a good thing

>>109468812
@ me next time pussy bitch, let's see your butthole gens then pal.

>>109468804
that's what i was wondering but they don't give much information perhaps to avoid getting nuked? I'm gonna download it in a bit and check it out.

HOW CAN I STOP H3 FROM TALKING OR ADDING MUSIC TO EVERY VIDEO FUUUUUUUUCK

>>109468836
prompting guide said now

"By default, if a prompt is about a cool action scene, the AI will automatically try to be helpful and generate an epic background music track to match it. Writing non_diegetic_music: N/A completely disables that specific musical background-score layer."
important info

>>109468802
yeh the porn stuf works without Loras and Loras will help of course and they will be more specific but this really is doing stuff LTX and Wan could not because it knows the physics better, it is pretty accurate. The problem with Loras is they can tend to set up generic actions based on the pretty small amount of training used whereas this is using all the training for the physics and will look more natural.

>>109468836
>non_diegetic_music: N/A

File: file.png (45 KB, 707x234)

45 KB PNG

>>109468644
>>109468826
Which version is suitable for 24GB of VRAM? The 19.5GB pruned int8 convrot version seems to take 99% of my VRAM already, I'm not sure exactly what fills the remaining 4.5GB/if any overflow will get offloaded without losing too much speed

>>109468281
whats the prompt look like in this case? no mention to video at all?

>>109468712
Oh look, its ZITFag.

Any tips on prompting? I'm struggling, I have severe prompt block.

>>109468847
>>109468852
Wait, prompting guide? First time hearing about that. So it was a skill issue on my end, damn. Thanks, anons.

>>109468865
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md

Read them, super useful, much better gens

>have you ever punched someone so hard he-
https://files.catbox.moe/34as5u.mp4

>>109468870
also can feed them to the TE
https://github.com/ethanfel/ComfyUI-H3-Qwen3VL-TextGen
and have it enhance your prompt

>>109468836
ez
non_diegetic_music: N/A

My current problem is now the character always talking gibberish all the time. Sometimes in gibberish japanese and korean language

alternate universe evil Goku (so black):

https://files.catbox.moe/ir5fov.mp4

>>109468878
Did you put something like "falls like a board" in the prompt

>>109468860

<subject 1> is a girl from <picture 1>
the completely naked short black man is standing with waist slightly bent over, body facing the left side, like shown in <picture 2>
subject 1, kneeling down on the floor beside the man facing the man.
she looks at the camera for a short time then turn her head to the man's ass, her right hand is on the man's right thigh while she move her left hand to hold the man's left thigh.
she open her mouth wide, stick out her tongue, she move her mouth and nose deep between the man ass cheeks, then move her head around repeatedly, licking passionately, giving a rimjob. she move her head up and down repeated with tongue out.

>>109468891
More like "turns into a wooden door"

>>109468823
Well at least the balls didn't touch. I did say Loras will help but this is straight out the box and it works and LTX (at least with the Lora I used) did do that type of action pretty good whereas say BJ's and facials it was very bad (again using Loras). LTX was bad at stacking Loras as well, as minimax already works I suspect Loras used won't even need much, just help along the actions req.

>>109468891
hello newfren :-)

>>109468739
>sol attention
How do I get this shit to work? I just get the same error as this post.
https://github.com/kijai/ComfyUI-SolAttn_triton/issues/5

>>109468900
do i need to put all the refs on 1 picture ?

>>109468878
>3 scene changes in 5 secs
are you Chris Nolan?

>>109468739
>int4 convrot
VRAMlet

>>109468931
nah, I want as fast as possible gens, then once I find a good prompt / seed then I can just upscale / 2nd pass it

File: 0805232307852-i9iauZPFYI2.png (588 KB, 1479x1122)

588 KB PNG

>>109468923

reminder to follow the ways of chudha.

https://files.catbox.moe/ggft2j.mp4

>>109468878
based kino but you need more than 5 seconds for multiple scenes unless you are a vramlet

>>109468947
>niggers don't tongue anon's anus
>anon tongues NIGGERS' anuses
i see.

>>109468947
This is the saddest thing I've ever seen

Rape All Newfags

>>109468947
thanks dude.

>>109468809
because people are making that level of quality with snake oils

>>109468949
peak cinema

i am new. i am being promptfiltered by h3. it hurts.

>>109468301
thank you for the eye bleach

>>109468969
Yeah prompting is so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so fucking hard. Its worse than LTX style of prompting

>>109468984
git gud, I2V really isn't that hard.
Or just enlist an abliterated LM.

>>109468903
I didnt use any Loras. Just 10eros 1.5. If i got worse results i just use 10eros 1.4 since its more stable

>>109469005
Blue. BOARD!

>>109468987
You have to describe every little action, for everything, in a more profound way.

>>109468996
Can yo do penetration ? Post it on /gif/

why are the face details so blurry?
minimax h3 ref, set up mp at 1.2
>>30998506
nsfw

>>109468984
get filtered pleb.
H3 is the first model to ever does exactly what it is prompted to do, in a quite literal sense of each word.

>>109468739
what node pack is this from? i've updated everything and don't see it anywhere.

>>109468997
All the fixes that Wan and LTX needed and minimax just werks out of the box to a level far far beyond what they could never dream of. And I actually do think LTX is valid still as it is faster and in specific cases like the audio in and using it within it's limits it's still good because that is how it should be viewed - a tool to be used for specifics. But it is clear minimax is going to be the model used and it will get better and better and faster. And LTX will start to see a fall off things being updated. But let's face it, people still use Paint for things. If you have things that work then you may well not see an improvement in minimax, sometimes the models will inject a style you don't' want.

>>109469055
no one knows

I plugged in the new gundam series promo (xarx-zero) art as a reference, it worked.
><Subject 1> is the anime robot in <Picture 1>.

https://files.catbox.moe/rt3qtm.mp4

File: 1758709748886371.png (2.08 MB, 1408x768)

2.08 MB PNG

>>109469038
Not really.
If you able to make this Megumin suck the popsicle WITHOUT eating it. Then i kneel. Because no matter what i do she always eating the popsicle.

File: computerworld.mp4 (2.4 MB, 736x576)

2.4 MB MP4

https://files.catbox.moe/q5k0q2.mp4

>>109469085
People adapting to the next big thing while the old thing still matures is baffling me.

>>109469098
lol'd this is what this technology was really made for
well that and bobs vagene but mostly this.
not a fan of the fast pacing but i'm sure that can be fixed one way or another innit.

>>109469037
https://i.4cdn.org/gif/1785943979273915.mp4

>>109469115
>that ass shape
holy cannoli muh dik its way too early in the morning for me to be diamonds

>>109469098
is this one gen or two gens combined into one ?

>wanschizo

>>109469115
Can you not post this shit on a blue board anon

>>109469109
minimax doesn't have enough of a hurdle for most even regarding the speed to not be used over LTX. The good thing of course is people can use both if they get in that mindset, but I know that minimax will have tons of improvements, Loras so I wonder how much longer LTX will be valid even for it's biggest fans?

Total Newfag Rapening

>>109468774
>LTX downscales by 2x
You don't have to lol. my workflow genned at the final res for all passes and it was still way faster than minimax.

File: 1550208279467.png (40 KB, 644x800)

40 KB PNG

>>109469007
>Blue. BOARD!
>>109469135
>Can you not post this shit on a blue board anon

>>109469127
It's one gen. You can apparently push H3 to 20 seconds and it still works fine.
Something is fucky with Comfy and H3 though. If I try to go past 0.4 MP, it starts to use less VRAM and is slow as hell. I think I have enough VRAM to go bigger, but it doesn't seem to want to use it.

>>109469156
faggot

>>109469115
Can you post more of this shit on a blue board anon (and share the catbox for this one)

>>109469159
99% sure this is an observed phenomenon everywhere, no one's vram is being fully utilized and i'm seeing 24niggerbyte cards going into shared memory.

File: 1641606438569.png (241 KB, 1440x1388)

241 KB PNG

>>109469164
>faggot

Not Wan, Not LTX, none of them could do it this good.
Xi Jinping I kneel
>https://files.catbox.moe/nks4xw.mp4

>>109469166
>Can you post more of this shit on a blue board anon (and share the catbox for this one)
fuck off stop encouraging him

>>109469096
If you prompt for an action similar to a BJ then it will work. Of course if that is the action you want or be more natural is how more detail in he prompt is needed.

>>109468739
>kijai's cache
what's this? link? I'm not going to join some trooncord server

google ai mode/grok/etc are good for making some of the more elaborate prompts without typing an essay yourself

ie, make a 90s infomercial for a pasta maker with hatsune miku.

>>109469186
tried that too. Failed.

>>109469219
it worked but holy shit I need to bump it to 15s or reduce some cuts kek, used a reference miku even though I know it knows miku natively, still in reference workflow.

https://files.catbox.moe/kcbk3e.mp4

File: MiniMax_H3_00291_.webm (1.85 MB, 736x736)

1.85 MB WEBM

https://files.catbox.moe/tac657.webm

File: 1782666941758968.webm (1.29 MB, 1440x816)

1.29 MB WEBM

>>109469186
forgot webm.

Also another Minimax weakness, When it detect anime artstyle. The animation is always in low frame rate. I know they trying to emulate anime but i prefer the realistic movement in Anime. And no matter what i do i cant change it

>>109469098
Awesome

>>109469263
It's August not July you fucking retard faggot

>>109468349
black male ass rimjobs is weird, but I'm glad that it happened because it helped divide porn where it's about the woman being degraded / ugly versus porn where it's about the woman being hot

>>109469246
okay, specifically said a 10 second prompt this time with a similar scenario. much better.

https://files.catbox.moe/y0z5u6.mp4

File: 1782626234787497.mp4 (1.8 MB, 864x480)

1.8 MB MP4

t2v is fun

anyone here using 4060 ti 16gb with sage attn for H3? is it worth it? I feel like I just dont want to install it and mess with my current comfyui. Something might break.

>>109469289
Would be funny if the logo said Comfy.org HQ.

>>109469294
to clarify I'm not using sage attn right now and it is slow as fck

File: 1673413841964530.png (266 KB, 785x1000)

266 KB PNG

>>109469277
>It's August not July you fucking retard faggot

>>109469267
I am going to suggest you didn't' put the actions of a BJ, so like rapidly moving the popsicle backwards and forwards between her lips. Minimax can easily do BJ's so you have to prompt for that like:

>she then moves her head towards the popsicle and she repeatedly slides the popsicle in between her lips and bobs her head back and forward in repeated motion all whilst a slurping and lip smacking sound is heard

>>109469310
and put
>her eyes roll into the back of her head in ecstasy
for added lols

>>109469263
God I love yuri

>>109469267
>but i prefer the realistic movement in Anime
There is no 2d anime that does "realistic movement". Just ask it to do 3d or game cg style. Or, suggest it to mimic the style of a specific animation studio.

I was getting cuda memory access errors whenever I tried to gen 1mp videos with h3 that were longer than 12 seconds. anyone else experience this? I had Claude look at it and it identified a potential issue and made a monkey patch that seemed to fix it and I'm able to do longer 1mp videos now (on an rtx 6000 pro). It's getting some strange stuff happening beyond 25 seconds and I'm not sure if it's just the model itself or the fix.

>>109468281
>all the transphobic chuds in the replies

>>109468611
not nativelyretar

https://files.catbox.moe/lyetbn.mp4

File: 1758809843297732.webm (1.28 MB, 1440x816)

1.28 MB WEBM

>>109469317
>>109469310

Same thing. This fucking bitch always eating it. Minimax just wouldn't listen.
They overtrained the eating part and anime part.

>>109469329
I tried that too. Failed.

>>109468611
Pretty sure only blackwell has int4 hardware acceleration

jfc ive been adding so much shit from these threads to my workflow that im getting dizzy and dont know what is right and what is unnecessary. can someone write a damn guide with the best practices already

File: MiniMax_H3NoAudio_00005_.mp4 (2.51 MB, 1120x832)

2.51 MB MP4

>That shoe
Fuck so close

File: Flux2-Klein_00526_.png (2.36 MB, 1248x1664)

2.36 MB PNG

she might press it

>>109469263
ack won't like this...

>>109469358
>can someone write a damn guide with the best practices already
We don't know what is best practices yet, and it's pointless without seeing the first generation of turbo loras since the guide might become invalid in a week, it's too early

<Subject 1> is the anime girl in <Picture 1>.

what I asked grok to make: make a 10 second video prompt for an anime girl making fun of shameless tiktok and instagram onlyfans girls in a satirical way. use <Subject 1> for all instances of her being mentioned

Cut 1 (0–3s): <Subject 1> sits in front of a phone on a tripod, dramatically flipping her hair and making exaggerated “come hither” faces while fake-whispering “link in bio ”. She suddenly breaks character, rolls her eyes hard, and deadpans at the camera.Cut 2 (3–7s): Rapid montage of <Subject 1> mocking common thirst-trap poses: over-the-top lip bites, awkward “casual” mirror selfies, and fake-shocked “oops my towel slipped” energy — all done in an overly theatrical, ridiculous way. She keeps interrupting herself with sarcastic voiceovers like “Wow so spontaneous” and “Totally not selling anything.”Cut 3 (7–10s): <Subject 1> drops the act completely, crosses her arms, and smirks at the camera. She holds up a phone showing a fake OnlyFans-style bio, then tosses it aside while mouthing “yikes.” Final freeze-frame on her unimpressed expression.

https://files.catbox.moe/i4u1q6.mp4

Whenever someone figures out how to merge the FL2VA and R2VA models, we'll be in business. I want to be able to use a reference, but also an exact first and last frame.

>>109469358
there's no need for any of that stuff at most you will cut down the gen times by 20%.

did anyone get pwned by the latest npm attack?

>>109469358
default workflow works for me?

>>109469350
alright hold the fort, i dug up an old gen of stocking anarchy sucking a popsicle from the trenches of my archives, gonna serve you up a successful popsuck. give me 30 minutes.

this shit is egregiously slow at 60s/it even with all the bells and whistles of optimization so something must be fucked with comfy in the same way a lot of other people are getting, the vram utilization. oh well. 2 more weeks until turbo.

File: 1779800401798464.webm (1.73 MB, 1280x704)

1.73 MB WEBM

>>109469317
>>109469310
>>109469411

LTX 2.3 (10 eros 1.4)
Night and day difference. Smooth 3D animation too. Cmon Minimax i know youre good but this is just pathetic.

>>109469361
> the belt

>>109469390
Sounds like an easy job, R2VA is like 99.9% the same as FL2VA, probably some block merging or transfers or some other discovering in the future could make this possible.
In any case I think the R2VA may be able to use a pic or video as a starting frame if you tell the model so.

File: 1779541490422125.png (129 KB, 1326x473)

129 KB PNG

>>109469411
try this setup, fairly fast at 0.3mp/10s

>>109468349
thats what you get for shit talking footfags. enjoy the floyd bbc rimjob spam

>>109469358
My current takeaway
>use memory efficient sage attention
>sol-attn degrades quality
>easycache degrades audio (it just skips steps when the rate of change is under the threshold)
>you can use other samplers if you change the sigma shift (6/6 is a good start)
>shift_video / shift_audio
>higher = better motion, more artifacts
>lower = more static motion, less artifacts
>if audio sounds underwater, increase shift_audio
>keep shift_audio as low as possible.

I have a workflow here for 00001mp and it beats anyone else's workflow in terms of speed

can all the anime posters here fuck off already? we don't want you here or your nsfw shit

>>109468281
why are mutts so obsessed with "eating ass" (literally licking poop)

>>109469443
>00001mp
0.000001*

>>109469390
You can prompt it in the rev workflow retard RTFM.

>>109469426
>In any case I think the R2VA may be able to use a pic or video as a starting frame if you tell the model so.
I've been doing this with completely lazy hand written prompts of just a couple sentences. it tends to change the camera angle slightly, otherwise yes of course it can use the image as the starting point.
perhaps more strict prompting following the guide regarding where the camera is etc. would help, but really it works anyway

>>109469414
Hmm curious how you made that gen 7 secs and the other 5 sec. Could it be you genned for 20 secs as in LTX and trimmed that? Also that looks pretty bad, not natural like some terrible vidya animation motion. I know full well Wan, LTX and minimax can do the actions whether it's a popsicle or a dong

Bro, rimming is actually fucking disgusting.

>>109469462
It doesn't work as FL. It just uses the images as references, not a pixel perfect start end ending frame. I am not a retard, actually.

sol-attn default settings, on a 3090, 0.5mp 8sec
21s/it vs 15s/it
but it creates visual artifacts and object ghosting.
really not worth it.

>>109469450
maybe he's a leaf tho

>>109469484
or nothern european.

Nvidia Rubin can't come soon enough, I need a proper hardware speedup!

File: 1769974512562398.webm (1.85 MB, 1280x704)

1.85 MB WEBM

Im not dissing Minimax, it just Minimax is terrible at making static animations.

>>109469470
The unnaturalness its what make it charming. Have you tried to watch Anime Hentai ? Shit looks unwatchable. This is why smooth motion helps. The point is Minimax is terrible for porn. For now at least. You dont need cinematic movement in porn.

Also i should change the popsicle color into white lol

>>109469449
anime website

https://files.catbox.moe/cmeksa.mp4

>>109469510
minimax has that wan rigid motion look. not all gens have it, but it's very easy to get it.

>>109469462
You can prompt it and it will get close with a good prompt but it's not exact frames.

>>109469450
you won't get it. it about degrading woman

File: file.png (6 KB, 201x45)

6 KB PNG

I'm tired, bros...

okay refining the prompt idea with grok worked nicely, just tell it to assign <subject 1> and <subject 2> to the characters.

<Subject 1> is the anime girl in <Picture 1>. <Subject 2> is the man in <Picture 2>.

Cut 1 (0–3s): Wide daytime shot of a busy New York City street. <Subject 1> and <Subject 2> face each other in the middle of the road as cars swerve around them. Both drop into classic Dragon Ball Z fighting stances, power auras flaring up.Cut 2 (3–7s): Extreme rapid-fire exchange of punches. <Subject 1> and <Subject 2> blur into a flurry of fists, each hit creating bright impact flashes and shockwaves that crack the pavement and rattle nearby buildings. Camera cuts quickly between close-ups of their determined faces and the lightning-fast punches.Cut 3 (7–10s): Final powerful punch from <Subject 1> connects with <Subject 2>, sending him flying backward through the air past skyscrapers. Dust and debris settle as <Subject 1> lands in a confident pose on the cracked street, aura still glowing. Freeze-frame on the impact.

https://files.catbox.moe/gmb6uy.mp4

>>109469414
why are you comparing a finetune to a base model. are you trolling?

>>109469267
>>109469414
ts is what's bugging me with H3, if you put an anime reference it'll keep on making choppy animations, if you put a 3d cartoon reference the movement and facial expression are exaggerated, I just want realistic smooth animation and facial expression damn it.

its up
https://huggingface.co/ssjenforcer191/Homelander_Minimax_H3_experimental/tree/main

File: 2026-08-05-123551_741x421(...).png (298 KB, 741x421)

298 KB PNG

if your gen has these, I don't want to see it.

>>109469561
why bother training character loras? references work well

>>109469572
will never work as well as being trained on them. Also reference slower

File: 1776130166851539.webm (2.61 MB, 832x1280)

2.61 MB WEBM

>>109469519
>>109469548

Until then, ill stick for LTX 2.3 for Porn
Minimax for making memes
WAN is dead. Im deleting it soon. Fuck those chink jews making their models cucked online only.

>>109469553
Yeah thats why in my opinion, Minimax is overrated.

>>109469471
it is, but you will never convince these people otherwise because they're just in denial at being bi or gay.
there's simply no appeal if you're straight, even if you are into degredation

>>109469572
Micro details

--enable-triton-backend worth on 3090?

>>109469548
NTA but some anons here think you don't need niche loras anymore for H3 because all you need is reference. Minimax still has flaws.

>>109469572
Model broad knowledge is like an unfocused picture, a lora will bring it into focus improving body mannerisms and voice.

>>109469563
I think using reference video gives you those because comfy is shit

>>109469600
yes?

https://files.catbox.moe/q6z6c3.mp4

>>109469607
no it's just the sampler shitting itself because you have too little res and too much motion. try with 6/6 sigmas

this shit is nuts
https://www.reddit.com/r/StableDiffusion/comments/1vg9od9/the_office_manifesto_h3nsfl/

>>109469542
potato rig or massive gen?

>>109469629
Man people really really want open source to get banned. First fucking uncensored model and everyone proves immediately why models get censored.

>>109469629
>Over 18? Please log in
fine, old.reddit.com
>Old Reddit? Please log in

fuck off

hmmm. 25steps+ez cache let's me bump the res to 0.7 but takes the same amount of time as 0.5

but the gen now looks better than raw 0.5

File: 1000019353.png (29 KB, 248x204)

29 KB PNG

>>109469542
You are like a little baby.

Which video model could I run with 8gb vram and 640kb ram?

Funny that like year and half ago i was waiting dozens of minutes to gen some 0.4 mpx crap with Wan 2.1.

>>109469669
billy mayes?

>>109469414
Sex with Megumin (14)

>>109469670
Lmao

>>109469669
>almost an hour
that's probably fine if you just want to cook a gen you already know is good at maximum quality but most anons just want to gen good enough shitposts quickly.

>>109469583
>LTX 2.3 for Porn
Even Wan 2.2 is better

>>109469639
yeah some people are genuine fucking idiots. like the ones who make political shit

any guidance on shot transformation in h3? i'm trying to prompt it to expand from letterbox to full frame from a reference image.

>>109469691
>Slower than Minimax and LTX COMBINED
>5s max
>Abandoned by developer

Lol no

we might see a singularity when the turbo lora is out cause things are already pretty good.

>>109469669
What the fuck.
>>109469632
3mp 5 second gen, 96gb vram and a barebones workflow.

>>109469674
bout 30 minutes for 720p

>>109469669
I'm happy with my 480p 15 sec gens @15 steps taking 20 mins on my 5060ti 16gb. LTX was producing far larger file sizes as well due to being 720p

>>109469692
>yeah some people are genuine fucking idiots
about 99.9% of society to be more accurate

File: Screenshot_3542.png (30 KB, 1463x226)

30 KB PNG

tried 1mp, 5seconds

>make a 10 second video prompt for an anime girl and an anime man sitting in a booth with Jerry Seinfeld and George Costanza, having a conversation with the joke being AI is a show about nothing.

https://files.catbox.moe/hj8fg6.mp4

Unlike us, we're smart.

File: 1762874866280618.png (277 KB, 744x421)

277 KB PNG

>YOOOO THIS MODEL CAN MAKE CONTRAVERSIAL THINGS GUYIZEEE IT CAN MAKE POLITICAL GORE PORN COOM DOOM DURING A TIME WHERE THE GLOBOHOMO ARE SUING AI COMPANIES LOOOOK LOOOK WHAT I CAN MAKE EASILY LOCALLY FOR FREE I CAN BARELY TELL IT ISNT REAL LMAOOOOOOO XDDDDDD HAHA!!!
goyim prove they are indeed nothing more than cattle with more certainty every day.

>>109469711
>abandoned by dev
but enough about LTX

>>109469691
the only reason someone would shit on LTX is because they're too retarded to use it.

You like wan because it has all those ultra specific loras that work every time but make all gens look exactly the same.

where do I get the mem eff sage attention node from?

>>109469583
Man people got so local shitty AI brained that they genuinely crave the shitty AI motion...

>>109469629
reddit be like:
>this is too disturbing, ban it immediately
then also be like:
>can't wait for gta 6 so i can go on multiple killing sprees

>>109469735
LTX 2.5 is coming on december

>>109469733
:-)

File: 638575472.mp4 (3.65 MB, 1504x640)

3.65 MB MP4

>>109469553
Idk, this one I genned seems to me to have the appropriate type of motion for the type of anime it's emulating.

I put this at the bottom for the style
>Style: Classic 90s hand-drawn anime aesthetic, detailed cyberpunk city background, dynamic motion blur.

>>109469692
>>109469639
>>109469629
The guy who posted that thread is from Slovakia. So I don't think he really cares.

>>109469747
Lol one is a deepfake of a beloved TV show and the other is a video game. Smh

>>109469734
goyim you say?

https://files.catbox.moe/az9ddx.mp4

>>109469743
>t never jerk off to Live2D

>>109469743
I'm not one of those guys who posted, but honestly I can see why they'd be mad. it should be controllable easily. if you want the janky frames that's fine, but if you want the fluid frames you should be able to prompt for that

>>109469734
lol who gives a shit the model is already on my computer
about to gen globohomo kissin my fat white ass cause thats all they can do

>>109469756
pretty good.

>>109469734
It's because social media brain rot makes retards broadcast everything online to strangers like we're close friends or that it's not completely public. Then the complete moral erosion to not think twice about what they post either like "maybe having The Office goes postal with R-rated gore" is something I shouldn't post online.

>>109469761
>my heckin wholesome beloved TV show... defiled!

every single finetuner working on h3 has said that the model responds well to training. and will be the new wan 2.2 for nsfw loras

>>109469763
shes doing that on september 10, right before... a lot of people took sick leaves for tomorrow, right?

>>109469768
>lol who cares about the impending foreseeable famine I have my apple right now

>>109469778
kek

>>109469768
>lol who gives a shit the model is already on my computer
the next, better one has less chance to be, lil tyrone

>>109469482
I think you need blackwell to actually make use of it
https://github.com/Saganaki22/ComfyUI-sol-attn

>>109469743
cope

>>109469761
okay so both are not real, why should i be outraged?

>>109469784
it is fucked up that you have to redownload the model after every use

>>109469759
He'll care when they restrict Slovakia from using models because they're apparently uncontrollable.

File: 1783720886476387.mp4 (3.43 MB, 864x480)

3.43 MB MP4

>>109469297