>>109614991
>No seriously, replace this shitty collage script that plays full videos and holds the last frame or some shit. this is literal jeetslop edition
lmfao

fizzledork

>Maintain Thread Quality
https://rentry.org/debo
https://rentry.org/animanon

>>109614991
shitbake

>Maintain Thread Quality
https://rentry.org/debo
https://rentry.org/animanon

>in the age of vibegoding, nu-/g/gers cant make a script to collage a bunch of videos together without it breaking or raping bitrate
kek

>>109615014
THIS!!!

>>109615003
>>109615014
>>109615025
seriously. why apply duration (with a fucking max of 10s) to all the videos at once? why is the aspect ratio a single number? fucking hell

>>109615013
Also, I'm giving you a massive hint how to prompt Stable Audio 3 in all my posts. You have to include the genre in the prompt, there's no other way around it.

>age of vibecoding
>complains rather than fixes it himself
typical of the troll baker lol

>>109615060
Unless you want free tier model garbage I am not paying for claude.

>>109614991
Purchase a better computer and the collages won't freeze. It's literally because you're a poorfag.
>>109615066
And there is the additional poorfag confirmation.

>>109615059
It doesn't matter. In the default comfy workflow for Stable Audio 3, it generates horrible-sounding noise no matter what you type in for the prompt.

>>109615010
>>109615012
thanks anon

>>109615073
I put in the default values and that's what you get. Please show me your example

>>109615066
get a job, fizzledork. oh wait... you can't lmao

>>109615073
>Purchase a better computer and the collages won't freeze
that's not how ffmpeg works kek

>>109615073
>Purchase a better computer and the collages won't freeze
I see the tech let author has arrived

>>109614991
hahah

>>109615084
Look at the past two years of collages and you will see other anons using the script just fine. You're creating video collages with a userscript, you need a PC that's not total shit for them to have a good framerate.

>troll baker
>low spec computer
loooooooool you cant make this shit up

>>109615095
he only picks the same length of videos so it encodes properly. videos of different lengths break the fucking thing because you didn't think of that you fucking retard

>>109615080
>default comfy workflow for Stable Audio 3

I'm assuming you're using Turbo. I haven't tried it on Comfy, but I recommend you try its Medium Base model on official Gradio (with Flash Attention installed so it's fast). That's the only place I've tested it. To be frank, Comfy is not really a place to do audio, even gguf ACEStep and Minimax versions are magnitude times faster than on Comfy, like a 5x speed difference.

>>109615102
You are incorrect. And again, look at the past video collages and you will see different length videos being used no problem. Post your specs poorfag lol
Also not the creator of the collage script. There were at least three different anons who edited it previously.

File: Test_00020.mp4 (3.28 MB, 1376x768)

3.28 MB MP4

https://files.catbox.moe/0kup95.mp4

>>109615109
I've used both. They both sound like utter shit. Only the example prompt generated music-like sounds.

>>109615100
We all knew he can't gen that's why he spends his time trolling /ldg/

>>109615088
its not ffmpeg its native browser APIs my little tech techlet friend

>>109615186
>native browser APIs
yeah i see your jeetness now. why the fuck would you do it that way? everyone here has ffmpeg installed

you wish i was the creator so you could continue to try and troll me kek but keep crying

Actually, if you search the archive for the first mentions of the script, I think the creator said why he did it the way he did desu

>>109615207
This is still pretty Jeet-coded of you for letting it slide but you are the author since you are crying so hard

>>109615217
V2 is tran vibe-shitting

Why is this troll baker still pissing and shitting himself ITT? He got the bake yet he continues to mald?

>>109615138
Yep, that's an indication you're prompting it wrong or perhaps you don't know what kind of music you're after.

>>109615233
tran can't help himself

>calls anon a jeet
>when his computer cant even make a video collage with a framerate higher than 1 frame per minute
kek

>>109615233
you can see how pissed he is with his reply >109615265 attempting to blame someone else. a classic move from this schzio troll.

File: ldg baking guide.mp4 (3.59 MB, 768x1376)

3.59 MB MP4

>>109614991
Almost there. you just forgot the last part. no worries. you'll get it next time! :)

Maybe this is a stupid question but if running the script is too intense for someone why not just toss a bunch of webms frot he previous thread into OBS overlapping eachother in a gigacollage, then record a few seconds of that and post that as a webm instead..?

>>109615334
The one who baked this troll thread is too retarded for that. But also, he's the only one whose ever complained so it's just a consequence of his computer being really old and shitty.

>>109615334
>why not just toss a bunch of webms frot he previous thread into OBS overlapping eachother in a gigacollage
an anon used to do that before the script was generated
i kinda miss the "handmade" collages honestly the ones created with the befunky website too

Still getting color shift with "Add guide for MiniMax H3"
what the node needs is a gradual guidance drop off. to hopefully make the color shift less jarring.

>>109615217
the guy crying because his pc cant handle a userscript DEFINITELY doesnt know about the 4chan archive kekek

Man, all I want is stable clip extending. is it too much to ask??

>>109614991
Is there any t2i model that can generate maps of cities or villages where I can specify I want a castle surrounded by so and so etc.?

File: int8_00021_.png (1.58 MB, 832x1216)

1.58 MB PNG

>>109615464
No, it just takes some elbow grease
https://files.catbox.moe/wkzq06.mp4

>>109614991
>mfw Resource news

08/21/2026

>MiniMax H3 Super Acceleration fast draft generation and high-resolution refinement, powered by Sol Engine
https://nvlabs.github.io/Sana/Sol-Engine/H3-Super-Acceleration

>4DAnyone: Create Anyone in 4D from a Casual Monocular Video
https://4danyone.github.io

>H3 Prompt Composer Version 5.37.1
https://github.com/BMB12d3/minimax-h3-prompt-composer

>LTX-2.5 Gemma-4 12B NVFP4 for ComfyUI
https://huggingface.co/Deadshot699/ltx-2.5-gemma4-12b-comfy-nvfp4

>MiniMax-H3 Pruned Ref-Delta Fused r1024 — ComfyUI Single File
https://huggingface.co/xmarre/MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-ComfyUI

>Phosphene 4.6.0 Adds Video Editor, H3, LTX 2.5
https://github.com/mrbizarro/Phosphene/releases/tag/v4.6.0

>H3 Optimizations: Standalone production optimization nodes for MiniMax H3 in ComfyUI
https://github.com/Zironic/H3-Optimizations

>MiniMax H3 known character lists
https://huggingface.co/datasets/malcolmrey/various

08/20/2026

>MiniMax-H3 Turbo 4-step v1.1 by Lightx2v
https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main

>MiniMax-H3 Turbo-SLA
https://huggingface.co/lightx2v/Minimax-h3-Turbo-SLA

>MiniMax-H3 Single-Frame VAE 500K
https://huggingface.co/iamkaikai/MiniMax-H3-Single-Frame-VAE-500K

>MiniMax-H3 RAVEN Streaming LoRA
https://huggingface.co/mvp-lab/MiniMax-H3-RAVEN-Streaming-LoRA

>Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models
https://pardistaghavi.github.io/SparsePR-website

>VA-Judger: Reward Modeling from Human Preference Feedback for Joint Video-Audio Generation
https://github.com/ShareLab-SII/VA-Judger

>Frozen DINO Localizes Image Edits Without a Localizer
https://github.com/VishalJ99/trail-image-edit-localization

>FiSeR: Fine-Grained Source Representations for Cross-Domain AI Image Detection
https://github.com/heyongxin233/FiSeR

>SkillNet: Create, Evaluate, and Connect AI Skills
http://skillnet.openkg.cn

>mfw Research news

08/21/2026

>VGI-BENCH: Probing Visual Intelligence in Video Generation Models
https://arxiv.org/abs/2608.19583

>WithEveryone: Unified Planning and Identity Grounding for Group Image Generation
https://arxiv.org/abs/2608.20336

>Stream4D: 4D-Consistency for Streaming Autoregressive Diffusion Video Models
https://banyuanhao.github.io/Stream4D

>Scale-Separated Conditioning for Style-Encoder-Free Diffusion Stylization
https://arxiv.org/abs/2608.19719

>BeyondMasks: Evaluating Causal and Physical Consistency in Video Object Removal
https://yigitekin.github.io/BeyondMasks

>Swift-Image: Exploring the Performance Frontier of Compact Unified Image Generation Models
https://arxiv.org/abs/2608.20334

>AvatarDynamizer: From Static to Dynamic Human Avatars via Generative Dynamic Textures
https://vcai.mpi-inf.mpg.de/projects/AvatarDynamizer

>DIFFCZSL: Compositional Zero-Shot Learning Regularized by Diffusion Representations
https://arxiv.org/abs/2608.19871

>A Plug-in Interpretation of Conditioning in Score-Based Diffusion Models
https://arxiv.org/abs/2608.19504

>Clustering and Token Denoising for Faster and More Robust VLMs
https://arxiv.org/abs/2608.19285

>Block3D: Efficient Text-to-3D Generation via Block-Wise Diffusion
https://alexandertsui.github.io/block3d

>Continuous Adversarial MeanFlow Transfer
https://arxiv.org/abs/2608.19540

>When Guidance Goes Off-Scale: Recalibrating Diffusion Transformers under Analog Compute-in-Memory Nonidealities
https://arxiv.org/abs/2608.19644

>DPC-Net: Dual-Prior Collaborative Network for All-in-One Image Restoration
https://arxiv.org/abs/2608.20141

>Orthogonal JEPA: Factorized Predictive States for Latent World Models
https://arxiv.org/abs/2608.20065

>When Saying No Makes Better Videos: Designing Dual Gatekeeping for Pedagogically Grounded AI Content Creation
https://arxiv.org/abs/2608.19812

sulfur 3 status? anyone in their trooncord?

>>109615387
If you used your brain nigga you'd realize people are genning while posting

>not being born rich so that i can just test out all of my ideas and goon 24/7
kek, meme existence.

>>109615507
I guess in my case what fucks it up is I'm doing 2pass.

how do I check what commit I'm on before I pull, in case things go fubar

>>109615618
There's something called git reflog, go learn about it.

>>109615618
>in case things go fubar
thats not a real thing, just pull

>>109615587
I believe the culprit was the turbo lora on the second pass.

still waiting for anon to show us how the latest extender cope worked

>>109615475
Some image models might obey that instruction sometimes, but you could use Ideogram 4 to layout your city with bounding boxes and have control where each part of the city should go that way.

>>109615844
I'm probably who you're waiting for.
without any cope nodes it just straight up works perfectly. otherwise I think turbo messes it up.

>>109615867
>turbo messes it up.
damn, i need turbo

File: deborror.jpg (121 KB, 768x768)

121 KB JPG

i don't like where this is going

>>109615844
nta but i need a clip extender that just makes sense, ie each extended section is its own prompt to steer the scene, just have that for t2v alone would increase my kino output

File: AnimateDiff_00033-audio_s(...).png (3.06 MB, 1918x1090)

3.06 MB PNG

>>109615867
If you can automatically incorporate storyboards of previous clips as input I think it helps consistency

>>109615867
one niggle, which isn't really anything at all but if you don't cut frames to a complete chunk length then it can cut the scene or do unexpected things
other than that, flawless

>>109615934
it may help character/scene consistency but it won't do anything for continuation coherence (motion) between clips

File: mini_00001-audio_storyboa(...).png (3.3 MB, 1918x1072)

3.3 MB PNG

>>109615945
The wf I use for extensions maintains motion consistency through latents I think, it's a total mess though

see here for like 5 extensions together >>109615507

Although for shots like pic rel it's probably not needed very much

>>109615867
>without any cope nodes it just straight up works perfectly
What is "it"? You need nodes to extend clips, if you use ref2va it won't be seamless (and you get the lower quality of ref2va unfortunately).

Personally I tried Contex-Loop nodes but it was fiddly and chaining still fried the clip very quickly. I've almost given up but it might be worth trying ComfyUI-H3-Motion-Context-MultiRef as it seems to use a different approach.

>>109615973
would you share the workflow?
also, has this worked better for you than the extender node which adds the frames in latent directly?

this is the key to looping yeah? just use the same clip for start and end

File: int8_00034_.png (1.22 MB, 832x1216)

1.22 MB PNG

>>109616013
I think this is doing the same thing, just in their own workflow. I found it on civitai a few days ago

This post has the wf and some details
>>109606936

vibecoded H3 node just got a big update
https://www.reddit.com/r/StableDiffusion/comments/1vuty3p/big_update_to_the_free_minimax_h3_prompt_composer/

>>109615587
Well, Nvidia is too, fricking "MiniMax H3 Super Acceleration" is h3 with a turbo lora, sol attn and LTX upscale

>>109615529
>https://nvlabs.github.io/Sana/Sol-Engine/H3-Super-Acceleration
was gonna ask if anyone had tried the equivalent in comfy, but even their own samples look like shit

File: 1774838182679087.png (129 KB, 360x360)

129 KB PNG

adding a character in h3 seems very difficult, if not impossible. i wasted 2k points on civitai: i got the right cosplay, but not the right face...

spending another day losing my mind trying to extend a clip.... I swear this can't be healthy.

>>109616139
post the character and ill gen you a SPH video

>>109616174
didn't you see? anon has it all figured out

>>109616191
I don't believe him that he's extending without flashes.

Like seriously, if you're extending on a cut that's cool but I'm trying to extend on the same static shot.

>letting ani bake again

>>109615556
?

>>109615529
>>H3 Optimizations: Standalone production optimization nodes for MiniMax H3 in ComfyUI
>https://github.com/Zironic/H3-Optimizations
would it kill them to give actual comparisons and numbers, holy shit

>>109615047
The max is 15 and you can input ints into the AR lmao

>>109616222
i dont either because he doesn't upload examples

>minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors
ok I just tried this and it's really ass. please stop using it. v4_step600_ema_pruned is MUCH better.

>>109615766
Did you prompt for a text like that?
I mean it being transparent and behind her.

>>109616299
>Across the background, there a fancy blue slightly transparent text with various graphical effects with the text "PSYLOCKE".

>>109616321
Thanks!

So what's the current turbo meta? 1.0 strength? 4 steps? 6? 8?

>>109616038
the slight buttchin reminds me of bad memories but I like the tight+gravity situation

closest I've gotten my cut to seamless
https://files.catbox.moe/1mxj3n.mp4 (NSFW)

>>109616354
The turdbo meta is to not use turdbo

>>109616361
I feel like if you encode the whole thing with h264 with shitty compression it might make the cut disappear

DO NOT PULL
https://github.com/Comfy-Org/ComfyUI/issues/15784

>gaypi nodes

>>109616373
I pulled master 2 hours ago and everything is fine for me.

>>109616373
>literally a fix right there in the link
im pulling and can never be stopped from pulling daily

I pulled so hard it broke off and now I have to go to a penis doctor

Does anyone here actually make anything useful?

>>109616403
some anon from here made an open source, node-based program called ComfyUI and became a millionaire

>>109616403
Define useful in the context of vid/img gen.

>>109616403
I make useful stuff for myself all the time.

Is any of it useful to you? I don't know nor care.

>>109616403
yes

File: MiniMax_H3__00239.mp4 (2.26 MB, 448x672)

2.26 MB MP4

someone gave her some witchnip

File: 00316-4194679814.png (766 KB, 832x1216)

766 KB PNG

I've barely touched any of this stuff in 18 months, what have I missed?

so whats the magic prompt to make someone NEVER SPEAK in H3

>>109616414
comyanon abandoned us as soon as he achieved the slightest amount of success. we weren't his foundation, we were just a gapfill

>>109616435
anima
krea
h3

basically the entire current meta

I mean a map of the globe made with gen ai would be useful, massively inconvenient and probably inaccurate, but useful in an extreme pinch

>>109616432
too bad the resolution fucks up her face
how do you make the speedup and slowmotion, you just prompt them?

>>109616414
Here as in itt not some fag turned reddit or grifter

>>109616448
>make program
>program works
what even is a grift at this point

>>109616446
At 5.60 seconds, she explodes upward again, this time launching herself across the room in a blur of motion, her robe trailing behind her like a comet’s tail, her staff blazing a trail of blue light through the air. At 6.40 seconds, she lands on the other side of the desk, back to the camera, and immediately drops into a fighting stance, one knee forward, staff raised like a spear, her hair whipping behind her in the wake of her momentum. At 7.20 seconds, she pivots on her heel and lashes out with a kick that stops inches from the camera lens, her leg blurring with speed, the force of it cracking the air and sending a gust of wind that ruffles the papers on the desk. At 8.00 seconds, she retracts her leg and spins again, her body a blur of motion, and as she completes her rotation she thrusts her staff forward, not at anything in particular, but at the concept of resistance itself. At 8.80 seconds, she comes to a sudden halt in a perfect pose, one foot planted, the other extended behind her like a dancer, her staff glowing brighter than before, her robes settling around her like smoke, her expression one of pure, unadulterated ecstasy as if she’s just performed the most beautiful act of magic imaginable.

>>109616435
>I've barely touched any of this stuff in 18 months,
I don't believe you.

>>109616455
Referring to putting all his chips in on you never owning new hardware ever again so you have to pay to suck his cock and some time on cumfartcloud and API node niggardry while selling your workflows and data

>>109616437
and i dont blame him, spend any amount of time here and you will see just what a retarded bucket full of crabs this place is. im talking some real abominable individuals that beggar belief operating here.

>>109616467
oh I see, basically you compressed the timeline so and added lots of actions in few seconds, interesting, thanks anon

>>109616472
I'm sure if you were king you'd do it much better

>>109616474
last I've seen him around here he got insulted by multiple tourette shitpost anons whenever he wrote anything, so good for him for making something useful instead of wasting time here arguing with bored schizos

>>109616474
Are you talking about catjack having doxxing meltdowns because Ani wants to make a c++ app?

>>109616476
yep

>>109616403
https://rentry.org/ldg-lazy-getting-started-guide#anon-guides-and-resources

>>109616437
he left because of julien tho

File: MiniMax_H3__00240-small.mp4 (3.76 MB, 832x1248)

3.76 MB MP4

bah.. fucked up the spin

uncozy breas

>>109616403
comfyui is for images/video. unless you're in the media industry, you're not making anything useful.

like are you expecting us to be making full blown anime on our (singular) consumer gpus, anon?

>>109616489
>wants to make
wasn't that years ago, did he never finish that wrapper?

>>109616566
yeah but only like 2 people ever used it. doesn't matter now anyway.

could probably get AI to write a wrapper in fucking assembly now. ani has no ground to stand on anymore.

4000 videos. time to go through and save the prompts i liked and delete everything

>>109616566
ani failed at everything hes tried
thats more noble that catjack failing to try anything at all

>>109616589
based, 4069 here

>>109616536
>832x1248
1MP seems to be the bare minimum for showing proper face.

Is zeta Chroma mothballed? I saw Kroma is a thing now and my tests with it have been okay.

I'm so bored of genning mindlessly
I wanna work on a project or something
but my rig is poor
life is suffering

I've clips extending I'm happy with now but I've been losing the seed lottery for the last hour.

Anyone been able to do decent cumshots with Minimax just with prompting ?

>>109616592
I want inspiring ani back. I want uggo tran to go

>>109616652
so far for me with ref2v
nudity without alien vagina/penis -> possible
sex scenes -> possible but copies a video
ejaculation etc -> nope

>>109616623
yeah.. anything less and it gets all LTX looking at any sort of distance or motion

>>109616664
Pretty much same, the model understands pretty much all sexual positions so if you feed it an image it will do the correct motions, if not handling the penetration correctly, but if you feed it an image of a girl in doggy style with a man behind, it will do the correct motions and it will look great

But not been able to get a decent cumshot going from prompt, mainly problems with the way it 'spurts', ah well, there's always loras

File: MiniMax_H3__00244.mp4 (2.7 MB, 416x736)

2.7 MB MP4

pretty annoying trying to get qwen3.8 abliterated to do anything that isn't NSFW with this pic..

>>109616654
ani got what he deserved
im not with the anti-ani schizos, but have you ever seen ani uplift anyone else?
he doesnt care about anyone else
i dont think he deserved what he got here, but i also don't think he got what he deserved
ani never figured out how to be a better version of himself
i think thats the worst part about him

>>109616681
I'm hoping the eros finetune can be used as a lora and tuned up and down to get me exactly what I want, aka not destroying the innate knowledge and qualities of the model but just add that nsfw it doesn't know about

>>109616683
that's nsfw as in weird

dis goofy ahh still yapping bruh

>>109616701
nah you are

>>109616701
>>109616695
(you)

>>109616703
ion you nga

File: MiniMax_H3__00248-small.mp4 (3.73 MB, 768x1376)

3.73 MB MP4

>>109616700

kino is cooking

can i interest anon in avideo of two girl kissing?

>>109616787
>>109616683
>>109616632
not as annoying as your yellow fever spam. nobody likes your ugly asian women

>>109616632
Pretty sweet.

>>109615323
Thank you, Hatsune Miku.

File: h3 embeddings.png (144 KB, 1258x443)

144 KB PNG

They added these to the Comfy repo.
They merged embedding support for minimax a few days ago.
Embeddings haven't been relevant since SD days, but still seems interesting. I would test but I don't want to pull just for that.

>>109616811
QQ

I still can't believe Stable Audio is so good. I have no idea what they did to their Github, it felt so shit when I first used it lol

>>109616867
did you try the soviet march?

>>109615118
reminds me of that japanese roguelike called Elira or something where you can be a little girl
gens like this make me happy knowing the videogames of my dreams will exist, eventually, i just need to figure out how to still give a shit about them when they arrive

>>109615561
if it makes you feel better i am this and can do this and i still don't and im bored all the time. maybe i should buy some sildenafil online and commit to goonmaxxing

>>109616888
Elona rpg and it's from taiwan

>>109616403
>Does anyone here actually make anything useful?
back in the WAN 2.2 era i generated thousands of clips and successfully diverted dozens to hundreds of pedo-hours that would have been spent masturbating to real children to AI generated children instead. i consider what i did to be the most ethical thing I ever performed in my entire and more impactful than if I donated my entire life savings to a children's charity.

>>109614570
which node to extract last 22 frames from a video?

>>109616899
chyna

>>109616851
>Embeddings haven't been relevant
Anon what do you think the TE does? Embedding files are just shortcuts for prompts so you don't need to type two paragraphs of text if you want a certain style or feature. They are always relevant.

File: jeeeeeeeeeeej.png (269 KB, 432x621)

269 KB PNG

>>109616908

>>109616623
>1MP seems to be the bare minimum for showing proper face.
yep. even 0.9MP is king of the copes. biggest turn off for me messing around with sfw and trying to make something "professional" like a short film / scene / music video

not sure but things seem a little faster since pulling

>>109616869
Here's two prompts attempting marching band using the description you gave last thread
https://files.catbox.moe/xyrvo5.mp3
https://files.catbox.moe/pzgqo6.mp3

Keep in mind the model can smoothly do every genre under the sun for EDM with seemingly no need for LoRA guidance, have gotten quite a few results I can't believe are AI.

who here is genning on a spark?

Has anyone gotten H3 to do good kissing? Like the sloppy, raw making out kind of kissing.

>>109616967
thanks. it does sound like marching music but it doesn't have that epic big sound that the soviets liked to compose with. it might help if you described it as "minor key, epic, dramatic"

>>109616982
It does that pretty well, otherwise you can always add a small kissing video.

mission: try to gen an h3 character shaking their head no without saying the word no
status: literally impossible

>>109616664
>nudity without alien vagina/penis -> possible
how? don't tell me you feed ref_image benis, and vagina

>>109616993
>g-guys look at me im an artist now! i can make abstract creative visuals please praise me!

>>109616992
yes, and it works, just feed the right angle otherwise it's weird

>>109616999
jelliboi

rapeman

i automatically skip every vidgen under 15s. if you're not breaking the multigen wall, you're not a real genanon

>>109616993
Very nice, almost reminds me of old limits/summoning in jrpg.

>>109617018
>multigen wall
what's that?

rtx6000 pro for $14k... worth it? that's still $2k off MSRP right?

>>109616985
Here it is with those words added in
https://files.catbox.moe/uqyuo0.mp3

>A grand Soviet military march in a minor key, epic and dramatic, blaring trumpets over heavy tubas with reverberant snare rolls, thundering bass drum and crashing cymbals, recorded in a vast hall 122 BPM

any update from that anon about animating doujins?

>>109617023
you don't vidgen, so dont worry about it

>>109617024
It's not worth it in any case.

>>109617027
hes in a goon hole, forget him

>>109617024
No. They were like 9k three weeks ago

>>109616993
Slop

>>109617033
i know.. i was gonna buy one for $6k last year but i didn't obv

Oai needs to die already. Same with oracle. If these fuckers get bailouts there will be riots. Gimme my hardware back

>>109617025
eh, i guess you can say it fits everything about the prompt except for the soviet part. i am not sure how i would describe it besides having a lot more dissonance compared to western marching music

>>109617014
spiderman piledrive most kino moment in cinema

>>109617024
why exactly is it worth spending that money when in essence you are not genning anything better than someone running a $500 card or even less? You might have better resolution but that's about it. So aside from the speed you are not gaining anything else. You can't really run any longer gens anyway because the model starts creating errors. This is like having a 5090 and running most AAA slop, you are still stuck to playing that same slop as everyone else.

File: debo_wr_k2_00102_.png (2.45 MB, 2048x1101)

2.45 MB PNG

>>109617046
oracle is already dead. its just a debt-corpse waiting for maggots to eat it

>>109617049
better resolution, longer gens, can run bigger models at the same time without unloading.. also good for running local llms

File: Flux_00511_.png (1.55 MB, 1024x1024)

1.55 MB PNG

>>109617014

perhaps kino
perhaps sovl

>>109617047
Is it some kind of subgenre of cinematic music? You can always train a LoRA on multiple songs if that's the case, that base result is close enough given its broad description

File: 2026-08-22-000846_383x451(...).png (55 KB, 383x451)

55 KB PNG

>>109616911

>>109617053
Yeh that would be great if actually the gens would be improved and by improved i mean something a lowly $500 card could not do. But the model isn't designed that way, it is designed to run on those lowly cards and the genning is going to be the same, the content is going to be the same. It's like Chris Nolan presenting his latest film in Imax and smiling whilst someone has already seen that same pile of crap in 480p on a phone. At the end of the day the model is going to reward people with ideas, not people with a $14k card and running off 20 sec gens in a few minutes that are complete garbage.

File: Unload Encoder.png (54 KB, 537x359)

54 KB PNG

Remember that you can unload the text encoder after it's been used.
It's just eating up a ton of memory for no reason hanging out in there all the way through.
No idea if this is of any use, but it's a load off your system with no downsides. Might benefit some smaller cards and ramlets.
This "H3 Free Text Encoder" node from the H3-multishot gets the job done.

>>109617079
why would i manually unload it when comfy would automatically unload it when the sampler model needs memory?

https://github.com/BMB12d3/minimax-h3-prompt-composer/tree/main

neat tool, not necessary but can be useful for the more complex video input prompts and so on

>>109616993
>prompt enhancer. Just fuck my shit up

>>109617018
Every time I try to make a longer, high-effort clip, I'm just disappointed by the quality compared to my low effort vibeslop t2va one shots. The referencing overhead murders the quality and it never flows together 100% right, that and it's a vastly slower, less forgiving process. The effort to reward ratio is so low in comparison.

>>109617068
i don't think it has its own genre. i asked an llm to help me describe it
>martial, tonal orchestral piece
>strong duple meter
>dotted rhythms
>snare-drum ostinato
>brass-led melodies in a minor key with singable intervals
>supported by low brass
>bass drum and crash cymbol
>woodwinds provide countermelodies
i will say that the anon who did a simpler prompt for the minimax music model got the right composition style out of the box

>>109617083

What if something else needs memory?
Will comfy be smart enough to unload it automatically then or will it prioritize itself and not give a shit?
Whatever the case, with that node my system went from using up 60gb of ram to hanging out at 44gb during generation and I find it really hard to think of anything negative to say about that.

wow, ltx 2.5 makes the ufo kinos look really cool

>>109617111
That would be me anon, the issue with Minimax is that the model is slopped and there's no real way to input audio or tune it. It gives outputs that sound good, but you're forced to use the prompting template they give you and there's little control over what it gives you. If you prompt it for coherent EDM, the model gives an output that may be good initially, but then it drifts. Unironically, as Minimax is right now, you get more coherent outputs out of SA3 (if you know how to prompt SA3). Of course, the only thing it can't do is vocals, but leave it to Minimax dev team to leave out the encoder needed to add vocals to instrumentals.

>>109617139
Also, notice a ton of people say Suno v4.X or v5.x is better than Minimax, and that's partly true due to how awful the model is at listening. With SA3 you can at least get instrumentals better than every model, outputs that are not just on par but better than all Suno models.

>>109617139
does SA3 take less resources than minimax? i have been too lazy to download any of the music models so far but i think i can try doing rapid prompt testing for the model that is faster so i can see if i can wrangle a soviet composition out of it

>>109617163
Fuck off yellow fever cuck

>another slow thread
this is what happens when you let ani the psycho bake

>>109617177
why do you discord troons pretend like you need this thread? go cry on your discord with all your banned discord friends

>>109617171
go find something else to do

>>109617185
The rentry threads are coming back, ani.

>>109617186
That's rich coming from the braindamaged autist who spends all day spamming yellow fever gens.

>>109617191
QQ moar

>>109617167
Significantly less, on a 3090 it's only a few secs per output whereas on Minimax it's several minutes (somewhere around 10x the waiting time). I'm using Flash Attention 2 and the Base model (50 steps).

>>109617163
at least make her cross her legs the other way

>>109617187
go back to your discord fag brigade and cry about your how your namefag spectre is haunting you. they're the only ones that will coddle your pathetic weakness. but please, keep crying here too. we all love to watch your intellectual poverty on parade

File: MiniMax_H3__00259.mp4 (2.25 MB, 576x896)

2.25 MB MP4

do a 360 and handcuff you

>>109617203
Ani, the rentry links are coming back.

>>109617208
parade

>>109616373
Just another open issue for the pile

What was Debo endgame ? Why he want to create /ldg/ for us FOR FREE ????

Go on debo slut, make us more threads You are our slave

>>109617200
cool, i'll go with that one then. is there a specific prompt format i need to use with sa3? or can i get away with natural descriptions

>>109617024
I mean, this is a ripoff, but that's a poorfag cope and it's only gonna get more expensive from here. If you can afford it, go for it.

Hi, I would just like you remind you all that I am still using ComfyUI, and I will never use AniStudio.
ComfyUI is tried and trusted software with plenty of extensibility. It's not perfect but it gets the job done.
AniStudio is untrustworthy shovelware, insecure, and could very easily carry a virus. This is on top of being very poorly coded, and just a bad wrapper in general.

I suggest you all follow my lead, at least in terms of never bothering with AniStudio. It's just terrible, and there will always be better options.

>>109617070
I tried add guide for minimax h3 and it does extend video, but then I get 22 overlapped frames when I use Image Batch nodes to combine to videos

>>109617229
This. Everyone here already knows ani's wrapper is god awful though, and is largely perceived to be a joke.

>>>/sci/17039733
>>>/sci/17039733
>>>/sci/17039733

>>109617163
it's creepy that her reflection doesn't move

Mental illness

>>109617239
What makes it bad? I think it's the only thing that looks unique among the vibeslop we've been seeing

>>109617217
No format that I'm aware of. I just initially go for general captions that look like the examples on their Github, plus get a bit more specific when the output is missing something. I think all my captions include the BPM at the end though. My captions are one line long at most.

>>109617252
how much time left on your ban? 2d 13hr~?

File: 1775299233748925.jpg (98 KB, 1003x1080)

98 KB JPG

>>109614991
>Hes being doing this for 2 years
Jesus fuck. Not even me have this dedication when shitting on /sp/

Anyway. Im out.
Maybe we see each other in Discord AI gen thread. This thread isnt fun anymore. This retard got too much VPN to waste and janny are too retarded to recognize him

>>109617262
okay, that's good. not a fan of the rigid formatting that some models need since it causes you to need to be very precise with everything which makes there be less variation and surprises between seeds

why is lil bro blog posting like he's someone important like ani

>>109617269
are we supposed to know who you are?

>literal who is quitting
holy shit im going to fucking kill myself

>>109617195
>QQ moar
Yellow fever anon is actually just Asian themselves, neat

File: MiniMax_H3__00260.mp4 (2.33 MB, 576x896)

2.33 MB MP4

>>109617191

>>109617304
himself* ESL bro

Why haven't you made money from this, anon?

>>109617311
because i have a job?

File: 2026-08-22-010933_509x327(...).png (37 KB, 509x327)

37 KB PNG

KJGOD already made the perfect node to make clip extending transitions seamless!!!!

makes the video very slightly washed out but completely flattens out the transitions

>>109617311
because i'm not very good at it. the average random gallery post for loras i look at on civitai mogs my shit

File: AnimateDiff_Quiet.mp4 (3.48 MB, 640x1152)

3.48 MB MP4

im genning

>>109617322
damnit i wish i thought of that first

>>109617322
>0:09
skipped

>>109617235
>I get 22 overlapped frames
Yes that's the point. you're supposed to remove them after.

File: Screenshot 2026-08-22 011504.png (132 KB, 788x889)

132 KB PNG

ultra uncensored heretic gemma-4 26B-A4B Q8 with web search enabled is pretty decent for image captioning. still requires some cross reviewing tho.

>>109617321
:/

>>109617344
the transition is really jarring, I think you can prompt to avoid that

>>109617353
>with web search
What is this for in this use case?

>>109617269
You're literally a discord transgender thread shitter and unironically proud about it. The only person that was in the wrong places for 2 years has been you and you were too retarded to notice this and you made it our problem. Wow actually kill yourself

>>109617308
there is no such thing as an East Asian male because they universally act just like women when raped and this phenomenon has been independently documented historically in both Mongol and Russian literature.

>>109617322
death to SFM "artists"

>he isn't generating at 60fps
are you even trying?

>>109617371
you've never even generated something worth sharing to ldg, why are you talking?

>>109617358
Why does it have the transition?

>>109617311
>Why haven't you made money from this, anon?
I'm gonna answer the spirit of this question instead of the actual question.

With the release of H3 and significant time with Seedance 2.5 I am regretting normiemaxxing and getting an attractive partner. I might have actually made a mistake losing my virginity in highschool and not just personal fetishmaxxing and building a community of like minded people because I could absolutely make $1000 a month without trying generating cuties I would generate for free anyways and sharing it with like-minded people
Maybe I'm just bored

>>109617376
who am i?

>>109617353
>MoE model
Are these usable for H3 prompt writing? I used to use Gemma 26B Heretic but I think I get better results from Qwen3.8 Heretic, even if it sometimes tries to ask follow up questions.

>>109617340
Alright, i ask clanker to make me a node..
this is overcomplicated for no reason

File: file.png (163 KB, 1275x759)

163 KB PNG

Why isn't the bypass taking effect? No matter the strength

>>109617383
ref2v where the reference is not corresponding to the initial prompt, something like that

>>109617385
>who am i?
the whole point is no one know who the fuck you are. the second I say "I have no fucking idea", you lose and I win

>>109617424
therefore, i win

>>109617408
stop making nodes for things you can already do.....

File: Screenshot 2026-08-22 013651.png (466 KB, 1486x1765)

466 KB PNG

>>109617407
I'm using this for image captioning to make a krea2 lora. Not sure how this would fair well for video prompting text for h3 format. The latest version should be compatible with latest qwen model.
this is what i use https://github.com/lostruins/koboldcpp

>>109617163
Stroking her knee just emphasizes the third leg

what to prompt... what to prompt...

How to use music as inspiration for the model without it being copied verbatim?
I use "weak_reference" but it doesn't seem to do anything outside of remixing the music.

>previous thread still isn't archived

File: 6262617171.jpg (1.9 MB, 2332x1556)

1.9 MB JPG

How do you anons upscale images with krea? My workflow is upscale latent -> turbo with 8 steps at 0.4 denoise, but this tends to fuck up the image in terms of style and some details vs the non upscaled gen, like in here. I could theory gen natively at like 5mp which doesn't suffer from this problem, but anatomy and details take a bit of a hit and it is more rng from what I have seen

When ready

>>109617657
>>109617657
>>109617657

>>109617655
>upscale latent -> turbo with 8 steps at 0.4 denoise
latent upscale requires higher denoise like 0.6 or 0.7 and cnets help rein it in

>>109617655
>picrel
Is that a lora anon?

>>109617667
I'm using the https://github.com/LoganBooker/SesquiLSR
>0.6 or 0.7
but this changes the image too much...
>>109617675
Like 5 loras on top of eachother

>>109617340
idk why the motion of the extended video is so slow and shit
I use same prompt when I generate the single video

>>109617687
>Like 5 loras on top of eachother
It's quite nice, good job.