how much better are your gens with fp16 over fp8 with wan 2.2?

>>106523197
>niggerjak in the op

15 minot

Comfy should be dragged out on the street and shot

File: 1167752492635495466-ForgeUI.png (1.76 MB, 1080x1344)

1.76 MB PNG

Blessed bread of bountiful frenship

>>106523246
It begins

File: image - 2025-09-05T185214.830.png (2.55 MB, 1072x1880)

2.55 MB PNG

is it time for Fall or Autumn gens yet?

I have a dream that one day anon will learn what "sloppa" means and will stop posting it

File: 00111-1111970897.png (1.58 MB, 1576x1080)

1.58 MB PNG

go play with your gay toys Andy, im gonna fuck your mom

I have a 5090 and 6000 pro running almost 24/7. I'm looking into solar with battery storage to offset energy cost...

Thankfully power is cheap here, like 10 cents a kWh

>>106523270
>nogen

>>106523270
I see no image attached to your post. Strange

>>106523270
that'll only happen once local catches up with saas models

>sloppers instantly seethe

one day we won't be using python slop UIs

>>106523233
>seething over a guy that takes months off because you're still not liked by everyone else
I was just the first to call you and your loser band of mental patients out. Also I notice that none of you made progress in anything
L
O
L
Fucking waste of space

Shoot comfy in drag

File: ComfyUI_temp_gvlez_00023_.png (2.35 MB, 1280x1280)

2.35 MB PNG

https://github.com/EnragedAntelope/Flux-ChromaLoraConversion
Anyone using this? I keep getting "The requested operation cannot be performed on a file with a user-mapped section open"

>>106523232
only noticeable thing I've experienced is that when generating a 10+ sec 720p video, it had little to no motion with the fp8 scaled model (even with using motion loras), whereas the fp16 model had proper motion even at such a length. if all you're doing is generating 5 second videos then just stick to fp8 as the differences at that point are minimal.

>>106523346
what about Q8 gguf?

>>106523233
>he still thinks the collage is handpicked by an actual person
Anon, I…..

>>106523357
i didn't try the gguf versions, sorry.

>>106523339
No. The fact that [name of repo I forgot] was never really able to translate 1.5 LoRAs to XL (or was it XL to Flux?) kinda soured me on the whole idea.

>>106523367
...am gay

File: Qwan_00006_.jpg (772 KB, 1984x2976)

772 KB JPG

>>106523346
Unless you're offloading substantial parts of the model, at that point there might even be a slight speed boost when using fp16 over fp8 (or only a very minimal slowdown).

File: image (35).png (2.98 MB, 1536x2304)

2.98 MB PNG

>>106523270
I absolutely know what slop is but I’m gunna keep posting because I like slop and I have never pretended otherwise. I also enjoy McDonald’s and kfc and Taco Bell and cheap beer and popular video games. Sorry it upsets you kiddo

chroma users are waiting 15 minutes for 1 gen?

>>106523409
no, just retards

you cant handle my kinosoul

>>106523283
why does local always revert to the same sd1.5 sloppaface greaseplastic style? sdxl, flux, qwen. it always arrives there in the end

>>106523265
I'm not sure which part of this gen is supposed to be Fall or Autumn

File: AniStudio-01450.png (1.2 MB, 1024x1024)

1.2 MB PNG

huggingface? more like "let me interrupt your 18GB transfer" face. amirite?

>>106523283
It's rather see anons local slop than a single piss filtered gpt image

>>106523460
bisgusting

>>106523457
it isn't, I was just asking the thread

>>106523460
progress report?

>>106523485
almost caught up to jan2023 a1111

>>106523463
okay grandpa keep living in the past

booo lame

>>106523460
Suddenly the seething makes sense, sadly I have nothing to do with you shitting the bed for months so now you're on your own.

>>106523515
>sadly I have nothing to do
yeah those 15 minute gens probably leave you with a lot of free time

>>106523531
At least I can gen and not seethe with nothing, How many years are you going to pick retarded argument vectors, lose and then cry for months after the fact?

drag and shot comfy

>>106523535
you've been shiddn and fartn 16 hours a day for a long time ron

Blessed thread of frenship

>>106523476
Sbux is selling pumpkin spice since weeks already, it’s a go for comfy autumn vibes.

>>106523460
We have Comfy
We have NeoForge
We have WAN2GP
We have invoke

Why we need you?

>>106523558
faggot

>>106523257
Retroslop KYS
>>106523265
2.5 doll face slop KYS

File: hotreloadsalloveryourface.webm (1.52 MB, 1920x1080)

1.52 MB WEBM

>>106523485
>>106523567

File: image - 2025-09-05T171320.564.png (2.76 MB, 1072x1880)

2.76 MB PNG

>>106523565
maybe I'll do walking home after school, autumn leaves in the background

>>106523586
krita diffusion mogs this futa-pilled nonsense

>catjak in shambles

Someone needs to post that open comfy/webui IP list again, I’m at work and wanna direct off some gens from my phone

>>106523586
Nobody asked for your garbage UI when ComfyUI and NeoForge already exist and work fine.
Why learn another interface that brings nothing new to the table except letting you attention whore with your name?
You're just noise at this point .

KILL YOUR SELF ANI, YOU ARE WORTHLESS, YOU ARE NOTHING

>>106523620
hint: system shock

File: 00432-1465474720.jpg (1.08 MB, 2048x2688)

1.08 MB JPG

>>106523586
>>106523612
This is getting sad, it's bad enough you shill this garbage but it's even worse when you try to fuck with the thread and go nuclear because anons are tired of you pretending to be anon and astroturfing this alpha software.
Go ahead ani samefag again as anon and get laughed out the thread

>>106523620
https://www.shodan.io/
just search common ports like 8001 or some shit

File: ldg_rockz.png (1.33 MB, 1184x880)

1.33 MB PNG

>>106523197
How much better is Qwen image compared to Flux Kontext and is it better with in the big tiddy waifu department?

>>106523586
We don't need you. You're delusional, and I'm not obligated to use your UI. You've been with this lie of a stagnant project for years, everyone knows it.

>>106523586
wwwwwWWWRRRRRRRRRRAGGGGGGGGGHHHHHHHHHH!!!!!!!!!!!!!!

>>106523586
based

>>106523664
I was aware of how bad it was but ani really will look at post like this and think it's one person doing it. I thought he would get better with me gone but holy shit he went crazy.
If you heard all the lies he made about his UI including having a fucking college use it when it's in this fucking state what 4-6 months after being dropped?
This dude ani is a full on schizophrenic and should never be listened to or trusted

>>106523586
You contribute nothing while larping as a dev, kys together with the anime sloppers faggot, nobody wants you.

>>106523567
havign more options is good

>>106523623
well it does do something new. I don't use pytorch for one. almost all the required libraries are included in the binary, the licence permits Vidya game development, shared libraries can be added during runtime and developers can see their changes immediately instead of restarting the whole app. ImGui is heavily customizable over the rest so the UI can look and operate the way you want. you can change filepaths without having to restart or use a ridiculous relative path as the filename. lastly, it's much easier to make custom tooling and uis for getting stuff done comfortably

>>106523630
both of you can fuck off

wan is fucking confusing, and all of the jeets pushing the borked 4-step distiller doesn't help

>>106523698
reminds me of when he claimed studio mappa was going to be using his shitty animatediff.

>>106523707
Or how about you fuck off to the thread you originally backed?
If find it adorable you think there's a seat for you here when ever old poster is well aware of your history and actions. It must burn not getting any support on your project either, but who wants to work with a lolcow like you.
Should I list off more of the tall tales you made before releasing this abortion of a UI?

>nodev glorifies poothon devs over C chads
I thought this was /g/? what gives?

File: 00432-1465474720f.jpg (2.35 MB, 2048x2688)

2.35 MB JPG

>>106523630

How good is qwen for try on clothes?
do you need lora?
I remembered base kontext could do it but it was very bad

>>106523736
much improved!!!!

>>106523735
>glorifies
Nah, I just understand why it's widely used in this field. Noticing and knowing isn't the same as glorification, fag.

drag comfy and shot comfy

Ani cobbled together his UI using AI, he has no coding chops and has very little actual dev work and it shows.

>>106523586
Nobody asked for this, your UI is trash.

>>106523744
but python is just a layer of shit over C libraries? why would you appeal to popularity like that? are you actually retarded? the reason researchers use python is because they can't be assed to learn C.

>>106523756
>Ani cobbled together his UI using AI
isn't that in the spirit of the thread

>>106523738
Pretty ok if you have the individual pieces of clothing on something like a white background. If you say something like "make me wear what this person is wearing", or swap clothes those don't really work reliably.

>>106523756
I can't imagine how infuriating that would be when llms suck ass at c++

>>106523735
>I thought this was /g/? what gives?

 pip install -r requirements 

filters most anons anyway so its not like they are able to fathom why its widely used like this anon >>106523760

>>106523759
Anons would have been more accepted if he was humble about how he would spread his UI instead of trying to force it in every general
>>106523763
Not when the schizo lies for months pretending that it would be some revolution in the space and it's DOA and the dev can't even get support because he burned every bridge around himself by being a schizo loser
>>106523776
Ani lacks the ability to think things through, if he had that skill he wouldn't have got in half the problems he put himself through

>>106523779
it's widely used because the world is 99% nodevs that need the hello world lang to do anything

>thread is fucked by some namefaggot
>back to UIwars
Alright I'll just go back to learning how to bake my wan porn lora......

>>106523638
Is there another one? That only lets you see like ten results before asking you to login. Some anon shared something days ago and we were trolling exposed comfy instances, shit I wish I had saved that post.

>>106523795
>the world is 99% nodevs that need the hello world lang to do anything
correct but im confident one day you will learn the actual reason

>>106523812
he had an account. he was using shodan. just use a burner

>>106523814
enlighten us anon, why is it widely used?

>>106523812
That was literally it. Just make burner gmail account

>>106523586
that's great anon but no offense, you really need to work on that gui as most people aren't weebs stuck in the 90s.

>>106523821
>enlighten us anon, why is it widely used?
He posts that same cryptic bs message every time the discussion goes to how python is shit but he never follows up. I think he’s just trolling at this point.

File: 3084736420.png (1.56 MB, 1824x1248)

1.56 MB PNG

Get it twisted https://voca.ro/12Ib4fVGiSiR

>>106523833
if I had more hand in the project we could make the most kino frontends but alas I still need to work on logistics. the big things being splitting out parts of the app to shared library addons that get packed in the release so you don't have to use memory to register shit you don't need and the node execution. I can make it "modern" looking but I'd rather focus on the functionality first. if you are wondering what's possible, check out everything else that uses this gui lib
https://github.com/ocornut/imgui/wiki/Software-using-dear-imgui

>>106523853
what is this final fantasy villain monologue shit

>>106523870
>Harass devs and model makers
>expose self doing industry poison
>annoy anons by samefagging and trying to inject your garbage in the OP
>surprised nobody wants to help you
You really should look in a mirror or take meds bro

>>106523853
jesus this bitch is a fucking BUMMER to be around she needs to get sodomized a few times

great job by the way

File: file.png (75 KB, 742x621)

75 KB PNG

wait until he learns about chroma

>>106523885
do you have any evidence of any of this happening or are you sticking with this schizo narrative?

>>106523902
We did this episode too many times the project can speak for itself, we're at a point all of his outcomes are speaking for themselves.

File: ComfyUI_01426_.png (1.21 MB, 768x1280)

1.21 MB PNG

>>106523821
>>106523822
Thanks lads just used an old gmail addy for it and found some in china already lol. I just wanted to do a quick gen for >>106523565 this guy but this one only has some weird pony mix model kek. Still, it’s pretty fast.

>>106523936
that he is making progress? I think that's a good thing. maybe you should learn something from ani and make something worthwhile instead of shidding and farting all the time

If nothing else, the fact that he isn’t using PyTorch if it’s equally performant should interest everyone. PyTorch is ass.

File: 1754496269451805.mp4 (2.46 MB, 704x1024)

2.46 MB MP4

>>106523498
damn this vramlet is zesty

>>106523962
>if it’s equally performant
BWAHAHAHAHAH

>>106523951
>Learn
>How to make a dead UI despite having a resource of devs that work in the space that won't support you because you shat the bed publicly too many times so your only public stan is a known thread squatting schizo that has so little reach putting your shitty program in his OP does nothing
>>106523968
He always speaks in the third person it's so cringe right?

File: 81359230-348940020.jpg (52 KB, 600x469)

52 KB JPG

so is Panchovix developing reforge again, couple weeks ago the repo was archived and he was to odepressed to continue developing or some shit but now it's open again and there's new entries

>>106523409
??? are you a moron ?

>>106523997
cant you just contact him directly to get an answer?

>>106523968
Yeah brother, I’m not some schizoid comfy only truther, if it works it works. I don’t care how the sausage is made as long as I can get it pain free and delicious.

>>106524008
no i don't use chroma

>>106523962
>he isn’t using PyTorch if it’s equally performant
Pytorch is pretty dang performant. Meta does have engineers that know how to implement low level code supporting high level languages quite well, Facebook would've kneeled over and died on its PHP codebase super early if that wasn't the case. And outside of most hyperscalers, most people still use Pytorch. What Meta doesn't have is scaling where they lack engineers with HPC experience. There were rumors about their training code being so shit on Pytorch that 10k GPUs running training was the same speed as their full 50k cluster. But that isn't relevant for a simple application doing image diffusion.

>>106524019
Neither am I the thing is it's not as performant. It's actually much worse.

>>106524019
If it works why is the only faggot using this huck of shit the dev himself?
Oh that's right it can't do most basic features still

>>106524016
no

>>106523997
>couple weeks ago the repo was archived and he was to odepressed to continue
i earnestly believe he came here, anon posted about it, and was assblasted so hard he gave up
cool to see hes back at it though

>julien speaks about himself again in third person
Actual schizo

>>106524031
>meta
reminds me of time someone was donated a working AIT fork and they made it worse, promised a rewrite before eventually abandoning it.

>>106524019
>>106524033 (me)
That's if it doesn't crash, mind you. Which happens frequently.

>>106524035
I literally don’t care anything about the guy making it bruv, any more than I cares about auto or comfy, you need to touch some grass lmao damn.

>>106523232
quite a bit better just look up some comparisons on youtube or reddit

>>106524033
Yeah I’ve never used it didn’t mean to imply I had I’m just saying if it ends up being a decent alternative then that’s fine, don’t know why this one schizophrenic seems so assblasted about it lol

File: ComfyUI_00004_.png (1.38 MB, 784x1216)

1.38 MB PNG

>>106523943
Another Chinese open comfy, this one only has a model called “dream shaper”. Seems like some kind of sdxl/illus based thing. These guys must be running on hpcs or something they gen so fast lol

File: 45747445.png (243 KB, 785x1000)

243 KB PNG

>muh chroma
>muh comfy
>muh anistudio
can someone explain these retard ass model and UI turf wars to me, I'm not autistic enough to understand, more options seem a good thing to me

>>106524085
>>106524100
>Ask same question that is always asked when he's under fire
>not getting the answer that you want
Is part of being schizophrenic is holding the belief you're smarter than everyone around you despite being proven on multiple occasions you are not?
You pulled this shit not even a few weeks ago when you tried changing the OP, I was alerted to the thread and still see the same low IQ canned post that never help

>>106524117
You can hover your mouse over the gpu usage graph to see their gpu.

What's the magic keyword to get a narrow jaw on Qwen NPCs?

>>106524124
theyre meaningless in our quest for the booba

>>106524124
The seething about comfy is from ani's pet schizo that is shilling his shitty UI, the other samefagging and advertisement is from ani himself. He keeps dog whistling things to get me angry but it doesn't work when everyone already knows his MO and patterns. He has made zero progress and can't even blame anyone but himself on how he and his UI are viewed

>>106524132
Ah I’m hacking the matrix on my phone here lol I don’t think the phone comfyui shows that stuff

>>106524117
>this one only has a model called “dream shaper”. Seems like some kind of sdxl/illus based thing. These guys must be running on hpcs or something they gen so fast lol
is this a troll or are there really anons out there who missed the 1.5 and xl era kek

File: AniStudio_output-00002.png (977 KB, 1024x1024)

977 KB PNG

>>106523962
it's getting there. or maybe it is faster? I wouldn't know, I got rid of all the pytorch shit off my machine but I remembered it was like ~7 seconds for 1024x1024 sdxl on a 4090. I'm not using lightning or anything

[INFO]: stable-diffusion.cpp:1523 - sampling completed, taking 4.86s
[INFO]: stable-diffusion.cpp:1531 - generating 1 latent images completed, taking 4.92s
[INFO]: stable-diffusion.cpp:1534 - decoding 1 latents
[DEBUG]: ggml_extend.hpp:1148 - vae compute buffer size: 6656.00 MB(VRAM)
[DEBUG]: stable-diffusion.cpp:1127 - computing vae [mode: DECODE] graph completed, taking 0.66s
[INFO]: stable-diffusion.cpp:1544 - latent 1 decoded, taking 0.67s
[INFO]: stable-diffusion.cpp:1548 - decode_first_stage completed, taking 0.67s
[INFO]: stable-diffusion.cpp:1682 - txt2img completed in 5.67s

>>106524147
Don't you think you should first apologize for all your faggotry and thread splitting to force your UI in OP before trying to pretend you're a regular?

>>106524117
>Seems like some kind of sdxl/illus based thing.
SDXL vae or something better, vae is the weakest point of everything XL based desu

I honestly thought about contributing and fixing this mess that is AniStudio but after seeing how he treats actual contributers who tried to help like hlky... nyo

>>106524147
cute!

>>106524147
You belong to the non tech pedos julien >>>/g/adt

>>106523896
lel, I make this extreme claim, I won't back it up with anything, just trust me bro

that there could be cp in datasets of millions or billions of images, yes, it has already happened with huge datasets like LAION and likely still does, but it would be so few that they would be drowned out during training

that the model trainers would create nonsense tags to train deliberately cp into model, that's insane, why risk jail time for something they could just solve with a bit of further finetuning post model release

>>106524162
Glad you're still here hlky
post some more dall3 gens maybe?

LDG is the forefront of local image generation

>>106523257
nice one

>>106524028
Yet you can't stop obsessing and lying about it, do something about your mental illness

File: 00602-1654927422.jpg (873 KB, 2048x2688)

873 KB JPG

Getting rejected by comfy broke him so now he's posting here as a signal of defeat.
You have your containment thread and the anime thread but decide to post here because?

>>106524176
Not hlky retard
The point is there's a reason no one wants to work with you and comfy distanced himself from you

>>106523233
and julien will be in the next

File: AniStudio_output-00005.png (1.05 MB, 1024x1024)

1.05 MB PNG

>>106524167
gao! <3

>>106524162
I'd honestly just wait until I wipe my ass metaphorically speaking. you can help if you want but I'd rather things get organized, linked properly and a dev guide is written. the hot-reloads will really help later

>>106524188
why do you post in the thread if all you do is get mad? take a break dude

>>106524127
>Is part of being schizophrenic is holding the belief you're smarter than everyone around you
no that just comes with being on 4chan

>>106524124
>muh chroma
mentally ill person obsessed with hating Chroma
>muh comfy
mentally ill person obsessed with hating Comfy
>muh anistudio
mentally ill person obsessed with hating Ani

They could very well be the same person

>>106524201
Your first post was seething at me followed by trying to shill the same program you tried to split the thread with not even two weeks ago.
Now you're doing that schizo thing when you claim others are angry despite you being the main force for that to happen. You're a fucking clown and it's a good thing they kept your crazy ass doing grunt work in the military for everyone else's safety

Well the one guy posted a super highres image with two girls while the other guy posted a lowres image with only a single girl so I think the first guy wins

>>106524187
huh?

File: 00083-672730187.png (1.51 MB, 896x1152)

1.51 MB PNG

>>106524207
personally I hate comfy, like chroma and don't care about anistudio

>>106524205
>Underrated post

finally glad there is another hat in the UI race going into next year. it's actually doing something outside of python which is a major plus

>>106524284
Cuckoo's nest

Reminder:
https://files.catbox.moe/g50shl.flac

$150,000, 105,000 H100 GPU hours, but not a single artist tag? Something just aint adding up here!

what's this schizo obsession with artists

>>106524348
I think he just didn't include them like the pony guy but didn't want to admit it publicly seeing the reception it had on pony.

Blonde fennec > silver cat slut
Silvers are always sluts.

>>106524372
>fennec girl plays in shit and mud
>silver cat slut has sex with you
yeah, you have shit taste

>>106524385
>silver straight up have sex with anon
Yeesh, what a fucking slut.
We need another schizo whose mascot is pink. I'd marry her.

File: Chroma_00008_.jpg (626 KB, 1160x1408)

626 KB JPG

>>106524304
Tweety bird!

First time trying to use any image 2 video model, downloaded WAN 2.2 and trying in Comfy under windows on my RX 9070 XT using latest ROCm. It works but my card OOMs with the 480p resolution, 512x512 gen finishes but it's slow to load between steps, fast when it's actually doing the diffusion or whatever. Meh.

>>106524411
bronze is next and it would probably be a dog. get the peanut butter ready

File: 3565050.jpg (93 KB, 1080x1037)

93 KB JPG

>>106524417
>RX 9070 XT using latest ROCm

>>106524348
Yes there are artist tags, but classic artists like like Alphonse Mucha etc

Just like all with other base models after SD1.5, since they don't want to get sued by contemporary artists like Midjourney, Stability Inc, Deviant Art etc have been

Probably a demand from whoever funded the training, since they don't want to be sued

File: Chroma_00011_.jpg (754 KB, 1576x1080)

754 KB JPG

>>106524435
>Yes there are artist tags, but classic artists like like Alphonse Mucha etc
some from anime booru also work

>>106524473
which ones?

>>106524417
you need flash attention
https://github.com/ROCm/flash-attention?tab=readme-ov-file#getting-started

>>106524475
Didn't make a list. I just noticed it when running old Illustrious prompts with and without them like month back

>>106524435
That line of thinking may have true before but now that models like illustrious/animagine is out in the wild with all of that fully intact and nothing has happened to them and pretty much all saas now fully on board too. It's just a weird way to shoot yourself in the foot now. Ahh well

>>106524480
Well I installed "sageattention and "triton-windows" - ComfyUI loads with --use-sage-attention now and seems to do something and is generating now...
>Error running sage attention: [WinError 2] The system cannot find the file specified, using pytorch attention instead.
fuck

File: ComfyUI-vid-_00079_.mp4 (1.26 MB, 720x1280)

1.26 MB MP4

strange things happen when I increase WAN i2v resolution

>>106524473
He trained on artist tags, but he also did a random token soup mask which means the model was never properly exposed to those tokens.

>>106524514
>illustrious/animagine
Those are finetunes specifically for artstyles, so yes, they need to tag the artstyles else they might as well not exist, are you retarded ?

>and pretty much all saas
Google, Microsoft, Meta, they can afford to be sued and fight it in court for decades, community funded models can't afford this

Note that the artists currently suing are not going for Google, Microsoft, Meta, they are going for smaller SAAS like Midjourney, Stability Inc, Deviant Art, because those companies don't have the money to spend decades in court with the best legal defence money can buy

>>106524577
>a random token soup mask which means the model was never properly exposed to those tokens.
Is there some benefit for doing this?

>>106524598
yeah you're a lazy faggot who doesn't want to do multiple captions so you cache a giant shitball and then put masking tape over random parts

>>106524577
Source ? Because again, classic artist tags like Alphonse Mucha work, so he must have selecting certain artist tags then

Is there any edit model that isn't slopped? It seems like even Nano Banana is slightly slopped. No matter what, it can not produce amateur looking images (similar to what you can get out of Chroma).

>>106524578
My brother, nobody is finetuning this shit.

>>106524367
>I think he just didn't include them like the pony guy but didn't want to admit it publicly seeing the reception it had on pony.
as if we wouldn't notice such a thing, this guy is such a snake

File: Screenshot_20250908_143737.png (413 KB, 683x816)

413 KB PNG

I'm going to try to make chroma loras, I have the hardware but will I be the roadblock in fixing this shit

>>106524609
Major artists are in the Flux model but improperly trained. So they're 1) more likely to be properly captioned in the soup mix and 2) they're easier to realign the existing image knowledge to the text tokens. I don't know why you people act like this isn't what happened. We know he had booru images with artist tags that he trained on. So yes, what I'm saying might be a hypothesis, but it's also very probable for what happened and we see it in the end result.

>>106523661
It's no contest, like Wan and Qwen Image, Qwen Image Edit is very easy to make uncensored, full nude LoRAs for and it has zero refusals.

>>106524578
>Those are finetunes specifically for artstyle
The flying fuck does that mean? It's trained on anime exclusively using danbooru tags which includes artist tags. That's like saying doing a realism tune that include captions of different camera used a camera style tune lol. Your teachers must have always returned your tests face down huh?

>>106524629
Major artists are in all the large model training sets, be they local or saas

Post SD1.5, they are increasingly seldom tagged with the artist name, and contemporary artists are more or less never tagged, for obvious potential legal reasons

It's not hard to write a script to remove booru artist tags and just train on the rest

Don't know WTF you are babbling about with your 'soup mix', it makes no sense, just remove the artist name part and train on the rest of the tags/caption

>>106524704
The evaluation prompts which he had live during training were a soup mix capiion.

LONG CAPTION + booru auto tags + artist tags, meta

No need to argue any more.

>>106524663
>It's trained on anime exclusively using danbooru tags which includes artist tags
Yes, so that you can SPECIFY SPECIFIC artist styles, which is the whole reason for finetunes 100% focused on artist styles

Like how fucking stupid are you ?

>>106524726
Fuck you are a fucking moron, well I don't want to lose anymore IQ with this shit, keep on knotting fur bro

>>106524724
Might be a dumb question, but did he use commas or dots as separators with booru tags?

niggle me this:
if chroma has 4x the parameters of XL, 4x the VAE resolution of SDXL, takes 4x longer to gen than SDXL, why doesn't it have 4x the knowledge of SDXL? It should have room to fit all booru characters and styles neatly in a corner while still having 3/4 the room left for realism, furry, and everything else. Why did chroma learn less in 50 epochs on H100s than what KohakuXL (the base for illustrious) learned in 2 epochs on a shitrig of 3090s?

>>106524563
there's no sage attention for rocm and
>windows
prepare for pain

>>106524724
You can easily remove artist names from tags/caption no matter what the caption strategy is

All contemporary artists have their art being trained in all models, local or saas, these days they are practically never trained with the artist name in the caption/tags

For legal reasons

Finetunes specifically aimed at artist styles will train on artist names, since that's their sole reason for existing to begin with, so they take that legal risk

>>106524766
Illustrious gets anally raped by chroma, you don't need tools like regional prompter and you get accurate text out the box. Niggle me this why are we niggling so hard when when we should naggle?

>>106524766
It's not knowledge, it's output coherence. More parameters isn't necessarily a larger dictionary of tokens but rather the output scene making more sense. SDXL is so bad when it comes to scene composition.

>>106524782
then why do chroma outputs look like a garbled mess that somehow does fishnets worse than sdxl?

>>106524810
lobotomizing the model at 512x512 res for 48 epochs does that to a model

>>106524766
>why doesn't it have 4x the knowledge of SDXL?
It does, but it hasn't spent that on booru characters, because it's not a booru character model, it's a base model to be further finetuned.

>>106524810
Not enough steps
This model requires beefy specs one of the biggest issues anons had with flux, but unlike flux this model loves higher steps which used to not be the case before the end

>>106524810
Skill issue, likely you are bad at writing prompts. It's simply a fact that larger models have better scene composition and comparing it to your cropped fetish 1girl anime SDXL model is retarded.

>>106524578
>Google, Microsoft, and Meta have money, they're not targeted
>Midjourney, Stability, DeviantArt don't, they're targeted
So why has nobody gone after NovelAI then?

>>106524833
probably because Turkish

>>106524831
I have no problem prompting pregnant women licking hag asshole in public squares with other women taking pictures in the background with chroma anon.

>>106524810
>somehow does fishnets worse than sdxl?
SDXL had more of its training data be 'fishnets'... it's not rocket science

All base models have biases, as in things it has more of than others, you will typically try to balance concepts, but some will be more important and some may be imbalanced as a mistake

>>106524833
The people suing seem to be western artists, isn't NovelAI like 100% anime ?

Only way to truly answer that is ask the people suing, IIRC they're like some collective of artists, go on, solve the 'mystery' anon

File: file.png (1.52 MB, 1024x1024)

1.52 MB PNG

With that all said, Chroma is shit.

>>106524931
with those settings I can see why your slow ass is struggling

>>106524778
Well it works fine without but it's just slow when switching between models and vae/lora/whatever while it's genning. It's like it's slow to load the things.

>>106524931
the minimum time to get a good gen is 15 minutes maybe you didn't wait long enough

>>106525002
This has been bothering you for a full day my unemployed friend

Who the fuck send Ani to my beloved /adt/? We don need him there. Send Ani to /sdg/

>>106525013
no it's yours now

>>106524766
Because you're retarded. Different architectures present different challenges. Chroma is inherently limited by the Flux model. The Flux architecture itself is not the issue, but the way BFL trained it is. They did not train it in a way that makes it easy or efficient to train artists into it. No one has ever successfully trained a variety of styles into a single Flux model.

So what is the issue? The Flux base is a strong realism model. That's why it's so good to give it additional realism concepts. It's also very good for general art, and typography. Teaching additional concepts in what it excels at will be simple. But what it won't do is a precise artist, because Flux is not an anime model. It also has very little control for styles/art. This issue has plagued Flux since its inception. It's not a Chroma issue, nor is it a finetuning issue. It simply is a very hard challenge to tackle in a giga model. Very few models since SDXL era has given artist/style freedom, SD3.5M/L being the last (but certainly not the best for anime/manga, that goes to the deprecated HunyuanDiT).

>>106525011
>This has been making you laugh for a full day my employed friend
thanks for the laughs my friend

File: file.png (1.98 MB, 1536x1024)

1.98 MB PNG

>>106524988
I don't think Chroma can do this.

>>106525038
Training separate base models, one for realism, another for anime, the way MJ does it (Niji), is certainly the way to go. But that's unfortunately never been done/considered. This one surprising aspect out of Chink models desu. They still are okay with completely slopping their models. Every image that comes out looks like it came from a Donghua. That's why it's pivotal to separate the two concepts. They would no doubt catch up to Flux or would've a long time ago had they done that, but instead we have to deal with the current state. It also does not help that they're lazy and use a bunch of synthetic data to train their models.

>>106525066
yes it can
I hope that isn't your final resolution after a upscale, this garbage belongs in /sdg/

>>106524565
720x1280 is supported, this looks like a cfg issue

in 2 more weeks someone (not me of course) will surely drop another 100k to finetune chroma on 1024x1024 and anime

File: cope.png (336 KB, 583x680)

336 KB PNG

>>106525087
>yes it can
https://youtu.be/6-9-FkwUrRo?t=38

>>106524766
>Why did chroma learn less
Because Lodestone trained it wrong. I tried to tell everyone this. Months ago I pointed to specific places in his training code where he is doing something highly suboptimally or outright wrong, and people just called me an anti-chroma schizo.

If he just trained the model normally using standard techniques, it would have learned more in 10 epochs than it actually did in 50.

>>106525038
it's excellent at buttchins though

>>106524994
Maybe try Wan2GP instead, but be warned, I have your setup (except on Linux) and it pretty consistently takes 1h08m to gen a five second video with Wan2.2. Trying to gen videos on AMD is pain, there's nothing you can do about that.
(Disclaimer though I don't know shit about shit, I'm as new as you)

every furry finetune has been stopgap copium until a superior anime alternative arrived. this was the case with easyfluff, pony, and now chroma.

>>106525161
>it pretty consistently takes 1h08m to gen a five second video with Wan2.
What? Takes 5 minutes for me on windows with latest rocm and comfy but 512x512 resolution. length 81

File: file.png (1.92 MB, 1536x1024)

1.92 MB PNG

>>106525087
Seriously doubt it. But I want you to prove me wrong.

>>106525085
>Training separate base models, one for realism, another for anime, the way MJ does it (Niji), is certainly the way to go.

Note, this should not necessarily mean just training anime. As I said, Flux base is a strong realism model. It was still trained with some anime/art in it. I think the same should apply to an anime model. That way, it's still knowledgeable in other domains, while still being strong a single domain.

>>106525166
At this point I just want qwen but finetuned with porn. Anime models are more or less solved.

File: 2025-09-08_15-27-13-train(...).jpg (51 KB, 1024x1024)

51 KB JPG

Chroma training is going surprisingly well

>>106525085
Wrong, it's been prove the more diverse the dataset the better the model is at generalizing.

>>106525214
very nice style. I'll try it when it's ready

>>106525196
I always thought some knowledge of realism in bigger anime models was useful to make the model understand basic anatomy.

>>106524565
happened to me, its it because the lightx lora?

>>106518830
>>106518688

>>106525172
Good thing I put that disclaimer on lmao
I'm genning with 1024x1024 fwiw but the discrepancy shouldn't be that big. Also according to other posts and my limited experience the animation and image quality at lower resolutions is awful, you get much better results at higher resolutions.

Please, we don't need Ani and his obvious samefagging in /adt/. The thread normally it's slow, but since he started posting we're at 20 replies in less than half an hour. It's clearly artificial posting.

>>106525200
That would be no different from taking Flux and finetuning it with porn. Qwen does not introduce any unique flavors that haven't already been done.

>>106525247
Based on the images posted there the thread was made for "men" like him

>>106525255
FLUX IS DISTILLED!!!!

>>106525135
I have no doubt we'll see finetunes, I would not be surprised if people were already doing tests for finetunes

Having said that, these take time both in preparation and training, as an example:

SDXL base model released August 22, 2022
Pony Diffusion finetune released January 2023
Pony Diffusion v6 XL finetune released January 2024
Illustrious 1.0 finetune released July 2024
Illustrious XL 2.0 finetune released March 2025

Keep seething

>>106525255
I just like qwen more in general compared to flux, it also feels more "intelligent" at least in following prompts, and they are less censored by default.

>>106525275
AYY-LMAO https://www.illustrious-xl.ai/sponsor

Illustrious 3.6 v-pred at the same stardust amount for ~2 months

>>106525147
kek you're so full of shit, you know nothing about training models

please, point out what he trained wrong and these 'specific places in his training code'

what an absolute retard you are

>>106525290
>they are less censored by default
BFL is just another western company with a boner against nsfw so...

>>106525275
No one is gonna train chroma. Might as well pour gasoline on your money and burn it lol

>>106525296
They really dragged that on to a point I fully understand why everyone is telling them to kick rocks

>>106525267
This faggot literally did nothing but wants clout in /adt/ like he contributed something. Meanwhile Panchovix and Homing are also weebs but actually made working, updated forks without attention whoring or samefagging.

File: file.png (1.21 MB, 1024x1024)

1.21 MB PNG

If you were going to spend money finetuning anything it should be Qwen Edit, which can do both txt2img and img2img edits.

>>106525275
>first usable finetune took nearly 1.5 years for a significantly cheaper model to train
lule mao

>>106525315
Him and his pet schizo already forced his UI in OP but it doesn't look like anyone wants to use it.
I don't know what drives a man to wakeup and take an absolute shit on his entire online and personal life on a daily basis. He was well liked at first then some schizo switch flipped and he started attacking people his drug addled mind perceived as threats or showing up up.

>>106525338
The software around training isn't even comparable. Pixart has been trained three times since then in several months each. The fact is no one is spending money or the good models (my theory) are staying private.

>>106525269
It was, but even in spite of this the state of NSFW is equivalent to base Qwen's. Turns out that overbaking your 20B model on slopped/safe pics is also not a good idea keeping an unfiltered model in mind.

>>106525337
Ideal but I don't think there is a easy dataset to use that is readily available for edit training. Could be wrong though.

>>106525358
Edit training works on white source images.

>>106525315
If you want a hint: you need to fight against the avararfags/namefags/attention whores
It will NOT get better by itself

>>106525200
thinking anything in ai is 'solved' is a sign of low standards. anime models are nowhere near close to being good

>>106525373
I can gen pregnant women sucking on giant bug ovipositors that ejaculate eggs and parasites and you're dooming?

>>106525308
It's a much better base model than SDXL, so if you want to do a new full finetune, it's easily the best option since it's already been de-censored and trained on tons of real world data, and training it is FAST compared to Wan and Qwen, which also needs to learn NSFW concepts and be de-slopped

Should stuff like 1girl/1boy, solo be tagged when making a character LORA/textual inversion?
I get the idea of tag only stuff you don't want the AI to learn, but does the AI meaningfully separate these tags from the character itself? Does anyone know?

>>106524994
>slow when switching
I have sageattention2 and loading off an ssd and it’s slow to switch models here as well. I think it can only load them into vram so fast, but I’m not a computer scientist so iunno

>>106525353
I don't even know what even means lol. We are talking about training here you dolt. Tuning on a distilled model just produces mustard gas. Getting the hardware is the big hurdle with tuning qwen no technical issue in doing it.

File: file.png (21 KB, 898x116)

21 KB PNG

>>106523460
>huggingface interrupt your 18GB transfer
That's a you problem

>>106525370
Besides being a complete retard, /adt/ was always a comfy thread where you could recognize the regulars. Now suddenly we have new avatarfags crawling out to defend this moron. This brainlet is definitely lurking ITT. Needs to touch grass.

>fp8
>not q8

File: file.png (1.61 MB, 1024x1024)

1.61 MB PNG

>>106525425
When the Spark comes out people will finetune the big models.

File: 1757307820052921.png (572 KB, 768x789)

572 KB PNG

can someone make a comfyorg edit?

>>106525373
I guess yeah, next step would be natural language understanding and zero shot proper limbs and positions.

>>106525290
>I just like qwen more in general compared to flux, it also feels more "intelligent" at least in following prompts

Qwen is good at following prompts, but it's not 20B good at following prompts. Besides, it also has the limitation that its results are rigid. Chroma is smaller than base Flux, just as intelligent, if not more than Qwen, because it's fully uncensored on top of that. Even if Chroma at is state is not perfect, it proves beyond a doubt that it's possible to get much better prompt following quality out of <10B models.

>>106525444
>regulars
Absolutely disgusting and completely anti imageboard culture. Take it to preddit or trooncord

>>106525444
As someone who originally tried to stop his kind and with the help of anon won because he's blocked here culturally
>He will blame you for everything he is doing
>He will lie over things easy to verify and will keep doing it for years
>He will use debo and his pet janny stells to aid him in his shilling and attacks
>he will randomly insert himself into your thread whenever a event that causes new posters to join, this is the most dangerous and be prepared for him to samefag in a pathetic attempt to gaslight new posters into thinking regulars are insane
>he will name drop others as a means of smokescreening and if he knows you by style will only reference you
>will try to falsely label you as a avatar fag in a pathetic attempt to bring you to his level
You have a life long schizo on your hands

File: file.png (1.29 MB, 1024x1024)

1.29 MB PNG

Using Qwen-Image when Edit exists. The Edit model, like Kontext, are fundamentally superior because like the video models, they have a better understanding of spatial features.

>>106525500
Can you prompt it like it was normal qwen?

File: ComfyUI_temp_axihh_030013_.png (1.93 MB, 1344x768)

1.93 MB PNG

Using differential diffusion, is there some way to force the denoising to go all the way through?

>>106525506
>Candid photograph. Snow White, a young woman with pale skin, dark hair, and red lips, is peacefully asleep on a white padded bed. She is wearing a blue dress with red accents and holds a bouquet of white flowers. The background is a muted gray, highlighting her serene expression and the delicate flowers. The image captures a moment of calm and beauty.

>>106525409
Kind of an indirect answer but fwiw I spent all weekend gachaing my 1girls and illustrious/noob based models definitely seem to “understand” that 1girl means one main subject. I was playing around with 2/3girls and “multiple girls” and it noticeably creates more multiple subject images. Not very well or controllable, lot of hallucination with the extra girls being in the background or other weird things, but certainly significant.

>>106525500
I am waiting for the nunchaku release this week to really get into it.

>>106525500
>The Edit model, like Kontext, are fundamentally superior because like the video models, they have a better understanding of spatial features.
and you can go for any character you like and make scenes out of them, no need to make characters loras anymore

>>106525514
probably fixed resolutions. check the documentation

>>106525525
I'm talking strictly text to image generations, Edit models have to learn spatial feature reasoning in order to do their task. So they're straight better at everything.

>>106525529
What?

>>106525500
Do you need to have premade white square jpegs in various resolutions or does it just eat normal latent?

>>106525538
it might only work with very specific resolutions. please just read the documentation

>>106525547
What documentation do you mean?

>>106525554
for differential diffusion anon....

File: file.png (152 KB, 933x837)

152 KB PNG

>>106525546
I haven't experimented, I just do a white image as the control image and an empty latent for the actual gen.

>>106525563
Can you describe differential diffusion to me, please?

>>106525565
Is this the comfy template wf?

>>106525565
lol wtf? you can use edit models txt2img out of the box. Encoding a white square/rect just adds more to the process

Qwen are we going to get a finetune. Get it?

>>106525516
Interesting

File: file.png (1.25 MB, 1024x1024)

1.25 MB PNG

>>106525586
The white box gives me booba and no box does not. Also objectively inferior.

>>106525593
I WAN one too.

>>106525519
Yeah I know that but I guess I just gotta test how the embedding behavior changes with the 1girl, solo removed from tags myself.

>>106525500
No controlnet support

>>106525607
>controlnet for a universal edit model

>>106525611
Is that supposed to be some sort of dunk? I need depth and pose controlnets for my workflow

>>106525529
>>106525547
>>106525563
Seriously though, why can't some janny get rid of this cunt shitting /lmg/ and now /ldg/ with his Nemo (or whatever) tune?

>>106525629
nigga just type it in

>>106525629
Why would you need a depth lora for a model that fundamentally knows the depth of a scene? Why would you need a pose controlnet when "match the pose" works as text? Have you even used Qwen Edit?

>>106525629
As far as I know you can just use the depth as the reference image and prompt as normal and it should work like a CN. Not sure about pose

File: IMG_2959.jpg (1.02 MB, 1142x1742)

1.02 MB JPG

>>106524117
Still clicking through more unsecured comfy installs and came across this absolute madman. That’s not a mistake homie is literally taking almost a full hour to generate on an sdxl based model. Grim. There are people really out there living like this.

>>106525628
kek

>>106525634
>>106525644
>>106525649
Being able to control the information available to the model is extremely valuable. I have an actual professional workflow and require characters to have a specific pose but need text guided generation in all other ways with nothing in the context window. Same for depth with backgrounds

>>106525650
Share his workflow.
I am curious what he is doing to take so long.

>>106525650
ain't no way this nigga is still at it

>>106525675
Scamming jeets for bitcoin dust with your coombait gens doesn’t make you a professional anon.

>>106525675
Anon you use SDXL, don't even talk. It's like a toddler who uses Duplos talking about K'nex not being good for complex projects.

>>106525628
lmao

>>106525650
you should make anti Chinese propaganda on their machines

>>106525682
I’m retarded I don’t actually use comfy at home so I’m not sure if there’s a way to see it on there like the old webui pnginfo tab.
>>106525683
He’s a gigachad with the patience of a saint. Imagine if you get a bad seed lmao

>>106525687
>>106525690
Interesting copeing mechanism

>>106525705
Give an example and I'll do it better with Qwen Image.

>>106525650
>>106525683
>man on a pentium mogging this entire general

>>106525675
Yeah, and with SD1.5 is great with IPadapter and cns! What's your point?
>I have an actual professional workflow
lmao

>>106525650
maybe because people are mogging his comfy session
also wtf is people with unsecure access like that, I don't get it

File: 1741763838352272.gif (481 KB, 220x228)

481 KB GIF

>>106525650
>unsecured comfy installs
Huh?

>>106525708
I've already tested qwen image edit, it's great for what it is but its instruct style prompting is creatively limiting and it can't handle the edge cases that controlnet can.

>>106525796
If you're a "professional" using AI and can't train a lora you're actually a poser.

>>106525822
>>106525822
>>106525822
>>106525822

>>106525791
he's gonna find u >;)

>>106525808
You should stop posturing, I've been around here since before loras were a thing in diffusion models. Qwen's T2I prompting is limited, by design even.