Blessed be the anon in this bread

official pixart bigma and lumina 2 waiting room, now with maybe not safe cat

How do you guys right click with the nodeUI, it opens two context menus with one click. Drives me fucking crazy

is there an image viewer for linux that'll show the metadata from auto1111/comfyui?

Blessed thread of frenship

waitin for the day i can masturbate to my pixarts

MADE THE OP COLLAGE TWICE IN ONE THREAD HAHA WOOOO

>the cat is real

petition to rename general to /lcg/ - local cat general

>>101383597
Maybe it has to do with your browser? I've never encountered that before.

fucking civitai fuck!!

>>101384530
>Maybe it has to do with your browser?
I'm using Edge

>gen a few images on civitai to test
>looks good
>download loras
>triple check every hash including checkpoint
>exact same dimensions, sampling method, steps, cfg, seed
>check prompt and lora weights
>gen locally
>completely different
this shit seriously makes me cry
and it's not just off by a little, like a different seed, there's something visibly wrong with the whole style

>>101383507
kolors sucks why is it in OP
why is it above pixart especially

>>101384976
what? kolors is way better than shitxart

>>101385882
DiT models are superior

looks like a new version of ideogram just came out yesterday
https://civitai.com/models/573014?modelVersionId=638761

>>101385995
>DiT models are superior
True, but Kolors is still better than pixart because it has been trained better, and has more parameters aswell

>>101386034
Kolors in it's final form will be nowhere near as good as Pixart in it's final form

>>101386045
time will tell

>>101386289
Almost great pic. I think there's too much purple. You can't have too much stuff between yellow-blue transformation

>>101386309
I loved the voice acting in the original bullfrog game. How the cyborgs would say the name of the weapon when they equipped them... Minigun! Leveling buildings was mind blowing

File: PA_0002.jpg (829 KB, 3328x1152)

829 KB JPG

>>101386353
I miss that game/company

>>101386362
Theme park was another amazing game of theirs. Pure soul. At least all these game can be played with modern hardware

File: ComfyUI_AuraFlow_00001_.png (1.6 MB, 1024x1024)

1.6 MB PNG

This Auraflow thing doesn't seem to know its fantasy animal anatomy very well.

File: ComfyUI_00139_.png (1.35 MB, 1024x1024)

1.35 MB PNG

>>101386411

File: PA_0005.jpg (921 KB, 3328x1152)

921 KB JPG

>>101386353
purple color removed

>>101386424
Now it looks like a painting. There is cobalt blue that goes to chromatic yellow. Very nice!

File: ComfyUI_AuraFlow_00005_.png (1.34 MB, 1024x1024)

1.34 MB PNG

>>101386422
That's marvelous and all, probably a great showing of prompt adherence, but it's still not great anatomy.

File: PA_0010.jpg (812 KB, 3328x1152)

812 KB JPG

>>101386309

>>101386511
>>101386526
is it the new 900m pixart model?

File: PA_0012.jpg (1007 KB, 3328x1152)

1007 KB JPG

>>101386537
Nope, just still PixArt and BooruMadness mix

File: PA_0014.jpg (731 KB, 3328x1152)

731 KB JPG

>>101386526
>>101386554
Silver is for...

>>101386554
did you try this? https://civitai.com/models/573014/900m-pixart-sigma

>>101386568
magic monsters? I just started replaying, bloody baron quest is next

File: PA_0017.jpg (830 KB, 3328x1152)

830 KB JPG

>>101386603
I'll check it out. Let me make some room by deleting Auraflow

>>101386603
>https://civitai.com/models/573014/900m-pixart-sigma
Got a request for first gen?

>>101386608
Stench city.

File: 1689191953651433.jpg (1.11 MB, 2731x1131)

1.11 MB JPG

>>101386673
this is a good theme ;)

>>101386673
it's still the same team that made pixart sigma that made this one aswell?

File: PA_0023.jpg (744 KB, 2560x1536)

744 KB JPG

>>101386603
since you didn't request anything >>101386673
Your prompt will be.
>magic monsters? I just started replaying, bloody baron quest is next

File: PA_0024.jpg (727 KB, 2560x1536)

727 KB JPG

>>101386859
I'm sad that I can't mix the models now.

>>101386859
>>101386882
these are pretty nice. Perhaps a giant demon toddler could go with the theme

File: rp0dn71py3cd1.png (1.12 MB, 1024x1024)

1.12 MB PNG

https://vocaroo.com/17tvudBFkcKm

[Verse]
He built AuraFlow, an AI dream,
To generate images, so it would seem.
But a problem lurked, oh what a gaffe,
A chubby cat would make us laugh.

[Chorus]
Oh, AuraFlow, what have you done?
Ideogram's data, it sure was fun.
He forgot to filter, didn't think it through,
Now every picture has a feline coup.

https://reddit.com/r/StableDiffusion/comments/1e1ktdh/auraflow_sure_does_like_making_the_ideogram/

damn I gotta try the new 900m

File: 1720829559326696.png (1.04 MB, 1024x1024)

1.04 MB PNG

>>101386159

File: mnsfw.png (1.4 MB, 1022x1022)

1.4 MB PNG

>>101386928

>>101387187
lmao, nice

File: redditmode.png (1.37 MB, 1010x1004)

1.37 MB PNG

>>101387303

>>101386052
>>101386034
>DiT models are superior
>True
Anon where is your disconnect.

File: maybe not reddit gay.png (1.4 MB, 1042x1046)

1.4 MB PNG

>>101387367

>>101387519
>>101387367
>>101387303
>>101387278
why are you spamming this picture anon?

It's already made more traction across the entire internet in the last week than you will in your entire life so don't worry about it anon.

>>101387597
I'm small fry, why not try to go up against banking ceos or something lol. Sad you go after the little guys just to get your sad waifu out here who will be forgotten in a month because your posts are garbage lmao.

>>101383661
kwrite/gedit

I was going to write an addon for this and found quickly that all the linux image viewers are very unfriendly to this.

File: PA_0039.jpg (401 KB, 2560x1536)

401 KB JPG

>>101387229
It's decent.

>release AuraFlow
>People spend a couple of hours amazed by it
>Immediately become persona non grata once it becomes obvious you just scraped idiogram

File: file.png (146 KB, 256x256)

146 KB PNG

sports car on a road

>>101386844
No they just stretched the weights from the original Sigma and trained a little on top.

>>101388386
vroooom

>>101388225
the only people amazed by it were jeets. the 'prompt comprehension' suffered the exact same issues SD3 did, where stuff looked like pasted together clipart. i tried it and right away i felt the slop

>>101388225
Who really cares about the cat thing, if he fixes the dataset like he said he will for 0.2 then it should be a good model.

>>101388581
Reddit is calling in DoA and their opinion matters a LOT more than yours.

>>101388853
>one guy = all of reddit
That was you wasn't it?

>>101388874
One upvote is more important the entirety of the 4chan archive.

brave take, I know, but auraflow is worse than sd3

>>101389202
Id say auraflow is far better at prompt adherence but is far more undertrained and so has that plastic undetailed look like pixart does.

>>101389272
he didn't train a custom clip. anything you say about 'prompt adherence' is just complimenting sd3

>>101389300
Its using T5 which was not trained by sd people either.

>>101389272
what do you mean by plastic look?

>>101389318
Missing details. Smooth / plastic looking. That is a sign of a undertrained model. Smaller details fill in as a model is trained further.

>>101389300
>>101389314
That said the model is simply better trained than sd3 at least when it comes to prompt following. Has nothing to do with T5.

>>101389321
like that?
>>101387945

>>101387945
That's pixart but yes. Pixart is also extremely undertrained. That's why people use sdxl as a refiner step for it but use it for a base with its prompt adherence.

File: PA_0075.jpg (424 KB, 2560x1536)

424 KB JPG

>>101389390
We need new PixArt. There is only so much you can squeeze out of this one.

>>101389446
Which T5 is better, original PA T5 or SD3 T5?

File: Capture.png (97 KB, 1208x515)

97 KB PNG

>>101389478
I use SD3 T5

>>101389514
>merging models
Do you have errors warning in logs?
I have errors but it seems still working fine

Missing UNET keys ['pos_embed', 'y_embedder.y_embedding', 'blocks.28.scale_shift_table', 'blocks.28.attn.qkv.weight', 'blocks.28.attn.qkv.bias', 'blocks.28.attn.proj.weight', 'blocks.28.attn.proj.bias', 'blocks.28.cross_attn.q_linear.weight', 'blocks.28.cross_attn.q_linear.bias', 'blocks.28.cross_attn.kv_linear.weight', 'blocks.28.cross_attn.kv_linear.bias', 'blocks.28.cross_attn.proj.weight', 
etc

>>101389478
They are the same. Neither group finetuned T5.

>>101388131
I like this a lot, wow

do you guys ever feel bad about losing

>>101389724
We lost?

>>101389724
? we have at least 4 upcoming models that are looking good. Lumina, huyaian, larger pixart, and now auraflow. And there is also another würstchen-like model in the works that I knew of as well.

File: 900mPixartSigma_anime_00062_.jpg (452 KB, 1024x1024)

452 KB JPG

This model is bad with hands, feet and thighs. And face sometimes I got very bad result too.

>>101389859
Makes sense. Pixart itself was extremely undertrained. 900M is just someone stretching it even thinner.

So now we know AuraFlow is a scam. Is it time to learn Chinese?

>>101389758
are lumina and hunyuan really baking new models?

>>101390039
as far as I know they are all works in progress. Pixart / auraflow were just proofs of concepts and are working on bigger models.

>>101389976
if you're not going to support local models just go back to sdg and stop ruining out general

We're eating so well my tummy full :(

>>101389976
It's not a scam it's a genuine good faith effort, they were just a little retarded about dataset cleaning.

not eating well enough to make gens. another 3 day thread inc

>>101390153
yeah so what

comfy bred desu

>>101388386
HYPE

Read previous btw

File: 57548.jpg (419 KB, 1440x3120)

419 KB JPG

i heard tell this place was less retarded than /sdg/. jury's out, but odds are in your favor!

>>101389478
SD3s clip is obviously less to load than the behemoth that is T5
>better
SD3s but only marginally. Likely not enough for most anons to care

>>101389202
they both used synthetic images in training so equally worthless

>>101383661
Allusion sometimes works

File: 57461.jpg (967 KB, 1440x3120)

967 KB JPG

turns out they're even more retarded. you hate to see it folks! SAD!

What kind of computer do I need to run AI locally? Can a normal laptop work or do I have to buy a desktop computer gaming rig?

>>101390640
sadly, ur laptop is probably not up to the task. it could possibly work, but it won't do so effectively. you need to spend thousands of dollars on 4090 to break into the "rich and retarded" club, aka "we've got all the its cuz we're stupid rich; eat shit and die poorfags, lmao"

>>101390640
nvidia 1080 is plenty to get you started

its == it/s
now watch it's gonna trim the slash between it and s because of course it will. big ghey

>>101390660
>>101390671
Thank you. AI is continually using up less resources per compute, I may wait a bit longer before buying a better computer.
I imagine next gen of laptops will start being geared towards running AI.

>>101390640
Check out how much a used rtx 3090 sells for on ebay, and see if that's within your price range (it'll be like 40% of the total cost of the PC, something like that).

>>101390640
>Can a normal laptop work
In burgerland, your local wall market likely has at least a couple for sale that would work.

>>101390735
>rtx 3090
It's pretty expensive! I guess you've known those kinds of things for a while, have they typically been getting cheaper over the years? (excluding inflation which inflates the price of everything)

>>101390153
>another 3 day thread
>>101390440

>>101390752
>It's pretty expensive!
Yeah that's why you buy a used one (second hand). They're pretty much binary in terms of whether they function or not, if they send you a GPU that doesn't work then you just use ebay's refund system.

I suspect prices will drop a lot in a year from now once the 5000 series comes out because 3090 will be two generations behind by then.

>>101390640
You need a computer with a beefy graphics card or else it's going to be very frustrating.

>>101391139
wow what a le heckin EPIC reply my dude xD

File: file.jpg (1.18 MB, 2560x1536)

1.18 MB JPG

>>101391139
kek, any reaction is a good reaction i say. cool gen btw

File: ComfyUI_Kolors_00836_.png (1.66 MB, 1216x832)

1.66 MB PNG

>>101391256
I do like your gens a lot. Don't get me wrong.

so for comfy: manager, impact pack.. what else?

>>101391316
WAS suite

>>101391316
rope

>>101391316
https://github.com/pythongosssss/ComfyUI-Custom-Scripts

File: file.jpg (1.24 MB, 2944x1408)

1.24 MB JPG

>>101391275
ty anon
>>101391316
plasma sampler and https://github.com/ClownsharkBatwing/RES4LYF

>>101388581
Will he restart the pretraining from the start though? Because the cat pictures have poisoned the model hard, and fuck AI sloppa pretraining, his model is "good" at prompt understanding only because he lazily used ideograms outputs for that, he just wants his model to be a cheap copy of ideogram. I'm not surprised for a chink though, the chinks are notorious to make cheap copies of something kek

File: ljj4defa86cd1.png (2.14 MB, 1024x1024)

2.14 MB PNG

>>101391539
>he just wants his model to be a cheap copy of ideogram
That's just sad man, local cucks only want to be cheap copies of the big guns, how about being your own thing for once?

Looks like google made a better model than GPT4V, now will it be opensource or not?
https://arxiv.org/pdf/2407.07726v1

>>101391539
>Will he restart the pretraining from the start though?
allegedly
>only because he lazily used ideograms outputs
nah, likely his dataset was thoroughly tagged maybe something with the architecture too

>>101391639
>nah, likely his dataset was thoroughly tagged maybe something with the architecture too
not at all, he just used ideogram images to train his model and because ideogram is great at prompt understanding, it gave the model this edge, it's not that deep

>>101391657
elaborate one how simply using idograms images in pretraining = better prompt comprehension
it comes down to how the images were tagged the only thing using those images did was turn the outputs into mega inbred sloppa

>>101391669
>it comes down to how the images were tagged
but anon... if he scrapped the ideogram pictures from the sites, it means he got the pictures AND the prompts by the users, and because the ideogram pictures matches the prompt we make well, it means it's already beeing tagged "well", do you understand or not?

>>101391681
>nd because the ideogram pictures matches the prompt
so it was due to the captions being better than other models, yeah thats what i said
if he used real images with the same accurate captions it would not look like sloppa

>>101391691
I agree, using AI pictures on the pretraining is blasphemous in my book, don't force that on your users and stop being lazy, use real pictures and tag them well by yourself or with an AI captioner, training an approximator with approximative images is like recording a VHS on a VHS tape, you just lose more and more accuracy
https://www.youtube.com/watch?v=nqy_hYDI0As&list=LL&index=3&ab_channel=JaphyRiddle

>>101391587
The inbreeding is incredibly apparent when you compare them side by side like this kek

Does anyone know a website that specializes into inpainting?
i've tries 5 ways from everywhere to try and make stable diffusion inpaint properly but all i get is a blurred area where i've set the inpaint zone

What's your solution to get tons of variation in the generations? I just want to see random shit.

>>101391916
>I just want to see random shit.
Truly? No prompt with a high-ish cfg.

>>101391927
How random is this? Do you get any structure?

>>101391916
you can try wildcards on civitai, but i find random stuff is pretty garbage. instead im curating my own prompt lists that get combined

>>101391988
How about you try automating something in your shitty life?
"Get combined" isn't what you actually mean. You mean you spend fucking eighty hours at a time trying to slap random shit into your stupid fucking textbox.
Fuck you.

>>101392016
trust me my life is extremely automated, I have nothing to worry about at all.
my very detailed comfyui workflow slaps the random shit into my stupid textbox for me, but yeah I did spend 80 hours setting it up.

>>101391956
>Do you get any structure?
no. maybe the solution your looking for is simply prune your prompt? hard to guess without knowing what tools youre working with

>>101392062
So share it.

>>101392016
>>101392098
>he doesnt know about wildcards
kek

>>101391916
>I just want to see random shit.
look at the latent images before they go into the sampler
literal random noise

>>101392159
Shit, not noise. Random things.

File: file.jpg (895 KB, 1536x2688)

895 KB JPG

>>101391916
>>101392177
>Random things.
Replace some words in your prompt with random symbols as in instead of "woman standing in an empty field" try "=}$~>]:-|_ standing in an empty }''/.@%$"
But also like >>101391988 says try wildcards

>>101392230
The fuck is the prompt

>>101392230
post modern beach sunset? nice

The default workflow should set seed to fixed. So many people have bought into "AI is completely random" because they don't understand what the seed is.

>>101392230
I recognize those symbols

>>101391256
good stuff

>>101392230
Run this prompt in your workflow, in seed 27671888707425
please

Animated Ghost breathing out an image of =}$~>]:-|_ 

File: 0.jpg (390 KB, 1024x1408)

390 KB JPG

>>101389695
thanks

>>101393169
How many seeds did that one take?

>>101389664

PixArt: UNET conversion has missing keys!
['y_embedder.y_embedding']
model_type EPS
Missing UNET keys ['pos_embed', 'y_embedder.y_embedding']

That's all I'm getting in mine since the update.

>>101393211
3 shots, all good

>>101393480
How it feels to chew 5 gum

File: SDXL_0003.jpg (538 KB, 1664x2432)

538 KB JPG

>>101393169
>>101393505
17 for me

>>101393564
This stuff will be real time in a few years, then it'll just be picking the winner instead of waiting.

>>101393654

By that count it will just run 20 times, compare each image to prompt and pick the winner for you

>>101391587
What a joke, it's not hard to scrape a 10 million image dataset.

>>101393860
he's just a lazy ass, he prefers to scrap AI pictures because it has the image + the prompt, but training an AI with AI pictures that mess up anatomy, perspective, lightning is such a retarded idea I just have no words...

>>101393924
With hardware resources it's not even expensive to caption a 10 million dataset especially with something like Llava 1.6 Mistral (which is pretty accurate and uncensored) . And this fag is acting like he's having some sort of super sekret captioning system.

File: 00146-457612265.png (745 KB, 960x536)

745 KB PNG

>>101387945
>>101389446
>>101393564
>burred background on literally anything that has a hint of photorealism and includes a person in the image
I hate pixart and sdxl so much it's unreal

>>101394445
It's the braindead aesthetic filters

>>101394445
Do you have an image that doesn't have that effect?

>>101395598
That's it

hibernation mode

>git pull
>nothing works
I never learn

File: SDXL_0004.jpg (410 KB, 1664x2432)

410 KB JPG

>>101393026

>>101388386
How long has it been in the oven?

>>101397575
Awhile, a month of near nonstop but I keep adding training data so it's probably not optimal.

>>101397607
Hero we don't deserve

>>101397690
But we needed

>>101393553
kek.

File: ComfyUI_00009_.png (1.28 MB, 1024x1024)

1.28 MB PNG

Auraflow is slow as balls. Is there any way to speed that up?

>>101399154
no, I tried everything.

>>101399154
comfyui's tensorrt works with it.

>>101399760
>>101399773
>>101399784
That's pretty good for a base

>>101400095
the key is to use artists' names

File: 00011-3707032091.jpg (291 KB, 1344x1344)

291 KB JPG

>>101398999
>>101399154
Excellent

>>101400505
900m?

>>101399433
>tensorrt
link to help a newfriend?

>>101391744
Adobe firefly is your best bet, but they only allow 25 images per month.
Also, Neural love may be better, but they only allow 5 pictures PER LIFE, and new accounts don't replenish that, which surprised me because I created them in a different browser.
But they knew it was me.

>>101391916
Have you looked at
https://perchance.org/fusion-ai-image-generator
The whole point of perchance was to create free generators of anything and they became the best at it by having people submit their best random lists, so when they got into the image generation business they allegedly became the best at prompt generation too.
It looks so complex I didn't even bother to touch it, but it's like wildcards on steroids.

>>101400716
>>101401034
Nice

>>101400844
Interesting.

File: PA_0176.jpg (641 KB, 3328x1152)

641 KB JPG

>>101393026

File: Grid.jpg (2.04 MB, 3840x2176)

2.04 MB JPG

>>101392930
good symbols
>>101393026
left (:2.0) right (:1.0)

File: file.jpg (849 KB, 1920x2176)

849 KB JPG

>>101401367
>(:2.0)
*(:1.5)

File: PA_0211.jpg (810 KB, 1664x2432)

810 KB JPG

>>101393482
decent prompt, I don't get but...to each their own

>>101401405

File: PA_0212.jpg (699 KB, 1664x2432)

699 KB JPG

>>101401405
>>101401412
forgot image

https://x.com/_akhaliq/status/1811864979710107843

>>101401034
What was the prompt btw

>>101401809

Watercolor and oil painting hybrid of a ghoul with distorted features, performing an arcane ritual, surrounded by Clutter-Mechanical (by H.R. Giger) that reflects on the surroundings, tendrils of fire weaving through the air with swirling smoke, Gabriele Dell'Otto and Bob Peak-inspired, bright saturated colors,

i shleep till bigma

>>101401199
>yellow to blue
instant kino

>>101401367
I'll take the left one.

File: 1705495765008.jpg (1.04 MB, 1920x1216)

1.04 MB JPG

it just drives me crazy
feels like there's consistently more detail in civit's gens
already found their safety stealth embeds and added them, still just something lacking

>>101403187
I was going to say some sort of loopback thing but it looks like something else hm
That is interesting wtf

>>101403187
might be related to how different model GPUs will produce varying details

>https://arxiv.org/abs/2404.07724
isn't this something that could be implemented fairly easily

>>101403187
can you drop a reference and workflow, I'll take a look?

File: file.jpg (938 KB, 1536x2688)

938 KB JPG

>>101403187
never used civit, didnt realize the difference was that apparent

>>101403629
civ: https://litter.catbox.moe/38w57h.jpg
local: https://litter.catbox.moe/4re3hv.png

Using Forge, everything should be in the image. I also tried Automatic1111 and the results were pretty much the same as Forge. Oh, I had to set randomness to CPU-derived instead of GPU to get them to look as close to civitai's as they do, somewhere in the settings.

Checkpoint is pony v6.
Loras:
https://civitai.com/models/188514?modelVersionId=211718
https://civitai.com/models/352581?modelVersionId=573829
The embeds are here:
https://civitai.com/models/99890/civitai-safe-helper (civit_nsfw)
https://civitai.com/models/222256?modelVersionId=250712 (safe_pos and safe_neg)

>>101403946
>https://civitai.com/models/99890/civitai-safe-helper (civit_nsfw)
>https://civitai.com/models/222256?modelVersionId=250712 (safe_pos and safe_neg)
Very interesting

File: 1702090823856.jpg (910 KB, 2496x672)

910 KB JPG

>>101403962
It says they're SD 1.5 embeds, and they don't show up in webui's textual inversions list, I don't know if that means they're supposed to "not work" for SDXL but they definitely still influence the image when you type the keywords in manually.

>>101403975
oh and this series was made using AutismMix_pony rather than normal pony as checkpoint just fyi

>>101403946
>>101403962
are these "unlisted"? cant find them with the search bar but i am retarded. do you know of any others?

>>101404007
Yeah I don't know what's going on there. I discovered the embed names in the metadata of images genned on civitai, I had to use google to find the actual model URLs.

>>101404007
nta but I'm logged in on Civ and I managed to download them. I can reup them to catbox if you want.

>>101403946
Lora values, Latent size, Sampler values (Steps, cfg, etc) are all matching Civit

I'm not sure if this is the correct thread to talk about it since theres so many AI threads now but

What's the best free and local video upscaler nowadays? I'm currently still on Real-ESRGAN

>>101404448
Pirated version of some Topaz labs software might be the best. I've used chainner + 4x-UltraSharp earlier, it was ok

What good hacks to the gen process do you use? Not a precise category of stuff but here's a list of what I mean:
Things that are clear benefits: Kohya hires/deep shrink, Euler SMEA Dy sampler, dynamic CFG, LLLite and models like "replicate" that work very well as controlnets (and are compatible with Kohya hires)

Sometimes benefit: Attention guidance (ex PAG), regular controlnet, detection-guided inpainting (but I usually just do manual masks)

And are SDXL mixes still the least shitty overall? I downloaded HunyuanDIT but haven't gotten around to working out a Comfy workflow for it. Have dabbled a bit with Sigma.

>>101404453
Nothing is more tiresome than people who say stereotypes don't exist. It's one of the most grander forms of gaslighting.

>>101404674
>Things that are clear benefits: Kohya hires/deep shrink, Euler SMEA Dy sampler, dynamic CFG, LLLite and models like "replicate" that work very well as controlnets (and are compatible with Kohya hires)
>Sometimes benefit: Attention guidance (ex PAG), regular controlnet, detection-guided inpainting (but I usually just do manual masks)
can you explain what these do?

>>101404763
Kohya hiresfix starts a gen at a downscale factor, then finishes it at a higher resolution. Leads to higher detail level than upscale models or img2img, but is more prone to decoherence. More detail and more decoherence the higher you push the resolution.
Euler SMEA Dy is just the best sampler I've used for both detail level and overall coherence, it especially works better with nonstandard gen resolutions so it's by far the best sampler I've found for Kohya hiresfix.
There's several dynamic CFG nodes/plugins, I use this one https://github.com/Extraltodeus/ComfyUI-AutomaticCFG Broadly, they all help prevent "frying" artifacts from high CFG while also helping you dial in contrast and sharpness and other stuff about the gen.
LLLite is a lighter weight controlnet
https://huggingface.co/kohya-ss/controlnet-lllite
https://civitai.com/models/136070?modelVersionId=164965

>>101404827
thanks I'll try all of these

File: 00204-304922023.jpg (692 KB, 1260x1680)

692 KB JPG

>>101404674
Negative Guidance minimum sigma, Latent Modifier, FreeU (sometimes)

>>101404750
Indeed. It serves as excuse for not releasing new stuff as open source. Tho I understand that these are nvidia guys and they have to toe the line

>>101404887
I'll also note for Kohya hiresfix/deep shrink I use 0.5 end_percent, the default in most setups is 0.33

>>101404914
It's a fool's errand because it's censorship of reality. It's why vision models won't say the gender and skin color of the subjects.

>>101404674
Deep Cache, increases iteration speed by like 30% at the expense of quality (so then you increase steps to get the quality back).

>>101404914
By latent modifier you mean something like this? Which parts do you use most often?
https://github.com/Clybius/ComfyUI-Latent-Modifiers

File: 00001-3852095653.png (1.24 MB, 896x1152)

1.24 MB PNG

Good afternoon

>>101405033
I use the Forge integrated version

>Which parts do you use most often?
Rescale Cfg Phi, Extra Noise Multiplier, Affect Uncond, Dyn Cfg Augmentation

>>101405007
Tinkering with deep cache is interesting, dunno if I'm into it.

did that hand fixer thing ever come out as a usable plugin for 1111 or comfy?

>>101405313
Just need that 16ch vae to perfect the hair

File: grid-0070.jpg (507 KB, 1792x2400)

507 KB JPG

>>101405421
teeth too

>>101403187
maybe civit have their own noise schedule like leonardo

>>101405615
>>101405501
>>101405313
ipadapter or celeb to get the consistent face?

>>101405731
lora + prompting for nationality

Here we go...

>>101405779
>>101405779
>>101405779

>>101405501
>>101405615
nuts. it looks real to me.

If its animated with unlocked chat 4o and locally stored (customised version at leas) it could be great experience.

>>101405904
>nuts. it looks real to me.
few runs with i2i + color correction and some artificial degradation and then these look real

>If its animated with unlocked chat 4o and locally stored (customised version at leas) it could be great experience.
and turn those into talking chatbots etc? could be interesting

File: file.png (95 KB, 256x256)

95 KB PNG

>1boy wearing jeans and a t-shirt

First time this started to resemble the prompt beyond being a floating pair of jeans.

File: 20240714_210240.jpg (403 KB, 1152x2048)

403 KB JPG

>>101383507

>>101407409
Nice