File: ldg3.jpg (946 KB, 1999x1999)

946 KB JPG

>beat me to the bake
Have my collage.

>not included in either collage
I'm going to kill myself

>>101813750
DO A FLIP
https://youtu.be/QibeKQ9W1UU?t=9

Blessed thread of frenship

File: ComfyUI_00297_.png (1.96 MB, 1024x2048)

1.96 MB PNG

>>101813658
schnell isn't good.
>>101813688
I don't fully understand it, but if you're doing the setup which lets you use negative prompting and you drop the CFG down from 6 to like 2.5 it becomes invisible. More steps can help too.
>>101813742
>2 of my gens in the last thread OP, 2.001 of my gens in this one
Feels good man

>thread blessing
we bac

>>101813742
>tiny miku
kek

since when does discord hook into comfy? the "Prompting [...] with Comfy" status thing

>>101813810
You can make a discord bot do anything you want

Your lack of respect for DALL-E will be your downfall when DALL-E 4 is revealed.

>>101813836
>>101793904

>>101813836
DALLE-E Snore more like

>>101813836
>Your lack of respect for DALL-E will be your downfall when DALL-E 4 is revealed.

We already have Flux. By the time Dalle 4 drops we will have Flux v2 either on the way or already here. There's nothing to worry about since open > closed, and Flux's release has rendered Dalle useless for anyone who's smart.

>>101813876
>Flux's release has rendered Dalle useless for anyone who's smart
That's what someone who is not smart would say.

fluxpro.art banned me for genning all those vaginas, lmao

>>101813750
Neither am I and I made one of them

>>101813886
kek

File: ComfyUI_00310_.png (1.56 MB, 1024x1024)

1.56 MB PNG

Okay Flux, we're cooking

File: flux-ups--2024-08-10-0306(...).png (3.59 MB, 1920x1536)

3.59 MB PNG

>>101813957
based

>>101813730
>>101813742
nice collage

File: ComfyUI_00321_.png (1.56 MB, 1024x1024)

1.56 MB PNG

You've converted me, mikuanon

File: Capture.jpg (35 KB, 844x442)

35 KB JPG

Does this shit do anything? I tried playing with it but it didn't change the picture

>>101814092
I saw people playing around with "realism" of the output by putting these parameters extremely low, like 0.3-0.8

File: Capture.jpg (339 KB, 3308x1695)

339 KB JPG

>>101814092
Yeah Idk man I tried a lot of values, even 0 0 and I always got the same picture, is there something wrong with my workflow?

>>101814147
your workflow is fucked

>>101814168
post yours right now

File: file.jpg (198 KB, 1672x680)

198 KB JPG

>>101814183

Are any of the DynamicThresholding, Guidance Neg/Pos, AdaptiveGuider, SkimmedCFG et c at all applicable to schnell? I guess not given the docs

>>101814225
Holy based

>>101814147
The purple node before modelsampling being "bypassed" means it's literally being bypassed, so any shit connected TO it is still going to the next node, am I getting this right?

>>101814225
How do you get the straight lines?

>>101814238
yup

>>101814147
why are there two line coming out of unet loader? that means your nodes are not all using the same config, you have one path that goes through ModelSamplingFlux and another that does not.
The node works for me btw

Do you really need to have a 24 gb vram graphics card to run inference of the biggest and better Large Language Models locally?
When will models become efficient enough and their size reduced without losing quality so that they are able to reasonably be ran locally on a 4gb card or even 2 gb?

File: Capture.jpg (269 KB, 3372x1175)

269 KB JPG

>>101814238
>>101814243
Ok I removed the purple node and nothing changed

>>101814250
what do you mean by that? Here's the workflow: https://files.catbox.moe/8l9d2u.png

>>101814270
Your basic scheduler is connected straight to the model itself, not the modelsampler

File: Capture.jpg (273 KB, 3221x1255)

273 KB JPG

>>101814250
Ok I think that's it, thanks anon I didn't know it wouldn't work like that

How well does flux make 2B art???
I'm going to start genning this weekend.
Ty!

>>101814256
>Do you really need to have a 24 gb vram graphics card to run inference of the biggest and better Large Language Models locally?
You need >100GB of VRAM to run Llama 3.1 405B at a low quantization.
Even LLama 3.1 70B won't fit on 24GB, anything smaller than that is too dumb to bother running.
>>101814270
I mean you have one line going to ModelSamplingFlux and another going to BasicScheduler, they need to be in series, not parallel.
>>101814285
Nice.

>>101814240
https://github.com/pythongosssss/ComfyUI-Custom-Scripts
Link Render Mode

>>101814285
For some reason Dynamic Thresholding breaks the image for me.IIt does increase the generation time like it's supposed to but all I get are blank white images.

>>101814290
Flux does not know 2B

>>101814311
>It does increase the generation time like it's supposed to
that's not the fault of dynamicThresholding, that's the fault of CFG > 1, that shit halves the speed

>all I get are blank white images.
Did you use this workflow anon?
https://reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/

File: Capture.jpg (351 KB, 3348x1411)

351 KB JPG

Holy fuck you can tell he's enjoying his burger, kek

>>101814290
>How well does flux make 2B art???
It can do pixel art, Miku and Trump, and that's it

>>101814290
>2B art
Try Hunyuan.

>>101814349
and Meghan Markle for some odd reason

>>101814365
In particular
https://desuarchive.org/g/thread/100907329/#100910377
Was meant to be 2B (I purposely didn't prompt for the blindfold and also prompted that particular style), freeway finetune should be even better at this.

>>101814408
>Was meant to be 2B (I purposely didn't prompt for the blindfold
if the model knew who 2B was it would've automatically put the blindfold

>>101814092
why does it have a resolution? why can't it adapt to the resolution you normally put on the Empty Latent Image node?

>>101814454
it doesn't have that info at that point in the graph

>>101814461
that's bullshit... there's probably a way to get that information before starting the graph thing, it's retarded to have to change resolution on each node

>>101814433
Not how Hunyuan works anon. I got many gens that it had blindfold and more of her features matched, also see https://desu-usergeneratedcontent.xyz/g/image/1718/07/1718073928017.png

I asked for Gojo's blindfold and just said "purple orb", the model inferred properly what they looked like and how the technique is used. 2B is a video game character, so if we imply that in prompt the model will get closer. Hunyuan works best when you prompt it this way, if it seems to forget to had amazing recall ability.

>>101814454
>>101814461
Just set up a couple of primitives called Height and Width and pipe them into both the latent image and ModelSamplingFlux, then you only have to set it once.

File: Capture.jpg (10 KB, 579x219)

10 KB JPG

>>101814472
this? how does that work?

>>101814496
right click -> convert widget to input
on the node you want to connect to

File: file.png (107 KB, 756x504)

107 KB PNG

>>101814503
>>101814496
Once you've done what the other anon said, drag from the primitive to the new input and the primitive will automatically convert itself to the right mode.

I don't even know if this is NSFW or not because I can't really identify the body parts
https://litter.catbox.moe/8vm4mu.png
just in case
jesus

File: Capture.jpg (185 KB, 1657x1364)

185 KB JPG

>>101814496
>>101814503
Ok I think I got it, thanks anon

>>101814469
>to had
it has*

>>101814329
> A picture of a bitch with a nice ass who is dressed like a maid who also knows how to kill you.

File: 94682d26fac980f764f7f8bf7(...).jpg (263 KB, 896x1152)

263 KB JPG

>>101814290
This is who it thinks 2B is

File: Screenshot_20240810_23025(...).jpg (588 KB, 1079x1078)

588 KB JPG

>>101814679
No

File: flux_00100_.png (1.67 MB, 1344x768)

1.67 MB PNG

*cough*

>>101814871
magnificent.

File: FD_00152_.png (1.89 MB, 1024x1024)

1.89 MB PNG

>>101814837

File: Capture.jpg (58 KB, 632x664)

58 KB JPG

For those using DynamicThresholding, I think you should change the threshold_percentile from 1 to 0.9, 1 is simply overkill and destroys the colors
https://files.catbox.moe/y2h7zd.jpg

Is tiled VAE decoding built into comfy or do I need a node?

File: 1716813091812757.jpg (58 KB, 738x1292)

58 KB JPG

>Heres you deep fry nipples anon enjoy

>>101814978
thats one thing that comfused me, I always used 0.97 or so and never 1

>>101814986
>double-click the interface
>type "tile"

>>101814989
https://files.catbox.moe/h8avmw.png

File: ComfyUI_temp_kzdxu_00041_.png (1.06 MB, 1024x768)

1.06 MB PNG

still gooning to nuns

File: file.png (13 KB, 355x128)

13 KB PNG

>>101814468
Of course there is.

>so many newfags on comfy
cute

>>101813836
DALLE is already impressive in its current version, but every time it spouts nsfw you get dogged.
It literally doesn't compete with local.

File: FD_00156_.jpg (1.07 MB, 4096x1024)

1.07 MB JPG

It doesn't like schizores :(

>>101813886
Did they ban you afterwards?
Is there some manual reviews?

>>101815084
let me guess, I need to install a new package to get that one?

>>101815133
I dunno but it works again so I must have been flagged for manual review and passed whatever test I was flagged for.

>>101815149
let me guess, you're even unable to understand the information given to you in the image

>>101814256
Anon, the bottleneck is real, why do you think most LLMs/quantizations magically need 24GB of VRAM ?
If we had a "4090 titan" with 64GB of VRAM, local LLMs would target that instead.

>>101815149
See the text in the green box above the node? it's the name of the custom node it came from. Comfyanon rarely ever does qol improvements so this shit is basically unusable without custom nodes.

>>101815169
>Comfyanon rarely ever does qol improvements so this shit is basically unusable without custom nodes.
can't wait to go back to forge, his software is fucking ass

>>101815175
Probably not gonna happen any time soon, illya is reworking the whole thing, and reforge is a half-working mess made by an amateur.

>>101814349
The way they didn't take out Miku is so weird.
It also knows Goku.
I guess they're way too well known or something.

>>101815175
illyas gonna abandon forge again. at least comfy is consistent.

bless flux for filtering a1111, comfy is so good, thank you for making me learn.

>>101815205
>comfy is consistent.
his spaghetti shit is consistently bad yeah

I think I solved the sylization of flux for painterly styles.
- One prompt with everything (style and subject)
- Gen with Flux
- Gen with SDXL
- Use SDXL Image as IPAdapter style only input.
- Gen with SDXL-Ipadadter model using Flux image and low denoise (0.5,-0.6)

>>101815224
works on my macheen
and has done so since ~December

>>101815237
>I think I solved the sylization of flux for painterly styles.
or you can just do that
https://reddit.com/r/StableDiffusion/comments/1enm9og/discovered_by_accident_a_trick_to_make_flux/

>>101815224
That's the least of the problems, especially with shit like get/set nodes and UseEverywhere that let you save anything globally and access it from elsewhere without any spaghetti

File: ComfyUI_00348_.png (2.96 MB, 4096x1024)

2.96 MB PNG

>>101815127
Skill.... issue?
I think with high enough steps you could get something decent out of it, but it's not really worth it given it gens like 5x slower than at 1MP

>>101815261
>https://reddit.com/r/StableDiffusion/comments/1enm9og/discovered_by_accident_a_trick_to_make_flux/
I tried it, it's very far from what I am looking for.

>>101815224
note how I did not include a quality tag describing the experience of working with comfyUI.
>>101815263
gotta admit tho, steep learing curve. my first weeks with comfy weren't fun.

>>101815224
It's just the way Blender works.

>>101815263
get/set node
The what now?

>>101815303
1) you don't spend your whole time doing spaghetti shit on blender
2) and? fuck spaghetti shit, is that clear enough?

>>101815320
I don't do spaghetti shit. My workflows are neat, square, and organized. I only build things once, and then I can import them as modules.

>>101815333
How do you manage your modules? Having them all in that dropdown seems like it'd be a nightmare.

>>101815320
I don't pass my time doing spaghetti on comfy, I set it once then generate.

>>101815347
finally, a vagina

>>101815320
okay nigger do you use flux with diffusers in python whats your point

>>101815361
the dropdown isn't that hard to manage, I think? That's just for select modules like IPA or different CNs. I always maintain the same base workflow
But I can either just copypaste them from a different interface, or just save them as screenshots of the workflows, with the workflow embedded in the file.
Just throw it into the window, ctrl+c, ctrl-z

but I rarely need to import things like that - i usually just have them bypassed somewhere.

File: file.png (122 KB, 1126x592)

122 KB PNG

>>101815311
Just another set of custom nodes, werks like this.
Notice how the model and positive/negative inputs glow in the ksampler, same with the clip input in both text encoders - they automatically take the value stored in anything anywhere nodes. Like setting a global variable or something. Get/set nodes connected to each other, you can name then however you want to save as many things as you want.

>>101815389
NTA but I do something like that. finally managed to get it to be faster and more stable than the UIs

>>101815411
I feel like its not a huge stretch from comfy

>>101815441
yeah. I definitely stretched it though

File: 4 Grids Collage.jpg (720 KB, 1999x1999)

720 KB JPG

The community is going to be retards and focus on schnell, because they are RAMlets, aren't they?

File: ded.png (3.02 MB, 1477x1855)

3.02 MB PNG

>CRITICAL HIT

>>101815477
What's that last model and doesn't schnell require the same amount of VRAM?

File: ComfyUI_00357_.png (820 KB, 512x2048)

820 KB PNG

>>101815477
Also the licensing bullshit, you can't do training on Dev and make money off it, so all the guys who have patreons will do Schnell instead.

>>101815523
https://comfyanonymous.github.io/ComfyUI_examples/flux/#simple-to-use-fp8-checkpoint-version
And Schnell does, but it gens in 4 steps as opposed to 20 so vramlets can have a chance to finish in under an hour per image.

>>101815571
i enjoyed this funny image

>>101815571
That's not what the license says. It says you can't use Dev to train a competing model. Donations become a grey area, because technically they can be seen as profiting.

I wonder what it takes to get the AI sloppa aesthetics out of a model.

>>101815588
Can I go back inside the womb?

>>101815604
>I wonder what it takes to get the AI sloppa aesthetics out of a model.
this? -> >>101815261
https://imgsli.com/Mjg1Nzk5

File: ComfyUI_00361_.png (856 KB, 2048x512)

856 KB PNG

>>101815571
>>101815127
Looks like the model doesn't care about total resolution, just that either direction isn't above 2048. Beyond that you run into the usual Cronenberg issues.
>>101815600
Sure, but why take that risk when the plebs will happily slurp up Schnell quality anyway.

>>101815600
>Donations become a grey area, because technically they can be seen as profiting.
So in simple terms it's DOA because no one is gonna risk it

>>101815600
>can't use Dev to train a competing model
Any finetune is a competing model

>>101815639
that's not important
the important thing is that you've found something that can help you feel superior to others

This so called realism LoRA doesn't do shit. The outputs are identical.

>>101815671
consider this
you're doing it wrong
or
everyone else is faking their outputs

>>101815642
>>101815655
>>101815629
>Non-Commercial Use Only. You may only access, use, Distribute, or creative Derivatives of or the FLUX.1 [dev] Model or Derivatives for Non-Commercial Purposes.
>Non-Commercial Purposes

>In the case of Distribution of Derivatives made by you, you must not misrepresent or imply, through any means, that the Derivatives made by or for you and/or any modified version of the FLUX.1 [dev] Model you Distribute under your name and responsibility is an official product of the Company or has been endorsed, approved or validated by the Company, unless you are authorized by Company to do so in writing.
https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md

>>101815696
I will translate this into retard. You are free to train fine tunes of Flux.1 Dev. Everyone is a dumb piece of shit who has a reading level lower than a 5 year old.

>>101815678
I very probably am because there's no instructions on this. Cranked the strength up to 2 for a 40x slower gen, but let's see if anything happens

File: ComfyUI_00362__.jpg (582 KB, 2048x2048)

582 KB JPG

>>101815657
Anon if I was capable of feeling superior to others I wouldn't be here on an imageboard.
>>101815612
There's no romb inside the woom
>>101815696
>You may only access, use, Distribute, or creative Derivatives of or the FLUX.1 [dev] Model or Derivatives for Non-Commercial Purposes.
I've read this like eight times and it makes no sense
>You may only... or creative Derivatives of or the FLUX.1 [dev] Model...
???
Was this written by an ESL?
What does 'You may only... ...creative Derivatives' mean?
Even if that's a typo and they mean 'create Derivatives', the next bit makes no sense either.

>>101815775
elitism sure tries

>>101815629
every model has some rare condition where it doesn't look like AI slop once in a while

>>101815762
>Cranked the strength up to 2 for a 40x slower gen
lora strength doesn't affect generation speed

>it's called flux dev
>you can't dev with it

>>101815775
You may only access, use, distribute or create derivatives of the FLUX.1 [dev] model for non-commercial purposes.

>>101815856
>it's called pro
>you actually pay money instead of making profit

>>101815856
>>101815714
retard-kun is right, you're free to train finetunes.
You just can't charge for them.

>>101815775
>Was this written by an ESL?
Yeah, Americans. I have to assume they mean 'create'. I cut some of the thing off but here's the rest of the Non-Commercial clause
>If You want to use a FLUX.1 [dev] Model a Derivative for any purpose that is not expressly authorized under this License, such as for a commercial activity, you must request a license from Company, which Company may grant to you in Company’s sole discretion and which additional use may be subject to a fee, royalty or other revenue share. [Email here]
If you want to use flux.1 Dev for commercial purposes i.e make money off it, for example host it on a site and sell token access, you need permission from BFL and you need to give them a cut.

More context for distribution and creating derivatives.
>in the case of Distribution of Derivatives made by you, you must also include in the Attribution Notice a statement that you have modified the applicable FLUX.1 [dev] Model; and
>in the case of Distribution of Derivatives made by you, any terms and conditions you impose on any third-party recipients relating to Derivatives made by or for you shall neither limit such third-party recipients’ use of the FLUX.1 [dev] Model or any Derivatives made by or for Company in accordance with this License nor conflict with any of its terms and conditions.
>In the case of Distribution of Derivatives made by you, you must not misrepresent or imply, through any means, that the Derivatives made by or for you and/or any modified version of the FLUX.1 [dev] Model you Distribute under your name and responsibility is an official product of the Company or has been endorsed, approved or validated by the Company, unless you are authorized by Company to do so in writing.
If you train a fine tune from Flux you must give them credit for the base model, but you are soley responsible for it, and any legal action that might come from it, for example you create a cunnytune, the feds will come for you, not BFL.

File: file.png (727 KB, 640x480)

727 KB PNG

>>101815902
It's written by Germans, Flux is a German model
It's amazing how fucking stupid Euros are

>>101815879
it's pro as in professional

>>101815886
this one is cool, reminds me of lice

>>101815860
Minus the cat ears this art style reminds me of a game I used to play back in the 90s on Mega Drive but the name escapes me. There was a barbarian, a wizard, and a sexy scantily clad woman who is the only one I ever played as.

>>101815604
Never understood the idea of AI slop, isn't it just when you see many variations of the same theme so you get tired of it and attribute that to the model being "slop" instead of you just getting used and bored to it?

>>101815991
When you see enough AI gens you realize how incestuous the outputs are, things like "sameface" and homogenous aesthetics. Every model has an underlying repeating style.

File: ComfyUI_02940_.png (1.09 MB, 1152x896)

1.09 MB PNG

>>101815984
definitely the vibes I'm going for, 90s pc games have a great style

File: Screenshot 2024-08-11 010457.png (101 KB, 798x251)

101 KB PNG

>>101815942
>Blattman
I just thought of a cool super hero

File: ComfyUI_00289_.png (1.06 MB, 1024x1024)

1.06 MB PNG

>>101816010
perfect timing, right as I'm going to bed

>>101816027
Yeah I can understand that, the lack of variations when you prompt without asking for an artist, style etc.

>>101816044
https://www.youtube.com/watch?v=qOXYnT3aMzU

File: FD_00009_.png (871 KB, 768x1024)

871 KB PNG

>>101816044

>>101816027
it is a skill issue

>>101816047
>lack of variations when you prompt without asking for an artist, style etc.
sometimes flux switches between realism and cartoon so it might just be a model size issue. 700 milllion parameters in SD 1.5 isnt a lot

>>101815942
stupid enough to make flux and Mistral huh

>>101816113
My guess is that it's because "styles" and "artists" are mixed and not/badly labelled.

>>101816115
It's incredible that the only competitive text/video models outside of the US ones are French (Mistral) and now German (Flux).
Where are the Japanese, Korean, English?

File: FD_00019_.png (1.23 MB, 768x1024)

1.23 MB PNG

>>101816092

>>101816146
one look at data sets of automatically captioned images, even the ones that used the latest and greatest VLMs, shows that's exactly the issue. the VLMs don't know much so the captions are pretty generic.
OpenAI trained their own caption model for DALL-E 3 (which was trained on 95% synthetic captions) and clearly it is leaps ahead of what is available

File: 1712532946467425.png (2.11 MB, 1040x1008)

2.11 MB PNG

>>101816194
OAI also didn't scrub mentions of "Character xxx", "Celeb yyy" and "Artist/Style zzz" or even a lot of nude/nsfw, instead choosing to do the moderation a posteriori by rewriting prompts and analysing outputs, which is the wiser choice imo.

>>101816168
Tencent is making one. Hunyuan is from them, and they are doing an LLM too.

https://github.com/mcmonkeyprojects/sd-dynamic-thresholding/commit/5d63447afbc44b377f706a5eb0430f85791dcf30
The script that allows AdaptiveGuider and DynamicThreshold to work at the same time has been integrated to the main repo, no need to download the script anymore on this tutorial:
https://reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/

>>101816247
Now draw her pregnant

>>101816258
>Hunyuan
Isn't that specialized into Chinese stuff only?

>>101816294
Not only, but it obviously has a Chinese focus, being Chinese and all.
Sigma is also a Chinese team I believe, but NVidia bought them so who knows.

where's a millionaire furry to make FluxDev the best it can be? It's been over a week already.

File: FD_00033_.png (1.14 MB, 768x1024)

1.14 MB PNG

I would do it but I am not a furry nor am I a millionaire.

>>101816319
Nice then, I want so much more competition than what we have now.

File: 1713415968419208.png (2.23 MB, 1040x1008)

2.23 MB PNG

>>101816293

File: 1240250997.png (922 KB, 1024x1024)

922 KB PNG

Just checked the latest threads to try my hand at the wuxia gens, flux may not do sword riding but goddamn does it make for some sick grinds

File: ComfyUI_02565_.png (1.8 MB, 1024x1024)

1.8 MB PNG

The red haired girl was supposed to be a tomboy, can't believe flux doesn't know that concept :(

File: FD_00047_.png (1.23 MB, 768x1024)

1.23 MB PNG

>>101816554
Just describe what she is supposed to look like

>>101816274
Nice, I could finally test it in my workflow, and I've gone from 70s to 50s generating an image (using the cuda0/1 trick too).

File: ComfyUI_02568_.png (1.7 MB, 1024x1024)

1.7 MB PNG

>>101816573
Not only it makes it faster but having cfg 1 at the very end kinda cleans the artifacts the high CFG made previously, it's really a win/win situation

File: file.png (61 KB, 499x629)

61 KB PNG

love the group map from the EasyUse node set

File: GUhoHqdX0AAxtCB.jpg (764 KB, 1463x2048)

764 KB JPG

Any artists or three-letter codes that will give me flat shading like this in pony derived models?

File: file.png (691 KB, 3072x935)

691 KB PNG

>>101816630
also stole this guys 2x flux upscale group and made few changes to it. can recommend

https://civitai.com/models/620294

>>101816630
How does that work?
How do you configure the nodes?

File: Rdg_101.jpg (161 KB, 1448x1280)

161 KB JPG

Morning

File: ComfyUI_Flux_0005.jpg (333 KB, 2048x1152)

333 KB JPG

>>101816664
>How does that work?
it auto-detects the node groups you created and allows you to turn them on/off. see >>101816662

File: FD_00059_.png (1.33 MB, 768x1024)

1.33 MB PNG

>>101816637
Junji Ito

File: 980226402.png (1.33 MB, 1024x1024)

1.33 MB PNG

>>101816667
Good morning

File: ComfyUI_02570_.png (1.16 MB, 1024x1024)

1.16 MB PNG

>>101816637
Seems to be really hard to do, I really did my best there, CFG 6, GuidanceNeg 10, boomer prompt made by Claude 3.5 Sonnet...

File: ComfyUI_Flux_0001.jpg (384 KB, 2048x1152)

384 KB JPG

>>101816664
>>101816681

the rgthree node pack also has group muting/bypassing nodes, in case you dont trust the chinese language on the EasyUse node set github

>>101816681
>>101816713
I installed the easy module thing.
It's really a useful idea

>>101816724
>yasuke-sama....

>>101815958
and it's professional as in doing this for profit

Why do random old nodes I used before (unconnected to current nodes) reappear when I refresh comfy?
Does that happen to anyone else?

>>101815158
there's literally soft cheese pizza on the gallery every now and then so I can't fathom what would be a bannable offense when you're genning vaginas

File: flux_upscaled_00062_.png (3.5 MB, 2400x1368)

3.5 MB PNG

>>101815992
nice

>>101814739
use a description of her from wiki it'll be closer

File: ComfyUI_02575_.png (927 KB, 1024x1024)

927 KB PNG

>YoRHa No.2 Type B from Nier Automata standing over a luxary car, on the bottom of the image, there is a white text subtitle that says "Why does Flux only knows Miku..."
I had some hope that writing the full name of 2B would trigger something accurate, oh how wrong was I

File: ComfyUI_Flux_48.png (1.13 MB, 1216x832)

1.13 MB PNG

>struggled a lot with tying to make a realistic Miku because it always force anime, or at the very least 3d shit
>mfw all it took was adding "a woman cosplaying as..." to the prompt

>>101816826
it doesn't work if you put "anime, 2d" on the negative prompt?

File: file.png (387 KB, 687x757)

387 KB PNG

First time trying flux. Not bad.

I have 16GB of VRAM. How over is it for me if I want to run this local?

>>101816748
Well, it's broken for now, will wait until the commit fixing it is merged.
https://github.com/yolain/ComfyUI-Easy-Use/issues/300

>>101816852
>I have 16GB of VRAM. How over is it for me if I want to run this local?
it's completely over, buy more ram my nigga :(

>>101816852
how much RAM? Even with 24GB VRAM people have to offload heavily.

File: ComfyUI_Flux_43.png (1.31 MB, 1216x832)

1.31 MB PNG

>>101816847
Miku is too strong for that, even the negative mumbo jumbo with 10 flux guidance doesn't work by itself

File: flux_upscaled_00063_.png (3.38 MB, 2400x1368)

3.38 MB PNG

Has anyone tried experimenting with max_shift and base_shift?

>>101816852
I have 12GB

>>101816870
16 as well. Would 32 be enough RAM?

>>101816890
Nice. How much RAM?

>>101816900
32GB but usually I need around 26-27GB

>>101816810
>>101816667
Very nice. Making see-through injection molded parts was an old fetish of mine, I have since given up on. Which model are you using?

File: a.jpg (1.64 MB, 2201x4096)

1.64 MB JPG

>>101816890
this is the only thing ive found on it

>>101816870
Yes, from what I see you need 24GB for model + 10 for clip.
So either one card > 34GB VRAM or two cards with 24+12.

>>101816935
>max/base schift: 0.1/0.5
wait, how can the max be lower than the base shift?

>>101816935
So basically that shift thing is an undercooker? nice, I'll try it out and see if it helps at high CFG

File: Rdg_131.jpg (203 KB, 1448x1280)

203 KB JPG

>>101816932
I'm partial to CrystalClear and PixelWave

File: fp045.jpg (382 KB, 1024x1024)

382 KB JPG

>>101816932
i'm using flux.1 pro on glif.app, not sure of the other anon

File: flux_upscaled_00056_.png (2.79 MB, 2400x1368)

2.79 MB PNG

>>101816935
thanks. It seems to have barely

>>101816982
any impact* is what I meant to write

File: ComfyUI_Flux_49.png (1.11 MB, 1216x832)

1.11 MB PNG

>>101816797
Well, Flux tried.

Now that flux is here, is SD kill?

File: ComfyUI_02581_.png (1.7 MB, 1024x1024)

1.7 MB PNG

>>101816935
Ehh that's not bad at all, do we know what the default values are when we're not using this node?
https://imgsli.com/Mjg2MjYy

File: ComfyUI_Flux_0009.jpg (407 KB, 2048x1152)

407 KB JPG

>>101816970
>>101816955
>>101816982
>>101817072
this guy also talks about the shift values in regards to his upscale workflow (timestamped)

https://youtu.be/L6_yi539V-s?t=362

>>101817123
Oh cool, I was tired on trying those values randomly without knowing what it really means, thanks anon

How do I use wildcards in ComfyUI?

>>101816797
>>101817026
I guess flux, like most virgin models, are still pretty bad at anime?

>>101817141
>trying those values randomly
90% of stuff I see in random nodes in comfy.
Zero explanation in guides, zero explanation in their own git.
Great.

>>101817184
I would say at 75% of the time you kinda get what one parameter means by just trying the max and min, but for that shift shit it's so subtle I don't really know what it does kek

File: flux_00123_.png (1.57 MB, 768x1344)

1.57 MB PNG

>>101817123
interesting. Probably only works when using flux for upscaling. I'll give his workflow a try

File: ComfyUI_00201_.png (1.81 MB, 1024x1024)

1.81 MB PNG

>>101817167
Tbh it's pretty good at anime

>>101817221
As much as I love flux with the bottom of my heart, I don't think I'll survive multiple months gening only Miku and Trump while waiting for a big finetune because Flux seems to only know 2 characters kek

>>101817221
where sonic?

>>101817123
So basically it's the difference between base and max that matters? so it's the same thing when you do base 2 + max 3 against base 3 + max 4?

File: ComfyUI_Flux_51.png (1.17 MB, 1216x832)

1.17 MB PNG

>>101817167
I just copy-pasted her entire design description from the nier fan wiki and added "A high-quality, high-polygonal 3d anime image featuring a woman standing in a desert".
Also that's one big fucking head.

File: 1708533602169748.png (267 KB, 683x645)

267 KB PNG

avatarsissies....... people think we are REPULSIVE FREAKS

File: FD_00099_.png (849 KB, 1536x512)

849 KB PNG

>>101817276

File: fp047.jpg (347 KB, 1024x1024)

347 KB JPG

>>101817308
purrfect

>>101817307
WTF SATAN WOJICKI IS REALLY DEAD, RIP BOZO LMAOOO

>>101817308
kek'ed

File: FD_00101_.png (774 KB, 1536x512)

774 KB PNG

>>101817261
What about Meghan Markle and Sonic?

>>101817361
yay...

File: FD_00108_.png (1.7 MB, 1536x1024)

1.7 MB PNG

>>101817434
Come on we can have some fun with that

>>101817468
Flux can't into gloves.

>>101817468
your picture is a bit overcooked, you can fix it with AdaptiveGuidance, dunno if you have it on your workflow yet?

>>101816890
32GB Ram and 16GB VRAM, I'm not having any major issues other than it being slow.
But if it works, it works. It's better a superior model that runs slowly than a fast model that only gives you what you are looking for by chance. Makes experimentation more tedious but you are better at the end. What I am saying is: run Flux no matter the cost.

File: ComfyUI_02588_.png (1.02 MB, 1024x1024)

1.02 MB PNG

>>101817307
dalle would NEVER!

File: FD_00110_.png (1.61 MB, 1536x1024)

1.61 MB PNG

>>101817562
This is CFG 1.0, Adaptive Guidance won't help. Also I can't be fucked.

>>101817588
I agree. It takes around 2:30 per image + a couple seconds to upscale. With 16GB it'll probably be around 2 min

>>101817641
>Adaptive Guidance won't help.
oh it can, you can really uncock the shit out of the output with it
https://imgsli.com/Mjg2MTI0

File: 01913cdc-c197-7573-a244-d(...).jpg (95 KB, 1024x1024)

95 KB JPG

pixel ass

>>101817665
Left looks like low cfg, I am using 3.5

>>101817698
it was cfg 6 + guidanceNeg 10

>>101817703
Right so I am not using Dynamic Thresholding, it's a straight flux gen at cfg 1.0 and guidance 3.5
I mispoke but yeah, adaptive guidance won't do anything because it's already at 1

>>101817719
oh ok, I get it now

>>101816681
>>101816662
What's the name of the node?
Searching groups gives me nothing after installing https://github.com/yolain/ComfyUI-Easy-Use

File: file.png (13 KB, 231x295)

13 KB PNG

>>101817764
Right click on an empty space and turn it on

>>101817781
ooh it's not a node
OK thanks anon

File: file.png (3 KB, 133x164)

3 KB PNG

>>101817764
>>101817781
also there should be these icons in your comfyui to make it appear. for me its in the bottom left corner

File: ComfyUI_02592_.png (1.32 MB, 1024x1024)

1.32 MB PNG

>>101817307

https://civitai.com/models/636355/flux-detailer
https://imgsli.com/Mjg2MjU4
https://imgsli.com/Mjg2MjU5

>>101817850
you made that lora anon? that looks nice

File: FD_00124_.png (966 KB, 1280x768)

966 KB PNG

OC Donut Steel

>>101817850
>soulless vs soulful
impressive

File: file.png (2.53 MB, 1024x1024)

2.53 MB PNG

>>101817803
lol

>>101817798
oh indeed

>>101817307
I remember when ldg threads were 2 days each and everyone was like "it's better this way"

>>101817946
which was like 2 weeks ago

>>101818058
Pre-flux, so it's before the current era

>>101817850
How did you do that?

File: Sigma_12541_.jpg (2.25 MB, 2048x2048)

2.25 MB JPG

>>101817946
>>101818058
Sigma bump

is there any prompt to get a character with a receding chin? I've tried everything to no avail.

>>101815942
>Average german RCD test

>>101817946
It kept the thread schizo away

>>101817946
slow threads are always the best since the only people that hang out then are the ones that actually care

>>101817946
yeah I remember when diffusion was stale

File: devschnellblockmerge.jpg (736 KB, 3072x1024)

736 KB JPG

On the Schnell vs Dev discourse. Schnell isn't great, especially at text, Dev can have issues with prompt adherence when CFG is 1. Schnell converges fast but has quality degradation. Dev takes more steps and can lose coherence on some details. (the text on screen in the prompt for example, prompt was: computer screen with text that reads "an image of what a gecko would see crawling up the side of a building." with a smug woman in a desk chair, view over shoulder, sweater, )

Think the best speed and quality tradeoff for people that aren't high VRAM chads is to do the block merge tip from https://blog.comfy.org/august-2024-flux-support-new-frontend-for-loops-and-more/

Or use a pre-merged diffusion model from https://huggingface.co/maximsobolev275/flux-fp8-schnell/blob/main/flux1-schnell%2Bdev_fp8_unet.safetensors that uses the workflow from the Comfy post.

Way better than Schnell as is when targeting 4 steps, but remember the license implications that this merge takes on the non-commercial aspects of Dev.

>>101815477
>comfy mix
what's that?

it's so easy to add text to an image after it's generated. why are they trying so hard to get text right, it's a waste of time.
the exception is if someone is holding up a sign or something with text on it but even then the actual text isnt the big problem

>>101819094
???
because text is basically in every image

>>101818936
>base schnell doesn't give women buttchins
Maybe it's not that bad after all...

>>101818936
>but remember the license implications
lol
sure thing
lmao

>>101819094
Yeah there could be different model for text that works like adetailer

>>101819569
gm

File: ComfyUI_Flux_61.png (976 KB, 1216x832)

976 KB PNG

Those are some weird Miku and Donnie

>>101819795
what does this have to do with local diffusion?

ran starting his schizo-rants again
this place was much better when you were gone, hardly any bitching at all

>>101819846
schizo anon from the other thread is trying to actively sabotage this one now that sdg is becoming less relevant. it's unfortunate for him because it had the opposite effect, at least on me. i don't even check the other thread anymore because of this.

>>101819893
you go to sleep and wake up thinking about debo, its incredibly pathetic

>>101815127
This is very cool promvt?

Let's hope this fresh bread can stay a bit cleaner...
>>101819993
>>101819993
>>101819993

>>101815860
>>101816034
prompt me up nigga

>>101815902
I'll use my gens for commercial purposes and there'a n9thing they can do about it.