🏞｜general-with-images

[scenic yew]

Nope, just a prompt. No lora, no upscaling, nothing else.
I don't have the exact prompt now, but I used an llm to create a new prompt: "The watercolor painting portrays a subdued and thoughtful depiction of Luffy, the adult lead of the One Piece series, standing amidst a lush and dense forest. His hand cautiously holds the hat that rests on his head, while his face is partially shrouded in shadow, suggesting a hint of disappointment or sadness. The atmosphere is slightly darker than a typical day, creating a moody ambiance. In the top right corner of the poster, the title "One Piece" is displayed in elegant letters that resemble the style from the show."

This is the image I get.

[formal wadi]

That's cool. I tried doing that without using any loras when I used stable diffusion and it just created a random looking character. What's an llm?

[royal monolith]

LLM = Large Language Model - like ChatGPT

[formal wadi]

ohhh

[brazen sand]

ah ok

[jade torrent]

/LAST WEEK on Kickstarter. still there is time to help us fund. Years have slipped past like a river flowing into the sea, and humanity has, in its own peculiar way, led the world to the brink of collapse, as if it were destined to be. But make no mistake—this tale is not concluded. Hatred brews beneath the surface, violence strikes without warning, and pestilence entwines itself in the very heart of our civilization. Now is the moment to forge order from the chaos, to seek out justice, even if it requires a touch of vengeance. The Sword of Justice is ready to be wielded, and its time has finally come. BACK NOW THIS SUPERBAKER MUST HAVE, MUST READ INDIE COMICBOOK. JOIN US
http://kck.st/3Bl6F3g

[hard escarp]

Hello guys

[unreal kayak]

Create an image for me

A fish with heads

How do I write prompts

[rain gazelle]

props to you if you know the reference

[sick gust]

Hi, can u draw me a fox girl

[steep sleet]

GGUF_Flux

[stoic fiber]

Hey there, anyone has any experience in recreating a subject with 100% fidelity runing it thourgh the k sampler ?

Here is what I want to achieve

[brazen sand]

I just finished some images for a music video, not perfect but I can only inpaint until certain point

Still images as I wasn't ever able to make img2vid

I upscaled x4 with Ultimate SD Upscale and got those upscaler artifacts, usually I try to upscale x2 and then x2 again and I think that way that lesser the artifacts, any way I inpainted the most important details that has the worst artifacts

inpaint is a tedious process, it can take several hours, even with much experience in it. At least I know when I just need to set some values and then just let it produce like 20 gens, select the best one, repeat.

[outer portal]

damn i need to rebuild my model generator

[stoic fiber]

Hey guys, I'am looking for an experienced comfyui individual (paid job) ! 🫡

[steep sleet]

£10 cash

[brave mortar]

/Halloween in Jar Coloring page, pixar style, simple outline and shape, easy and big coloring page, large print, single coloring page, coloring page black and white --v 6.1

[steep sleet]

https://discord.com/channels/851231576494702613/1294323099155693568

[languid pebble]

The real reason why they had to sink the Titanic ... 😄

[steep sleet]

... I can see Leo and Kate ... !!!

[languid pebble]

YaY! 😄

Must have been before she tried to fly ...

[south temple]

well friends its been forever 🙂

is there a latest something that everyone is loving as of late?

[steep sleet]

Flux inside PortraitMaster

[south temple]

how new / old is that?

and WHAT is that lol

comfy?

[steep sleet]

https://github.com/florestefano1975/comfyui-portrait-master

Use Flux as the checkpoint, but do not use the ControlNets

[south temple]

that... is in depth

[steep sleet]

OpenPose. Canny. Depth. Flux does not work with SDXL ControlNets.

[south temple]

portrait master is deep lol

[steep sleet]

But it does great portraits using Flux otherwise

[south temple]

I need to read up on flux next

its been a while 😉

SD3 ever exist with any merit?

[steep sleet]

SD3 has been largely 'forgotten' since Flux arrived

[south temple]

AH so FLUX is new to me

thats SD3 is where I stopped playing for a bit

pulling flux in now

[languid pebble]

The state of the art right now ...

[south temple]

Looks fun

HUGE files

neat

gets pretty specific

is the main better or is schnell a betetr use of time?

[languid pebble]

Schnell if you don't have a decent GPU and VRAM

[wispy nest]

I think most people would find that one flux image is worth 10 Schnell images TBH

which is roughly the ratio that it is

4 steps for schnell vs 40 for dev

I think its worth waiting

[south temple]

running some tests now

[granite mesa]

Hey guys, I'm trying to finalize my first AI influencer, and I'm close but I think it's still missing something.

Anyone have any tips for how I could improve these images? ( I have more, this is just a few random samples)

I'm not sure I have consistency 100% down but I think it's close. I'm wondering if I need to add filtering/tweak the lighting/contrast/etc?

I think I'm just too close to the problem at this point

[south temple]

sorry... ai influencer?

[granite mesa]

https://www.instagram.com/fit_aitana/?hl=en is the most famous example

[south temple]

interesting.

[wispy nest]

this looks too much like flux

it shouldn't rly be possible to tell which model you used

if its done well

[granite mesa]

Yeah that's super fair. Not sure how I could fix that I guess

[wispy nest]

its hard, to remove the flux look

[south temple]

euler or bust eh?

im testing all the randoms

[scenic yew]

Use schnell or some loras instead, schnell is more creative imo.

these are all 1-step btw

[wispy nest]

wow that's one step

we could take this and refine with SD 1.5 or SDXL

[scenic yew]

yeah with 2-step or more, schnell makes it too plasticky as well. You definitely do need to refine it to make it better.

its quantized as well, so you can probably get even better imgs with normal schnell.

[granite mesa]

Crazy for one step. Lemme give this a try, thanks!

[brazen sand]

yes, usually better everything and more original or unexpected in a good way

sadly my PC in the middle of the generation goes OOM, so it is not only that I have to wait for it to generate, I cannot watch yt or watch movies while at it, and working on generating means to check how it is doing from time to time, changing the prompt etc.
If only the performance was improved, I wouldn't mind the long generation times not that much

[wispy nest]

have you tried the smallest quants>

they get pretty small

[rain gazelle]

Been trying to make 'vintage travel posters' but.. I'm not terribly good at it. 😐

[steep sleet]

img2img these Public Domain Posters https://vintagraph.com/collections/wpa-posters

[ionic vault]

anyone know the name of this style?

[radiant turtle]

@ionic vault

[old delta]

Keanu Reeves sitting inside a Haas F1 car in the pit garage, wearing a racing suit with both Haas F1 and Toyota GR logos. The Haas F1 car features its signature red, black, and white livery, while a Toyota GR logo is visible on the side of the car. Surrounding him are team members in Haas uniforms, some working on the car, others discussing race strategy. The setting is a bustling F1 pit stop, with tools, equipment, and a Toyota GR banner in the background. Keanu looks calm and composed, preparing for the race

[steep sleet]

https://vintagraph.com/collections/travel-tourism-posters

[languid pebble]

Good morning coffee!

[still rose]

The urge to buy a 3090 or sum with 16gb of vram or higher just to learn comfy Ui with flux

[fair quartz]

hello

[steep sleet]

ComfyUI and Flux run well on my 8Gb VRAM RTX2070 with 64Gb RAM

[celest sigil]

Define "well" 🤣

[rain gazelle]

It's less about the style, and more about trying to capture specific places. (In this particular case, locations from a game.) There's some loras out there for the style (and it's even understood to some degree with things like "vintage poster, flat art" type descriptors.

[still rose]

was gonna ask that. Ill assume a 4060 ti 16gb is the best budget option

or would a used or refurb 3090 be (that would be prolly £100 to £150 more

[still rose]

bro said HAAS 😭

[still rose]

Hyper model doesn’t require that many steps

Iirc it’s like 4 steps or sum and like 2 to 4 guidance

[queen river]

definitely not. I just tried a simple image and they all turn out significantly better with 15+ steps. Its not worthit past like 30

and dont get me wrong but it can make GOOD images but it somehow only excels in portraits

like this is what i got with just "goth girl"

[pale steeple]

I would like to install pony to comfyui. What would be the easiest way to do it?
and what's a good workflow that works with lora?

[still rose]

Interesting for me hyper model is like 5 steps for me

[lavish garden]

Prompt: https://civitai.com/images/33788681

Happy (early) Halloween! 💜 - Garfield (Gijinka)

[pale steeple]

why is there no background in the image?

[steep sleet]

Hey! Where'd Dice go to?! 😄

Flux/Florence2

[languid pebble]

He has his own Discord Server now ...

[languid pebble]

Hungry now ^^

[clever oar]

👋

discord blocked in my country now 🙁

bypassing the block slows it down so much that emoticons are sent with a delay of 3 seconds 😃

[languid pebble]

Aww... sorry for you!

Hope you have a good VPN ...

[clever oar]

its not vpn i think

GoodbyeDPI

[languid pebble]

I don't remember my VPN ... I have 50 gig/Month ...

Windscribe ... you can check whether they still have some free amount ...

[clever oar]

why you need vpn ?

hm strange i off my bypass and discord still work

[languid pebble]

VPN is just another bypass ...

You connect to a Virtual Server somewhere in the world and it connects to the server blocked in your country

[clever oar]

in you country why you need vpn all works

[languid pebble]

I don't need ... would give you my account if it couldn't cause problems ...

Claude A.I. wasn't available here for some time ... so I used it ... it's been more Claude wanted to do a local test first ...

[clever oar]

work message but picture i cant see

if i off bypass

[languid pebble]

I'm sorry for you!

[clever oar]

💀

[languid pebble]

Today we've had a demo in my city ... people willing to install a kalifat ... I have to notice I'm loosing my patience ...

[frozen pasture]

do you think i can generate imagtes like this with stable diff?

[languid pebble]

Sure you can ... but Flux might be the better choice ...

[frozen pasture]

thank you... ill look into it...

[languid pebble]

Flux needs a strong hardware but will follow your prompt more ...

[frozen pasture]

as in good pc?

thats fine

[languid pebble]

SD3 you should search for a version fitting your needs

[languid pebble]

I'm using a 4090 ... it also works with lower settings ....

[frozen pasture]

ok thanks for the help!

[languid pebble]

It's not really easy ... so many different ways you can make it work ...

[frozen pasture]

noted

[languid pebble]

https://youtu.be/Ad97XIxaBak?si=Sl4_GwIRNyVXEp-o might be a good point to start ...

[deft walrus]

For those interested (if anyone) I'm working on a full set of Trump Cards from the Roger Zelazny books, the Chronicles of Amber.

[pale steeple]

are these numbers with these models normal for a beefy pc with a 4090?

it seems to take a while to generat

[stoic fiber]

guys, anyone knows why inpaint with brushnet and sdxl model changes the background color and how to fix it ? (Look at the flower on the left, you can cleary see the mask seem)

[wraith plank]

yo what u use for this, this looks insane

[languid pebble]

Old SVD technologie with some tricks ...

[wraith plank]

nice

technologie? german? im german

[languid pebble]

Ich bin aus Hamburg ...

[clever oar]

mojo

[nocturne oak]

@languid pebble

[pallid ruin]

Nice!

[clever oar]

thx

hi leo

[pallid ruin]

Hello!

[steep sleet]

Flux with PortraitMaster

[brave star]

two images side by side, A cute white tiger wearing Overalls, wave to the camera, Bubble Mart style, clean and simple design, IP image, high-grade natural color matching, bright and harmonious, cute and colorful, detailed character design, behance, Shanghai style, Organic sculpture, C4D style, 3D animation style character design, cartoon realism, fun character setting, ray tracing, children‘s book illustration style --niji 5 --style expressive --ar 3:2

[steep sleet]

I only got this far!!! 😄

[celest sigil]

They're not 2 images side-by-side though 😄

[steep sleet]

Teach me a good prompt to do that? 🙂

[celest sigil]

I just used this one #🏞｜general-with-images message

[steep sleet]

I used the original one as well ... using Flux

[celest sigil]

Same 🤷🏻‍♂️

Flux dev

[steep sleet]

Flux-dev-Q8

[steep sleet]

Portrait Master and Flux

[languid pebble]

Sweet! Thanks!

[celest sigil]

All of the double images are a single gen.

[languid pebble]

Good morning coffee! 😄

[ocean cloud]

Any idea what tools can generate such cool videos if fed with great ideas? https://www.instagram.com/jayprints?igsh=MThheGtiMGM0azFuaA==

[languid pebble]

You could try: https://venturebeat.com/ai/new-high-quality-ai-video-generator-pyramid-flow-launches-and-its-fully-open-source/

[still rose]

Tricked some1 thinking this was real XD

[clever oar]

👋

maybe due to the wrong version torch consumes more vram?
before i can generate with ip adapter without problem now i have out of memory

[vapid crest]

how to avoid this happening with ultimate SD upscale?

[quiet current]

dont feed it crack.

What settings are you using ? (screenshot the whole webpage)

[quick lagoon]

Creazy patato wering sunglass and smile fun serious around

#artisan-1 hello world

[nimble mason]

Here is the image you requested.

[vapid crest]

or you could drag it into comfyui

[quiet current]

Nothing indicated that you used confyui for it. I m not an ComfyUI guy but I don t see anything inherently wrong there. Have you tried another lora and model combination to be sure it s not specific to some model/lora/vae combo ? Also maybe try asking in #🤝｜tech-support or #🧣｜comfy-ui instead.

[vapid crest]

alright

[steep sleet]

PortraitMaster+Flux

[steep sleet]

Prompt: "An Art Nouveau style storybook page depicting the Whispering Woods scene. The page features elegant, flowing borders with organic floral and vine patterns. At the top, the title 'The Whispering Woods' is written in a beautifully ornate script, surrounded by decorative elements. The scene shows the shadowy forest with graceful trees, fireflies lighting a winding path, and curling roots, all framed by the decorative border. The colors are soft and magical, with deep greens, silvery moonlight, and glowing gold from the fireflies. The layout feels like a page from an enchanted storybook, with a mixture of imagery and text space."

[clever oar]

what he do with crocodile?😮

[steep sleet]

Skateboard!!!

[clever oar]

oh !

why not standing)

[steep sleet]

The boy is an amputee

[steep sleet]

Flux+PortraitMaster

[wispy nest]

thought they were smoking 😲

[steep sleet]

Hedgehoggery

[languid pebble]

Good morning coffee

[languid pebble]

Has to look up "Buzz" 😄

[steep sleet]

In Brit English - buzz = uptempo, positive excitement or vibe! As in "There is a buzz about the place!"

Flux+PortraitMaster

[steep sleet]

Phew! Flux+PortraitMaster

Almost Albino

Pies 'n Fries'll do that to you!

[granite mesa]

Do these look real? I've spent way too much time looking at them lately to tell

[jovial tiger]

No. too many anatomy things wrong with each one

[nocturne oak]

Exactly. First image, the knee on the ground is at an angle that looks very unnatural. Second has an issue with the rear foot placement. Third has a thumb in a very strange place in the odd looking gun. The fourth is probably the best of the bunch but her left eye (our right) is an issue).

[granite mesa]

Appreciate the feedback

[nocturne oak]

The tools are getting better all the time and you should keep working at it. Hell, even some of us who have seen insane amounts of these images still sometimes miss things. But if you trust your eyes and take a moment to think about what's natural, you'll be able to pick up the ones that are better than others over time. It's not impossible to produce passable results now, either; it just takes more effort than it will in about a year or two. (Keep in mind that image inference of any degree of complexity is only about 2 years old. That's staggering growth in capability.)

[granite mesa]

I'm still upscaling these ones, so the finish isn't there yet, but I feel like anatomy is better

[jovial tiger]

the best way to not have hand and leg issues, is to just not have hands and legs. 🙂

[nocturne oak]

It's always better when you cut off the extremities and 3 of the 4 here don't show hands or lower legs.

fest and I are cut from the same cloth

[granite mesa]

lmao true that

Barefeet are the worst though

[jovial tiger]

it's funny, but every model has limitations, just have to work within them

[nocturne oak]

One of the first things I do when a new model comes out is to begin to push various limits to see what the model is capable of. How can I break it? What does it repeat in the images over and over and over? Do I see similarities everywhere I look if I've generated a thousand images? Once you begin to know those things, you can work around them, play to the strengths of the models, etc.

[granite mesa]

Cleft chins in flux being the easy example

[nocturne oak]

This is why I don't worry too much about what some dub "Flux Face". SD1.5 was WAAAAAY worse with that. Flux can be manipulated with words to avoid the repeated facial structures.

Also, I knew what you'd say. 😉

[granite mesa]

I find a properly trained lora fixes the chin as well for the most part

I've seen it sneak through in a couple though

[nocturne oak]

Sure, but you don't need a LoRA to fix it.

Words will do the trick.

[granite mesa]

nah that's true

Honestly the bigger issue I've found is that it adds Jewelry constantly

and without a negative prompt it can be hard to avoid

[nocturne oak]

Jewelry can be toned down a bit, but yeah...that and beards on guys is one of the common issues.

Not even really full-on beards so much as 5 o'clock shadows.

All of it is solvable with words, though.

Something with Flux that is a real problem is the grid/bars.

[granite mesa]

what do you mean by grid/bars?

[nocturne oak]

Nearly every image from Flux has a grid or bar pattern to it. It's less noticeable at 1 megapixel than higher resolutions is most cases, but it's still there on nearly every image Flux produces. You can take steps to mitigate it, but once you see it, you can't unsee it.

[granite mesa]

interesting. can't say I've noticed it in my images, but I also do a ton of refinement so that might be why

[jovial tiger]

So you basically have to generate at 1mp, like 1344x768 in this case, and then just ultimate sd upscale at 1.25x, 8 steps, deis/beta, 0.18 denoise, 1024x1024 tile size, repeatedly until you get to the desired resolution. It's the only way to never have bars/stripes. I'm sure others have mentioned it previously, but I figured it out the other day. The downside is that there isn't much hi-res fix to be had. The other solution is to train lora at higher resolutions. I did one the other day after Clownshark kinda figured it out, and that indeed solved the bars issue. With that particular lora, I can now generate at 1792x1024 res etc and no more bars. When I trained at only 1024 and below res, bars all day long. The issue of course, is that almost no one trains at higher than 1024. Even 1344 barely fit in 24 gigs, and anything higher needs a larger video card, so fat chance of having that on most lora. So this one however was not trained above 1024, but with this iterative upscale, never have a bar anymore.

[celest sigil]

a very odd cult

[crimson mango]

What if Leia and her clone became empresses of the empire? 😂

[still rose]

Flux 👍

[serene gate]

Darth vader in a coffee shop

[sterile kiln]

https://www.youtube.com/watch?v=5tXh_MfrMe0

[dapper lava]

#😊｜co-creators people

[clever oar]

i eat potato fries now😎

and 2 pizza

[languid pebble]

Yummy!

[clever oar]

cheese pizza with pepper

[languid pebble]

pepper better than paper 🙂

[clever oar]

😆 lol

bad english

[languid pebble]

That happens ...

[clever oar]

but i feel i say something wrong )

[ionic vault]

do guys know if can do something similar like this on SD

didnt find any lora on civitai like that

[languid pebble]

To much text ...

[clever oar]

are you publishing your book?

[languid pebble]

Just a cover of the Leonardo A.I. Magazine 🙂

[clever oar]

this word magazine-in my county is shop🙂

[languid pebble]

Leo could be a Mario Brother, too

[clever oar]

ip adapter not work now out of memory, before it works

i change something(

[languid pebble]

Sometimes it's just an update ...

[clever oar]

😭

now work with low ram and 256 on 256 res

[languid pebble]

Awww... sad ...

[clever oar]

why i cant have minimum like before(

life is pain

[languid pebble]

yeah ...

[clever oar]

what update you think?

[languid pebble]

Sometimes an node in ComfyUI gets an update and becomes incompatible to other parts

[clever oar]

last i play with install cuda and torch maybe

[languid pebble]

There has been an update for torch that didn't work with xformers ...

but I have no idea whether that's still a topic....

[clever oar]

gpu can use ram for generate?

if i buy more ram it solve this ?

[languid pebble]

I don't think so ...

AFAIK with Flux the LLM uses RAM and the model the VRAM

[clever oar]

sad\

[languid pebble]

You can use online services ...

[clever oar]

when i start generate it takes 15.8 of 16 gb memory

not all 15 gb for it +system too

[languid pebble]

I'm sorry but I don't know much about your problems ... cause my PC is stronger ...

[clever oar]

i think need buy +16 gb

[languid pebble]

There are videos on youtube how to use Flux with 6GB VRAM ....

[clever oar]

i have 4)

[clever oar]

i use it but i like play sometimes with loras and adapters

love freedom

you dont know sd3 still bad?

[languid pebble]

Yes, they've never released a new model as promised ...

[clever oar]

sad

[languid pebble]

I think it's all about milking the cow at SAI at the moment and the people who loved the idea left ....

[clever oar]

I was so hoping that they would finish it to an adequate level🙁

[languid pebble]

There are some finetunes @ CivitAI

[clever oar]

It’s strange to me that one picture with an adapter requires more than 4 gigabytes

on sd 1.5...

[languid pebble]

You have an input picture to analyze and work with ... sure it needs more VRAM

[clever oar]

strange time i want buy gpu not for games🙂

[languid pebble]

A used 3090 could be an idea

[clever oar]

i check it now)

[languid pebble]

But ... your power supply and case will have to handle it ...

[clever oar]

hm not bad price

[languid pebble]

Iz's not the fastest but has a lot of VRAM

[clever oar]

but can was in mining

i want something new not big price lot vram and so that it doesn’t consume a lot of electricity

am I asking too much?😃

[languid pebble]

Well you definatly will want an NVIDIA ....

These old graphic cards where made for gamers ... so not really fitting to you 100%

[nimble mason]

yes lol

[languid pebble]

Hi Clown!

[languid pebble]

This world is not fair ...

[clever oar]

I once told a friend I have a video card with 1 gb vram and he told me why do you need so much😃

it was before ai

[clever oar]

make a public page for your city and write about it. I took a photo of how a local plant dumps waste.

[languid pebble]

A.I. is changing a lot of things faster than we are used to ...

[clever oar]

probably not a single game will reveal the potential of your video card🧐

[nimble mason]

alan wake 2 will

[clever oar]

lol

this game gpu killer

[languid pebble]

Already waiting for 5090 to be released 😄

[clever oar]

It seems to me that no serious movement in game graphics is expected for another 100 years

maybe that why we won't see hl3

[languid pebble]

They are not doing what they can, cause they have to sell to the masses ...

And the most don't have 24GB VRAM

[nimble mason]

if i were president, i'd make it a felony to sell, manufacture, or possess with intent to distribute, any graphics card with less than 24gb of vram

punishable by 1 year in prison for every gb shy of 24gb

multiplied by the number of gpus

[clever oar]

we would have to build prison cities😃

[sterile kiln]

https://civitai.com/models/818292/akira-toriyama-freezer-era-style

with the right lora it's possible

[pallid ruin]

For AI the best for the price is an used 3090, and second best maybe a 4060ti 16gb I think 🤔

[golden pebble]

how is this done?

[celest sigil]

With a prompt and AI generative model 🤷🏻‍♂️

[pliant prism]

for the comic book style yes

I will try to re-find it

with text no

[south sonnet]

All of my attempts to us Flux seem to fail, the preview is a grey pattern shown here. The final image is just all black...what am I doing wrong? I'm using Forgeui and the flux1-dev-bnb-nf4 model.

[wispy nest]

could you switch to comfy?

[south sonnet]

I could...but I'm not as used to that platform. I know it's better and I should learn it but I just haven't taken the time yet.

[pliant prism]

are you using a vae? the model you're using apparently already has one so you don't need to use one

maybe its that? im not sure

[south sonnet]

No vae...I tried with and without but results all failed.

[wispy nest]

its just that its not worth spending any time troubleshooting a forge issue
its better to just switch to comfy or diffusers

[steep sleet]

ae.sft?

[south sonnet]

I tried that and same result...but to my knowledge it's not needed

[south sonnet]

I think my issue could be related to CFG, what should I set distilled cfg and regular cfg to for the nf4 model?

[steep sleet]

CFG 1

Conditioning 3.5

[wispy nest]

it depends on the workflow

[steep sleet]

Usually CFG1 and 3.5 for Conditioning

[clever oar]

i want 4060ti now 🧐

car made from a blanket

warm in winter but damp in rain 🧐

[clever oar]

quadrober

not legal in my country

[verbal sparrow]

Hi. Relatively new to SD. Have experimented with A1111, ComfyUI, and ForgeUI and using the latter most of the time. Awhile back I did some images on a free AI site and I'm trying to reproduce the style and can't even come close. I've tried the most popular models/checkpoints on CivitAI, starting with SD and moving to SDXL and also Flux. I've tried addd Loras and using styles as well. With much experimentation I can get digital art and photorealism but I can't figure out how to do the style of the image below. The prompt was simply Lara Croft (i think this is SOTTR) wearing an Indiana Jones hat. I've seen many checkpoints with similar art, downloaded them, but nope...just can't get it.

Anyway could provide insights on a suitable model with or without Loras (and/or style for the prompt) It would be greatly appreciated. Also if this is wrong channel or wrong server, if you could point me somewhere more appropriate. Many thanks.

[verbal sparrow]

ok this was a total fucking waste of time as usual.

[crisp stream]

what was a waste of time? There isn´t any image attached btw

[clever oar]

he mean clowshark image

no model or lore comes close to his level realism

[crisp stream]

yet what is he referring to? And are you even sure because he said the prompt was Lara Croft and that doesn´t really look like Lara 😄

[clever oar]

he said image below

[crisp stream]

how could he lnow if it wasn´t already posted by someone else?

[clever oar]

only one image below -clownshark image😆

[crisp stream]

🙂

[clever oar]

good question but he can dont know how😄

[crisp stream]

still not going with that theory of saying he ment that image anyway at all 😄

[clever oar]

it's just Lara Croft but in a regular dress🙂

clown why can't you help this guy get the picture right level?

[clever oar]

medieval anatomy textbook?

[crisp stream]

yep

[clever oar]

umbrela shinobi

[clever oar]

strange flux cant make hybrid

hybrid of human and shark standing on a street

[clever oar]

its shark(

where my hybrid

[clever oar]

Soviet chronicle 40🙂

[clever oar]

my ipadapter now work out of memory

before it works without problem

sad

[crisp stream]

reboot already done?

[clever oar]

yep(

[crisp stream]

hmmm

[clever oar]

low vram xformers 256 256 try(

nothing work

[crisp stream]

not familiar with it

[clever oar]

i play with audio ai last time

and install cuda

maybe change torch version

[clever oar]

🇲 🇴 🇴 🇫 ℹ️

[royal monolith]

try "humanoid sharkman standing on a street"

[royal monolith]

now you need to create a video and tell his story

[celest sigil]

Some nice illustrations...

[steep sleet]

Quadrober?

[steep sleet]

Flux_GGUF_Q8

She's got LEDs in her pants!!! 🥳

[celest sigil]

Pre-warmer has been activated

[celest sigil]

https://tenor.com/view/rain-gene-kelly-singing-in-the-rain-raining-trendizisst-gif-18130024

[unique trail]

#🌠｜show-and-tell
Under a clear sky, there are Buddhist grottoes built along the mountain in the valley.

[steep sleet]

Flux_GGUF_Q8 + Magritte LoRA

[unique trail]

#artisan-faq
Under a clear sky, there are Buddhist grottoes built along the mountain in the valley.

[steep sleet]

Magritte LoRA in Flux_GGUF

[steep sleet]

Magritte-y

[steep sleet]

"Make me one with everything!"

[celest sigil]

I wouldn't eat that if I were you

She's contemplating it!

[fierce shuttle]

Do not use any robotic GPT! Order only human written essays and homeworks -> https://extraessay.com/?key_wpg=5wpgrd

[dawn seal]

Those are HOT pants!

[pale steeple]

getting an error when i try to run cogvideo

anyone have any ideea why?

when the cogvideo model was installing it free so i had to restart the program

now it is not even downloading the model anymore

it is giving me this erro

any help pls?

[steep sleet]

Have you downloaded https://github.com/Clybius/ComfyUI-Extra-Samplers.git

[clever oar]

yes

[pale steeple]

no, how do i download this? github is always fkin confunsing to me

[steep sleet]

The green button

Flux_GGUF and Magritte LoRA

[pale steeple]

and where do i put the files?

[steep sleet]

Custom_nodes folder in ComfyUI

[pale steeple]

cheers

it is freezing at this spot when downlaoding the model

[steep sleet]

Server is busy - or your disc has too little space?

[pale steeple]

is there any place where i can download this model?

[steep sleet]

What is the name of the model please?

[vagrant dust]

He is trying to get cogx Video the 5b model (image to video). The download works fine but these files are large...

[pale steeple]

how do i download it?

[pale steeple]

so how do i do it?

how do i download the model?

[vagrant dust]

your screenshot form above showed that you already downloaded 7% of it. While in total you would need 14 files.

[vagrant dust]

at the end you should have a new folder in the models directory called CogVideo. Within you find depending on the model a folder called for example: CogVideoX-5b-I2V
All files from the huggingface and folders are copied (Transformers, VAE, ...)

[ionic vault]

did u find it?

[clever oar]

What do you think the 4060 ti is worth the money? or the difference will be insignificant and you need to take 3060 ti

[midnight bough]

im going to violate you

https://tenor.com/view/star-trek-cigarettes-the-orville-bortus-500-cigarettes-gif-12658448946227274013

[clever oar]

ok do it

[midnight bough]

https://tenor.com/view/ratcockroach-gif-9933737926074927923

[clever oar]

you rat?

[midnight bough]

https://tenor.com/view/mrbeast-patting-tummy-stomach-patting-stomach-full-gif-27423087

[clever oar]

you need sleep

[midnight bough]

https://media.discordapp.net/attachments/873729689049518120/1132863148211314699/ezgif.com-video-to-gif.gif

[clever oar]

💀

[midnight bough]

https://media.discordapp.net/attachments/854559754331357184/1146937748163399690/ezgif-5-b7daa74b5a.gif

[languid pebble]

3090 has more VRAM ... at first it's about VRAM ... second point is how long it takes to calculate a picture ....

[midnight bough]

how much vram can i shove up my ass

[languid pebble]

420GB

[clever oar]

no 3090
3060 ti and 4060 ti

3090 expensive

[languid pebble]

VRAM let you work the modell .... the rest is about ir's speed ...

[clever oar]

i know but new expensive for me

3060 ti and 4060 ti big different for ai?

[languid pebble]

Do whatever you want ... I am just sharing ideas ....

[clever oar]

hm i need wait

when 3090 it will be cheaper

[languid pebble]

I don't know all the graphic cards ... I have a 4090 ... so I don't know the problem with lower ones ...

[clever oar]

ok

your trump shut up the troll 😎

[languid pebble]

And I don't really know what you need ... there are smaller models of flux working with less VRAM ... it's just my idea not to trust to them cause future models might become bigger ...

I don't think it's a good idea to invest in yesterdays technology ...

[clever oar]

in 3090 i can run flux shnell?

[languid pebble]

I am pretty sure

[scenic yew]

you can run flux in 4gb vram if you can wait, but 8gb vram is needed for a moderate speed.

[clever oar]

i need use sd 3 ,flux, sdxl models without long wait

4060 ti give me this ?

[scenic yew]

Just use schnell with q8 quantization, it should fit in 16gb vram with a nice speed.

[languid pebble]

That's 2 questions ... can I run the model ... and how fast ...

[clever oar]

i have 4 gb i cant run sdxl)

[languid pebble]

There are about 40 different workflows for Flux ...

even with low VRAM

[scenic yew]

You can run sdxl too, its just going to be very slow. You have to offload on cpu ram which will make it far slower but fit and work at least.

[clever oar]

i try one time

but fail

one question more-can i use ram if i run out of vram

i have error CUDA out of memory. Tried to allocate 20 MiB

20 mib ...

[clever oar]

no its only for you

funy?

[clever oar]

good choice for trolling sd channel...

[peak sapphire]

If i have a reference and i only want something inside of the black how would i set that up in flux?

i use forge with stable diffusion

[clever oar]

dont understand do you need to color the letters?

[compact aspen]

Hi everyone, I'm new to stable diffusion. Even though I have python, git, pip, torch and all the settings installed, I still can't create the image. It always gives me this result. Can anyone help me?

[vagrant dust]

Hi there,
i would use the controlnet depth for flux to achieve such a result.
But you would need some abstract prompt otherwise it might look like my result where you simply could user photoshop or any other tool to use a mask on the letter and cut the base image with the same mask....

[vagrant dust]

Well two ideas for you as you are new.
First better chances to get a reply if you post your question in the #🤝｜tech-support channel and second add some usefull information like the workflow you want to use, models, settings, hardware, .... so you might at least find someone who can help you.

[peak sapphire]

@vagrant dust i did use control net and got nothing like this

[nimble mason]

dude is a legit flamethrower

[vagrant dust]

Well it seems that we used different kind of workflows. Here is a link to one that should work:
https://civitai.com/models/709352
But replace the Loader with the one from InstantX

[bleak plaza]

"Create a realistic portrait of Timothée Chalamet styled as Philip J. Fry from Futurama. Capture Fry's distinctive hairstyle and casual outfit, blending Timothée's features with Fry's playful expression. The background should be colorful and slightly futuristic, reflecting the essence of Futurama while maintaining the realism of a photograph."

#🏞｜general-with-images "Create a realistic portrait of Timothée Chalamet styled as Philip J. Fry from Futurama. Capture Fry's distinctive hairstyle and casual outfit, blending Timothée's features with Fry's playful expression. The background should be colorful and slightly futuristic, reflecting the essence of Futurama while maintaining the realism of a photograph."

"Create a realistic portrait of Timothée Chalamet styled as Philip J. Fry from Futurama. Capture Fry's distinctive hairstyle and casual outfit, blending Timothée's features with Fry's playful expression. The background should be colorful and slightly futuristic, reflecting the essence of Futurama while maintaining the realism of a photograph."
#🏞｜general-with-images

[brave pagoda]

We are a commercial company, and we currently have a paid project requirement. We need assistance with editing some images by replacing the heads of different people into specified positions. Could you please let us know if there are professionals available to help us with this task?

[viral frost]

It's too bad that Flux likes to put planets in front of clouds in scenes like this, because otherwise they are pretty nice.

[wispy nest]

I love the sense of size and scale

[viral frost]

Flux doesn't know what travertine terraces are, though.

"Dude, I just spent a fortune on this crazy new subwoofer. I can't wait to blow out my eardrums!"

[wispy nest]

LOL that really is a giant speaker

[pale steeple]

anyone know why i am getting this error in cogvideo?

i changed a bit of the size of the image

any clues?

[deft walrus]

Weird fingers is still a pain, even with Flux

[scenic yew]

Yeah flux is still a diffusion model which will always have issues like these, but flux is a big step ahead from other models.

[deft walrus]

Heck yeah, i've just been getting spoiled with the quality of the hands.

[scenic yew]

Me too lol, I think kolors was the best at hands before flux.

[deft walrus]

I just love flux and today was the first I've tried PuLID. It's pretty awesome.

[harsh dune]

Prompt? 💯

[tepid temple]

#artisan-1 green square

[steep sleet]

DecadeTW-AutoPrompt-LLM

[languid pebble]

Ups ... I dropped it again. Good morning!

[languid pebble]

Have you seen this? Does anyone have the Workflow? https://www.youtube.com/watch?v=3WViQ2pL6ks

[steep sleet]

DecadeTW Auto Prompt

[languid pebble]

Strange style ...

[steep sleet]

SuperFlux w/f - Its working for me 😄 (Copied from YT Tutorial by Olivio Sarkas)

[steep sleet]

I used SDXL Checkpoint = Artfusion XL Surreal

"Super-Flux_GGUF"

[languid pebble]

Looks good! I wonder whether Flux Dev will have a benefit too

[viral frost]

The workflow is in the image, so you can load it yourself in comyfui, but the prompt is:

beautiful detailed photograph of a glistening tall massive futuristic sci-fi technological device made of metal and plastic built over idyllic travertine terraces with rising steam, incredible dramatic lighting, clear water, majestic solitude, lone figure walking, alien planet in the sky

[steep sleet]

Super Flux

[viral frost]

I’ll check out his workflow when I get a chance. If it requires a lot of rigmarole, though, I might just stick with my simple flow. I like the ability to go to large sizes without much trouble.

[old cypress]

/好好

[languid pebble]

Super Flux Dev. 🙂

[viral frost]

From looking at the "super" workflow, it seems to me like the only slightly novel things are 1) using the beta schedule (which won't affect anything for someone using an adaptive sampler, as far as I can tell) and 2) doing the 1st stage generation in two sub-stages with a different noise seed for each. I don't think that is likely to do much, since it will just end up averaging the leftover noise, sort of creating a different seed. I think he'd do much better by upscaling to the second stage using a model rather than just using basic lanczos.

[wispy nest]

its just a form of Restart sampling

its good because Restart sampling is good, essentially

[spice pelican]

https://i.imgur.com/VFi7UNJ.png

[scenic yew]

Schnell gives much more different images then flux dev it seems, both these images are schnell compared to flux dev which the ppl above give pictures for.

[viral frost]

I’ve only used dev, but it doesn’t seem to have the image variety of SDXL. Though it’s a bit hard to compare, given that images take me about 4x longer to generate with flux, so it’s harder to compare equal size sample sets.

[wispy nest]

yeah its lower

[viral frost]

What makes it good? What’s the evidence? I’ve seen people do this sort of thing before and it always seemed a bit silly to me. The examples above seem “different, not better”.

[wispy nest]

the original paper has a bunch of benchmarks

[viral frost]

I’m a complete novice with AI training, but I thought about the training method that SD3 used (where they randomly chose between the base CLIP caption and an AI generated caption) and the lack of image diversity in FLUX and some other models. I think it would be interesting to try training with AI-generated captions, but set to three different detail settings. One would be simple, capturing just the basic aspects of the image, one would be moderate, and one would be “this image is worth 1000 words”. Hopefully this would help the model learn both how to follow prompts well and how to add unprompted details and variations to the image.

[viral frost]

I didn’t know there was a paper. I’ll try looking for it. Thanks!

[wispy nest]

this would be cool yeah, maybe a level which is just tags also

[scenic yew]

Yeah flux dev is distilled and most distilled models do seem to have lower diversity, llm prompting can help make it more creative. Might be because of the distilled guidance too, not sure.

flux schnell for some reason gives more creative outputs(even though its also distilled)

[nimble mason]

stochastic sampling can really help with the diversity issue

[rain gazelle]

nice

[clever oar]

thx 😎

you like sonic game?

[rain gazelle]

Well I like the style there ... I liked the original sonic games lol

[clever oar]

I completed almost all parts and new one on PC as well

[viral frost]

Isn't Flux incompatible with SDE samplers?

[nimble mason]

Nope

Its trickier to do but it can absolutely be done

I've got a whole repo full of flux sdes

https://github.com/ClownsharkBatwing/RES4LYF a lot of stuff is getting overhauled atm but if you want to see it in action real quick try SamplerRK with the rk type set to res_2a and eta to 0.5 with hard noise mode

Flux really benefits from sde sampling imo

[celest sigil]

My own workflow, not "SUPER" 🤣

[languid pebble]

Good morning coffee

[viral frost]

I’ll check it out, but I’m getting confused. According to what I read, SD3 and flux were supposed to be ODE on some fundamental level. Plus, RK samplers are ODE as well, thus why they are suited for the application.

[nimble mason]

Yeah that's what they said... Not quite entirely true

They need to be variance preserving that's the most important thing

[wispy nest]

you could strap an ancestral step onto an RK sampler

RK and DPM are not that different its mostly that DPM solves the linear part algebraically

[nimble mason]

In some ways they're the same

I actually implemented some of the dpmpp samplers as RK

Which is kinda wild

[edgy nebula]

does anyone know how to help with face swap quality in #🤝｜tech-support ?

[nimble mason]

Reactor uses insight face

The authors of insight face are paranoid babies and limited the resolution to 128x128 out of fear they would trigger some kind of deep fake apocalypse

So it will always look that way

[edgy nebula]

😭😭

[wispy nest]

you can write DPM in butcher's table I think yeah

[nimble mason]

Train a flux lora on a character and inpaint the face

It works way better

Yep that's what I did, got the table from the RES paper

[edgy nebula]

hmm sounds interesting i would love to try, but don't really yknow what that means xd

[edgy nebula]

is it veryy complicated ?

[wispy nest]

you can use a lora service like Fal or Civit

if you are willing to use comfyUI, they have a workflow in Impact pack called facedetailer to put your lora

so its just 2 steps really

[edgy nebula]

hmm

just had a look on yt

[edgy nebula]

and this looks crazy complicated aha

what about "fface fusion" have you guys heard of this before ?

[wispy nest]

its just a swapping app

its not gonna compare to a flux lora

[edgy nebula]

i see

[edgy nebula]

and this works with img-img?

[wispy nest]

yeah

[edgy nebula]

If you have the right toutorial on how to set this up, or know anyone that could help that would be amazing 💗😅

[nimble mason]

I think you can pay a couple bucks to train one on civitai

[hybrid acorn]

today is my 1st day trying flux and it's the best ai model i've used yet.

[clever oar]

portal game?

[languid pebble]

Yes ... I really like the humor ...

[clever oar]

I hope part 3 will have the same graphics

[languid pebble]

Will there be a part 3?

[clever oar]

i hope

[shut sinew]

Flux is great yeah

[shut sinew]

Valve doesnt 3rd games

[languid pebble]

Bad luck but maybe a good decision. Better no 3rd part than a bad one ...

[clever oar]

Gaben give me hl3 and new portal!

[celest sigil]

HL Alyx 2 would be nice

[clever oar]

I don't have VR for alyx

stable diffusion in vr multiplayer 🧐

generate pictures with friends

[languid pebble]

Let's have an ...

[ancient bronze]

anybody happen to know what this means?

[fair vapor]

You don't have the VAE that is named in the workflow you downloaded. So just choose the right vae.

[lucid meteor]

@royal monolith

[royal monolith]

? i'm not sure what sort of response you're looking for from me?

[wispy nest]

looks like a lora, if you want to make something like this

[languid pebble]

Good morning coffee

[wispy nest]

flux colours are so nice

[lucid meteor]

I am trying to recreate the style of the character portraits close to the top of the image

[wispy nest]

you could go lora

or have a go with IP adapter and control nets (much harder)

[lucid meteor]

Do you mean finding a specific lora?

[wispy nest]

no I mean training a fresh one

[lucid meteor]

Oh I haven’t done that before

I don’t have a stockpile of images like the ones I want unfortunately

[wispy nest]

you can do flux with just 1 image apparently

[lucid meteor]

Interesting

How are the results?

[celest sigil]

Would you join my cult? 😄

[wispy nest]

seemed good to my eye at least

I have some issues with flux (distillation, ret flow loss, positional embeddings) but it is still very impressive

learning a lora from 1 image is really cool

[inner grotto]

#🏞｜general-with-images

#🏞｜general-with-images 123

[sterile kiln]

Rimworld ❤️

[steep sleet]

SD3.5 anybody? https://huggingface.co/stabilityai/stable-diffusion-3.5-large

[pallid ruin]

Yoooo

[clever oar]

can i use online?

[pallid ruin]

Im downloading it rn

But 13,9 size.... 💀 my 3060 will suffer, I´ll have to offload to ram ... and maybe even to swap (and this is fp8, the original fp16 is like 16gb)

[steep sleet]

Colab

[clever oar]

hate colab)
tried to start but there were errors

[wispy nest]

mhh sd3.5

hands

[sterile kiln]

They used the meme "a woman laying on grass" as avatar 😂
This wont save them

[still rose]

That’s really nice what’s the prompt if u don’t mind sharing

[crimson mango]

upscaled SD3.5 with a very chaotic prompt

[languid pebble]

The image is a movie poster for the film "Black Forrest, The Real Story". It features ((Forrest Gump)) with an african skintone sitting on a wooden bench with a suitcase next to him. He is wearing a beige suit and a blue shirt, and has a serious expression on his face. The background is a light blue color, and the title of the film is written in white text at the top of the image. Above the man, there is a white feather. The overall design is simple and minimalistic

[still rose]

Flux?

[languid pebble]

Ich glaub das war Flux, habs aber auch mit SD3.5

[still rose]

I’m sorry I don’t understand German 😕

[languid pebble]

This is SD3.5 the other was Flux

[fleet goblet]

"I think that was Flux, but I also had it with SD3.5"

[still rose]

Thank you

[still rose]

Ah ok anatomy needs to be fixed

[fleet goblet]

Euclidian Knight

[languid pebble]

Not yet convinced by SD3.5

[still rose]

Which does text better so far SD 3.5 or Flux?

[fleet goblet]

They both use T5xxl as their encoder

[still rose]

I’m waiting for the community stuff for it

[languid pebble]

I still prefer Flux

[fleet goblet]

at this moment

3.5 should be better

should

[still rose]

I’ll say it defo has major potential specifically with community stuff

[languid pebble]

Haven't seen anything better, yet

[still rose]

I will try it but I’m waiting for the MacBook software to get updated

[fleet goblet]

Well I got the fp8 variant of 3.5 running

and i got llama 3.2 making prompts for it

[gaunt elm]

here are some SD3.5 images I made on replicate.com

[still rose]

I still need to know for to prompt LLM to make prompts for me XD

[still rose]

I really like the 3rd

I might need that prompt also

[gaunt elm]

photo from below of a standing, one hand on hip, seductive, Japanese girl, female cyborg, cyborg implants, metal choker, prosthetic arm, headband, defiant, (latex suit:1.2),( on top pastel wind breaker pvc:1), (long hair, red, blue, green, gradient hair:1.3), black lipstick, freckles, on chest small crop-top with brand logo "NEO" .
subsurface scattering, ambient occlusion, bloom, detailed background, extremely detailed, dark, soft light, cyberspace, (iridescent purple background), Soft focus, Soft lighting, anime, cyberpunk

[still rose]

Very much appreciated

[sterile kiln]

SD 3.5 =

https://tenor.com/view/dumping-trash-daniel-labelle-frustrated-garbage-trash-gif-27186872

[still rose]

Crazy

[fleet goblet]

3.5 isn't bad

[still rose]

I’ll be happy with using it with better anatomy the style variants are cool

[crimson mango]

it's not bad but the anatomy is still janky probably due to the censorship

face can be fixed with post processing, but there is a lot of AI "noise" in the image overall. I think for "standard AI" style images SD3.5 isn't very good and FLUX is still better

[still rose]

Flux is still very good but like I always say for these models I wanna see what stuff the community cooks up to improve it

[sterile kiln]

it tends to furnish well, where Flux will remain simplistic, but Flux doesn't bring bullshito.

[still rose]

True I have not tried recently but is Flux good with anime styles without Loras?

[gaunt elm]

yeh .. with some Loras and checkpoints SD3.5 might be pry good

[still rose]

Yep and that’s what I’m excited for

[crimson mango]

Yeah so far I think SD3.5 is better than FLUX for non simple images. Or images that are more surreal/dreamy

[languid pebble]

I think SD lost a lot of fans who prefer to create Flux stuff now ....

[crimson mango]

Okay this one is kind of fun

[random acorn]

I'm a bit surprised. Why is every image people are showing somehow better than what they picked as the promo image for the model? 🤣 her face is all fucked up, for a moment I thought the model was terrible, but the images you guys are showing are giving me hope!

[clever oar]

https://huggingface.co/spaces/stabilityai/stable-diffusion-3.5-large

[crisp stream]

well, thats actually quite a simple image, one person, one even background