🏞｜general-with-images

[quiet current]

(and additionaly also use controlnet canny/depth/etc)

[lament rain]

Hey all. Anyone interested in helping me with some generations?

I have been running some tests on CivitAI using some of the web models on there, though am not happy with the results. I unfortunately lack a PC capable of running SD locally, so CivitAI has been my best free-to-use option so far.

My goal is to re-create some images like the one attached here. As far as I know, this is a real, life-size prop animal and is not AI generated.

[atomic coral]

man danging in moon

[copper matrix]

🌛

[copper matrix]

#artisan-faq

[fleet goblet]

making liminal entities is always unusual and interesting

[viral frost]

Nothing like rolling up your skin and cooking some dogs on an open fire.

[meager kelp]

Draw a young man with white hair on a chair, next to a cold white cloth

[copper matrix]

👴 🪑 ⬜

[copper matrix]

#artisan-faq if you wanna generate in this discord

[wispy nest]

flowers attacking her

[wispy nest]

she tamed the angry flowers

[desert ocean]

railway and data

[pale steeple]

quickest wan workflow at the moment?

[wise turret]

Not the exact same but I mean the appearance

[violet solstice]

Hazme un logo divertido con colores púrpura blanco negro gris y blanco el nombre del negocio es The most virals products shop

[wispy nest]

Google gemini

Gemini

[wispy nest]

SD 1.5

[cold bramble]

how to move from this...

to this?

im trying my best with prompting but failing all times

[quiet current]

I don't understand .... what are you trying to do here ? your prompt is for asuna at the beach => you get asuna at the beach.

[shrewd peak]

please roast my lora of my dog (sdxl base image ) 🙂

[wispy nest]

IDK this is well above average SDXL lora so its hard to criticise it

[dawn seal]

Curious. Did you use "blue Police box" or "TARDIS" in your prompt?

[celest sigil]

You must be a detective! 😄

[shrewd peak]

sorry was away, are you talking about my lora ?

[wispy nest]

yeah its nice

[shrewd peak]

ah thanks 🙂 was not sure I was on the right path

[wispy nest]

ye this is way better than most SD 1.5 lora
the main thing to be aware of is to try not to overfit it

if you overfit it then it can start making the model less flexible, or have strange visual effects

but its a tricky subject there is not like one clear definition and metric for if something is overfit

[undone dock]

You might want to delete this one for two very specific reasons.

[wispy nest]

the social definitions of SFW versus NSFW are not very logical
by the usual social standards that is SFW

[undone dock]

When I think "SFW" I think "Rated PG" or "kid friendly"

[wispy nest]

oh I interpret NSFW too literally maybe then

[shrewd peak]

Yeah and there is so many parameters, the images, the training parameters, the base model, it is quite hard tbh

[shrewd peak]

Because i have another dog lora too 😜

[wispy nest]

LOL I like the cinematic lighting

yeah lora training has many param

tricky

[fleet goblet]

intel arc is slowly catching up to cuda

latest 2.7 pytorch nightly now includes triton

not working yet but still

[topaz brook]

Hi

[shrewd peak]

hey, nice one, which model are you using ?

[celest sigil]

Thanks! I'm constantly changing, but I think that may have been this one https://civitai.com/models/998740/shuttle-mixes

[halcyon grotto]

new gemini flash 2.0 EXP on aistudio

[steep flower]

idk if anyone can help, but idk how to get from this to this, i've been trying literally ever setting, and the left pic is as far as i can get

[undone dock]

Where did you get the pic on the right? Did it mention Loras? Did it mention using custom nodes, scripts, or extensions?

There's all sorts of utilities to get from left image to right

[rain gazelle]

change your dimensions.

that's probably the easiest advice

[undone dock]

This too, the image on the right is using 560x840 while your image is 512^2

[steep flower]

i got embededs, 1 lora, and 2 models

ahh ok ty for that, i'm new to this

[rain gazelle]

dang spam bots

[undone dock]

Hey @wispy spindle or @wheat girder, two bots hit all channels

[copper matrix]

a second bot has hit the server

[royal monolith]

@wheat girder this is another one to add to the bot and block

[wheat girder]

what left pic?

[undone dock]

It seems to be a spamming bot, it's copying alissa's message from above.

[royal monolith]

a little more details? are you trying to do a video or something else?

[copper matrix]

i think hes struggling with detail and pose i suppose

but he needs afterdetailer

maybe

maybe also a vae

used to get those grey wash images when i had none

[wispy nest]

looks like SD 1.5
you can make good image with SD 1.5 but there are a bunch of things you need its not easy

compared to SD 3.5 or Flux where you can just run basic workflow

for SD 1.5 you essentially need

1. FreeU or some other tool that deals with the same issue
2. one of the fancier guidance methods
3. one of the methods for dealing with CFG-burn effects

but it all needs manually tweaking
so I can only advise to use modern model like SD3 if you are still learning

[regal panther]

I tried doing some inpainting for a D&D map.
Does this look plausible for a dragon's breath attack?

[royal monolith]

yup. looks very good.

[regal panther]

thanks, I'm still trying to get it more snowy and perhaps a bit less chaotic though

unfortunately this is pretty slow. large image size involved. takes almost 5 minutes per image

[dawn seal]

If the dragon is aiming at a particular target, you would tend to get those smaller scorch marks.

If the dragon is just strafing, you will get very long, more narrow scorch marks, as he flies overhead and just shoots at everything in front of him.

[undone dock]

And don't forget arcs if the dragon lands somewhere and then breaths fire.

[teal plover]

Maintain a human face. Wear a yellow safety helmet with the letters "KFB" on it. Wear glasses. The person is 40 years old. The background is modern architecture. Clothing: a white T-shirt with a collar. engineer

[halcyon burrow]

A few pictures I've made recently. Working on making character art for some of the pathfinder table top role playing games that I'm in.

On the left is my creepy summoner and on the right is my pyro goblin alchemist.

[halcyon burrow]

I love how it's like, cartoony pony but the background is like an artistic painting. ^_^

[pure cedar]

A chibi-style character standing outside in the morning, looking at the ground covered in fallen cherry blossom petals. They have a thoughtful or slightly sad expression, holding a tiny petal in their hand. The background features a soft, pastel-colored spring garden with gentle sunlight filtering through the trees. The petals swirl around in a light breeze, creating a dreamy and poetic mood.

[viral frost]

These labels are hilarious!

[dry rover]

What do you mean with "the same issue"

[dry rover]

You can probably just save the image at a lower resolution, inpaint, then add it back to the picture

maybe downscale and then later upscale as well

[regal panther]

yeah. Probably the wrong toolset. anyway. I heard gemini has some really good automatic image modification thingies. maybe I can try that.

[exotic otter]

"A highly detailed 3D model of Enkidu, the wild man from The Epic of Gilgamesh, in a full-body pose. He has a muscular, untamed build, covered in wild hair and fur, with an intense and primal expression. His eyes are fierce, and his posture is strong, standing confidently in a wild, natural environment, surrounded by animals like bulls and wolves. Enkidu's body language conveys his raw power and connection to the natural world. The image should be ultra-realistic, with dramatic lighting emphasizing his wild nature. The model should be presented with no background (transparent PNG)."

[copper matrix]

Thats oddly specific lol, youd need a lora for enkidu

[undone dock]

🦹‍♂️ Here you go

[undone dock]

https://tenor.com/view/monkey-dancing-monkey-dance-gif-13114207

I did the best I could on such short notice, sorry.

[wispy nest]

the unet is imbalanced and the skip connections need balancing with the output of deeper blocks when the model is used

its just how it is

[dry rover]

First time I heard about that

Do you have anything to link to for this? Might be time to go try SD1.5 again

[wispy nest]

its a weird thing yeah
that the models have this
if you haven't read FreeU paper that is best place to start https://arxiv.org/abs/2309.11497

but skip connections get mentioned in a lot of other places

some methods exploit them for editing and stuff

[slate creek]

1.5 is prone to making artifacts which are basically bright spots. No way to really get rid of it since it's a flaw with how the VAE was trained.

[wispy nest]

lol yeah I remember when that went viral on reddit

[undone dock]

The bright spots and blue spots are both problems that occur when someone merges checkpoints and doesn't know what they're doing.

[wispy nest]

they are but there is a second strange issue

the vae can have "hotspots"

[undone dock]

The only VAE issues I've seen are JPEG pixelation issues (artifacts) if the VAE was badly trained.

Too many people train on lossy JPEGs not knowing what they're doing.

[sterile kiln]

@celest sigil

[ornate osprey]

sorry for bothering, i want to ask, so.. now i try to learn basic comfy ui, then i try to generate the image and iam confused why the result is so weird.. so i try use AUTOMATIC1111 with same model, prompt, cfg, resolution, and the render its good.. so why the render in comfyui so weird?? is my node is wrong? or what? need some help and advice haha. sorry if my english bad, and thank you

[undone dock]

For starters, A1111 & ComfyUI use different prompt weighting systems. Check this out for more info: https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb

[south temple]

Have we seen anything new in terms of video or is Hunyuan still the recomended king?

I never got great results off it.

[spring mist]

Wan2.1 is a new capable series of video models

[halcyon burrow]

Speaking of wan2.1, I just installed it today and made some of my first videos. Exciting!

[undone dock]

That's awesome, and gives me some hope for my future usage of wan2.1 What GPU do you have?

[halcyon burrow]

4070

It was easy to install and run and I'm smooth brained. Though i had to ask a few questions, poke gpt, and watch a video

It takes forever to generate videos though. Like 5-10 minutes to generate 1 second

[undone dock]

SD1.5, no bells & whistles. WIP 1080p base model. Too high weights killed my subject making this completely SFW by accident. (No Prompt included)

[royal monolith]

if you animated that, it might look like a waterfall

[mighty wren]

an artwork that reads the name سحر in Arabic typesetting font

Final Gemini did it after a few tries. a correct name in Arabic

[royal monolith]

gemini is google, and with google translate available, i would expect it would do the best job

[mighty wren]

do Google Translate have Arabic vision glyphs? I would say Google fonts more precisely than Google translate but still It should have parameterised them in the model and any font page can render Arabic text, this is not something extraordinary.

[royal monolith]

give me a phrase

[undone dock]

The only words I know, you can blame 4chan for. "halal" and "haram" (they're not curses)

[mighty wren]

Well that font is everywhere, the arial or the arabic typesetting standard font but still google produces other fonts:

I gave it 10 images of the name with different fonts, look what It produced:

[royal monolith]

to me, since i can't read that - it's pretty, squiggle lines

i have hard enough time reading latin letters and english words

[mighty wren]

still correct Arabically

[royal monolith]

i'd love to have a language talent, but my brain just doesn't do that

[mighty wren]

language is culture and culture is language and power is language and language is power

[royal monolith]

i mean, i can't even count to 5 in arabic - forget reading anything more complicated

[undone dock]

If you ever find yourself in need of it or wanting to learn, I recommend Pimsleur above all others for spoken language learning.

[royal monolith]

i'll try it. i failed duolingo spectaculary

[undone dock]

It's what the CIA uses to train their Analysts so it's highly recommended.

[mighty wren]

Fair. Arabic is even hard for me native speaker as It has many forms but thankfully, I was attached to it since my childhood so I know a lot of under the hood rulings and uncommon classic words and how to make Arabic poetry

[royal monolith]

always best to teach kids multiple languages when little if you can. that's when the human brain is most tuned to learn languages

[mighty wren]

Actually AI is getting better when talking about Claude Arabic performance, It needs only better vocalisation, grammar and correct typography in terms of image generation

[royal monolith]

but it's an LLM, not an AI image or video generative model

[mighty wren]

yeah and I don't think AI will break language barrier as culture is very dynamic thing

[royal monolith]

all this technology we're running around making images and videos with is computer vision - that's been around a long time - and wasn't really developed for what we use it for. if you haven't researched that yet, you should

[mighty wren]

developed for other things? non-cultural usages?

[royal monolith]

sure :) medical stuff comes to mind

https://youtu.be/2w8XIskzdFw?feature=shared this is a good spot to start

[mighty wren]

maybe but before Chatgpt there was an AI called Spleeter for music separation and I think AI has been directed in to this evolution from the start by looking at the tools I were using before 2018

[royal monolith]

we've had AI since 1957 - and computer vision has been being worked on for decades

[ornate osprey]

sorry for bothering.. do u guys know why segment anything not showing?. so i just installed comfyui layer style.. and then i try to search segment anything and is not appears?

[celest sigil]

There's nothing in that node pack called "segment". Did you install "Segment anything" pack?

This is what layer style has for segs

[ornate osprey]

umm i dont know actually.. but after install the layer style is this included right isn't it?

[celest sigil]

Try searching for that name, and not "segment".

[ornate osprey]

umm.. still not appears

[hearty violet]

hi guys, I will do today a presentation about sd to generate architectural images for a university

it's possible I keep negative prompt on some way on seetings?

for example, I cant generate NSFW images, can I put it on some way on settings or only on negative prompt?

I wanna avoid show this words

[celest sigil]

Check startup in terminal to see if they failed to load.

[quiet current]

or you could install a NSFW censorship extension eg : https://github.com/AUTOMATIC1111/stable-diffusion-webui-nsfw-censor

[hearty violet]

thanks ^^

[quiet current]

You could also use/create an embedding/style if your intention is simply to hide nsfw words from showing during the presentation. Or both the "nsfw" negative embedding and the extension. That way you re minimizing the chance of getting something nsfw AND censoring it if you still end up with some ranchy output.

[outer violet]

"Rainforest terrarium without water, Portland cement rock formations "Hyper-realistic geological formation with volcanic rock layers, f/8 depth of field (foreground branches at 1/2 meter), dual soft lighting (5000K main/4000K fill), weathered texture (Ra 3μm surface roughness), moss-covered areas (30% coverage), dry branches system (crown radius 8cm), environmental fog (30% concentration), 16K resolution with 32 samples/pixel, photorealistic style (75%) + illustration enhancement (25%), CRI 95 lighting, natural color balance (4500K), volumetric dust effect, rock layer thickness gradient (40-60cm), Fibonacci spiral composition, subsurface scattering on moss (#7CB342), environmental color (#1A1A1A) backdrop"
Require higher details: promotion"sample rate 64 samples/pixel" + " texture detail 2048×2048"

[hearty violet]

the link of my presentation
https://www.youtube.com/watch?v=eAHk1r5zFQA

Guys, this was a presentation about SD to a Architecture course here in Brazil. The presentation is on portuguese but I thanked the SD discord server and the whole community for their help and company 🙂
I left the link to the server as well and invited people to join in

[quiet current]

Dang that was fast, you were asking bout the censorship only yesterday. I'll try to watch it later on, my portuguese is pretty thin tho :p so I might have to turn on subtitles. Skimming through I see that you're showcasing txt2img, basic img2img, controlnet, styletransfer. That's a lot of stuff to explain in a single course.

[hearty violet]

thanks ^^ At last I don't show nothing NSFW 🙏 Was not a course, was a presentation of AI to generate images for architecture for the university. I introduced SD and tried to show a bit of the whole picture

[clear cedar]

generate a really funny picture of a friendly and happy ghost that is sitting in a chair while being interviewed for an executive job

[undone dock]

I'm surprised you went with that, I would have just done a GIF of Ghostface from the Scream parody "Scary Movie"

[undone dock]

https://tenor.com/view/whats-up-scary-scary-movie-scream-gif-16966640

[restive cloak]

lol

[clever stag]

For flux, could you guys reccomend better Ultimate SD setings?

[royal monolith]

i wouldn't use that sampler with flux

[clever stag]

Thanks, could you reccomend something else?

[royal monolith]

note that the the sweet spot for flux is 672x1024 or 1024x672 - so you'll get the best look to your images if you stick with those for the image sizes. try huenpp for the sampler and simple for the scheduler

[clever stag]

My input image is 512x1024, so I'm just squeezing under, just trying with your suggestions now 👍

[royal monolith]

i know, but there's a reason for the specific numbers i suggested

[clever stag]

I'll try and find a doc on why that is, I need to have the base 2 to compress these images properly.

[royal monolith]

the reason it is - when flux first released, a number of us did some indepth testing. that's the ratio that works the best. we didn't bother to write up what we found so you won't find a doc on it

but it has to do with the size of the images in teh data set it trained on

you can always crop the images before you compress them if you need to

[royal monolith]

here's my flux latent space walkthrough https://docs.google.com/spreadsheets/d/1VDt_9U59QXwXUhFP0T32HhMg6_Iem76aW4C2u_B3CsI/edit?usp=sharing you can see clearly how it reacts to words and what it knows

[clever stag]

I fully understand the trained resolutions, I just assumed you meant flux was different than any of the other generative models, testing with 672 to see if it changes much.
As for the heupp2 and simple, it takes longer and is almost identical (it doesn't look better) to the previos settings I have, so I'll keep mucking about with these.

Cropping or rescaling will break the aspects I need.

For instance this is dirt that is good, and the upscale weird "dirt"

[royal monolith]

i'm sorry. i will point out this is the stability.AI discord - and that SD 3.5 large is much much better than flux in all respects - and suggest perhaps you switch to it

[clever stag]

I've used 2.5 and found flux d to be way beter, I have no problem trying it out, are you sure its way better? 😄

[royal monolith]

not sd3-2b-medium. sd 3.5 large

and yes, i'm certain it's better. by miles

[clever stag]

I'll give it a try then

Thanks!

[royal monolith]

let me know if you get stuck. remember - that with every AI you first want to learn how it thinks - so one or two word prompts, generate multiple times, study the results

[clever stag]

Hate to ask here 😄

But anyone have a good workflow for 3.2 I can test out, I just searched and nothing good came up.

Also, the two differences between 512 and 672, they are roughly the same in terms of detail,

[royal monolith]

it's 3.5 not 3.2 and sure.

this one allows you to prompt the 3 encoders seperately

[clever stag]

Thanks mate, greatly appreciate your help, this may have solved my entire issue!

[royal monolith]

they each have their own strengths and ifyou give them all the same prompt, they'll battle with each other

[royal monolith]

and this uses sd3.5 large as the main model with sd 3.5 medium doing the upscaling

[clever stag]

That was a slog to get HF to work with me and the token.

For me it's hard to decide what one's better, as they both have their pros.

But I will say, the encoding is 3x better

[tiny ingot]

anybody have any luck generating pixar-like images?

[undone dock]

3D Animation Diffusion is what you're looking for.

[tiny ingot]

oh seeing some nice results, cool ty

[coral pond]

Can someone help me find a yolo.pt file for detecting full body. I am really struggling to find it 😦

[alpine crag]

"A retro sailboat docked at a small beach pier at sunrise, with coconut trees swaying in the breeze, gentle waves lapping at the shore, and a large, glowing sun rising over the horizon, in a 1950s-inspired illustrative style with a serene, pastel color scheme and sharp silhouettes.

[dry rover]

Just tried it, is pretty nice

[steep edge]

A picture showing Two main types of bone include spongy (trabecular or cancellous) and compact (cortical) bone. osteons in compact bone and trabeculae in spongy bone. Figures showing osteons in compact bone and trabeculae in spongy bone, including osteogenic cells, osteoblasts, osteocytes, and osteoclasts and blood vessels.

[pale steeple]

can anyone give me a simple wan workflow with loras? please? i beg you ! 😊

[clever stag]

Give me the model requirements and I'll try and make one.

Is it on HF?

[pale steeple]

i have 4090

but would like under 5 minutes for a generation with lora

[clever stag]

No I mean the specific modules / models.

Is there a HF repo?

[pale steeple]

loras and model you mean?

[clever stag]

Yeah what you'd need to have it runing on comfy

[pale steeple]

wanAIWan21VideoModelSafetensors_comfyorgI2v480pGGUF

BouncyWalkV01

[clever stag]

2.1 example

[pale steeple]

thanks, where do i load the loras 🙂

[clever stag]

Just add a load lora node in between the text encode:

hook up the model and the clip

[clever stag]

Do you mind if I DM you?

[pale steeple]

kinda

i am not dming sorry

but do appreciate the help

i am trying to

load the loar

so basically a third node from the clip to lora and i connect the model node to the base model and lora yes?

[clever stag]

you'll figure it out 👍

[pale steeple]

thanks bro all my hearts ❤️

[primal iron]

designing textures for video games might be easy nowadays with AI

[clever stag]

?

[copper matrix]

Hmm not really, it might help but youd still need to edit a lot

Using it for inspiration sure

But without the base textures like metal etc its gonna look out of place

[primal iron]

ohh... mmm...

[pale steeple]

any good prompts for wan making a character turn around from i2v? thanks!

or loras

[royal monolith]

i2v is almost impossible to get any AI to do a turn around with - you'll have better success if you start with an image of a character partly turned but it'll still likely rotate the character to be looking directly at the camera

[undone dock]

I could forward one, but it's NSFW. I have seen some that were more than successful.

I just don't know what combination of tools they used to make it.

[wispy nest]

Stable Virtual Camera and SV4D 2.0

both launched this week lol

[spring mist]

a little extreme but there is this https://civitai.com/models/1346623/360-degree-rotation-microwave-rotation-wan21-i2v-lora

[empty topaz]

Can someone point me towards a good tutorial to get the best possible likeness while training my own likeness lora?

SDXL

[wispy nest]

no one ever takes this advice but my advise is to use Stepfun even if you have to block-swap it constantly

its an upgrade to wan

but you need to take care your power consumption does not exceed what would be worth it

if your vram is below a certain amount then the amount of block-swapping would make this the case

but if you have 24GB vram its probably fine

[wispy nest]

these were my personal notes: rank and batch size as high as hardware allows alpha = grid search from alpha = 0.5*rank to 2*rank learning rate = grid search from 1e-6 to 1e-3 weight decay = grid search from 0 to 1e-4, possibly up to 1e-2 gradient clipping = grid search from 1 - 2, possibly up to 5 optimiser = adamW schedule = cosine warmup = 2.5% to 10% of steps

grid search just means try numbers in that range

[sweet crane]

hi im just wondering what style and prompt strucutre is used to create images in this style? thanks

[wispy nest]

Picture of a jet

[undone dock]

✈️

[fleet goblet]

video of an idiot

https://tenor.com/view/dumb-huh-stupid-idiot-gif-15033206007095816163

dude even looked up to read the text

[copper matrix]

i mean a lot of people use google translate in here

[copper matrix]

i just point them to #artisan-faq

[true star]

Can 4060 run this?🙏

[copper matrix]

Ti? Maybe, normal ver has 8gb vram so hmmm either out of memory errors or its gonna take a very, very long time

Hm 5080 takes 10-30min

[true star]

I can proceed it before I go to bed, afterall AI thing is lucky draw lol
Unless there is no out-of-error I don’t mind to wait but at least allow me to wait lol

You know, waiting isn’t a matter, yet out-of-memory-halt is the real hidden devil

You know, running hub shit can project a preview of the prompt for quick and dirty via this https://www.runninghub.ai/ai-detail/1901651290458316801

If I can wait and something sexy, then I can get this shit

[pale steeple]

what is most used and most compatible wan checkpoint for loras

480p? 720? gguf?

I am not good with these choices

[dim tiger]

hello i would like to generate images in this style of drawing with stable diffusion could someone please help me when i do it it makes an image that has nothing to do ?

[buoyant bane]

"Aesthetic wallpaper inspired by Oriental Five Elements, gold and wood fusion, soft green leaves with intricate designs, shimmering golden branches, interplay of emerald and gold, tension of balance, delicate mist, hidden golden glimmers, minimalist abstract, ideal for mobile, 4k"

[willow walrus]

Design a vibrant and eye-catching Spotify playlist cover featuring a dynamic collage of musical elements. In the center, a vintage vinyl record spins, surrounded by colorful musical notes and abstract sound waves. The background is a gradient of electric blues and purples, evoking a sense of rhythm and energy. Scattered around are small illustrations of various musical instruments like a guitar, keyboard, and drums, adding a playful touch. The overall mood is lively and energetic, capturing the essence of diverse music genres coming together in harmony.

[dry blaze]

Curious as to why checkpoints/loras always transfers so poorly

And this is from one nvme to another.

[undone dock]

Because file transfer for single files is high, but you're transferring 1,554 of them. Cropping won't save you here.

[feral lagoon]

Is there a way to use AI to digitize a logo from a photo?

[dry blaze]

Yes, but each are 300MB. So it should easily be few 100MB's at the minimum

And cropping?

It's 150GB worth of loras, so it should be quite high, even for nvme's at sequentials. It'd be way slower if it was few 100KB per file :P

[zealous mulch]

what am i doing wrong?

i tried just a normal img2img enhancement

[dim salmon]

#🏞｜general-with-images pixar style

[iron solar]

anyone knows if this image was generated usingstable diffusion?

[wanton nacelle]

Anything

[restive trench]

@feral lagoon Of course. Use the IPAdapter style transfer.

[iron solar]

DM me

[feral lagoon]

And this can basically just take the logo in the photo and remake the same logo digitally? Is there a good guide you know of for this?

[celest sigil]

Just look at the hands 😄

[iron solar]

Every AI messes up hands, that's not really giving it up

I'm trying to create images similar to that one, but can't find the right AI for that

[shrewd peak]

MF

guys

it have been a rollercoast tonight

won a bid on ebay for a 3090

(bear with me)

received it today ! total surprised ! (the mail is shit in the UK you pay for next day delivery, you are happy if you got it the week after)

anyway

got the card

got a minute for my previous card that serves well

ope the egpu (x chroma) ( yeah i got a laptop)

clean it quickly, all good

try to put the new card in

try again...... try again.... do NOT FIT .....

I repeat do not fit !!!!

oh god depressed

google it, the fu%$£%£$ing bracket too large

had to use a bloody plier on my enclosure to make it fit, I was thinking I got no fucking idea what I am doing, I might just have destroyed the egpu and the card trying to make it fit

but guess what

card recognized and soon you gonna have more picture of my dogs !

sorry for the rant, it just happened and I had to say it so someone

[shrewd peak]

on this, gonna mind my own business
https://www.youtube.com/watch?v=Enf2P1_Mofs

[pearl coral]

How many steps?

[zealous mulch]

20

[pearl coral]

Denoise?

[zealous mulch]

0.7

[zealous mulch]

i reinstalled stable diffusion now because

first time i installed it my interent was cutting off often so idk if there may have been corrupted files

i will try again

[restive trench]

@feral lagoon Inside Comfy Manager search for IPAdapter. Click on the github link and scroll down. There is how-to video on the page.

[pearl coral]

That's likely too high denoise. Try 0.1 and increment it gradually with successive gens until you get a better result. If that doesn't do it, then I don't know, I don't use i2i that often.

[zealous mulch]

will try thanks!

[copper matrix]

@slate drift leave that discord
Its spam/scam

[undone dock]

I'd recognize baby yoda's influence anywhere, those eyes and ears...

[sterile kiln]

what model is it?

[shrewd peak]

which model is that ?

[wispy nest]

woah

cool

comfyui?

[viral frost]

Thanks. Yep, Flux.

[wispy nest]

do you by any chance know what Euler a is?

is it a name for Eular_ancestral?

[tired spindle]

Lovely!

[viral frost]

Yes.

[sleek ore]

good. anime waifus this is the type of thing we all love to see. Waifudiffusion has a special place in my heart

[undone dock]

WaifuDiffusion is ancient now and garbage. Download Aniverse v4.0 and thank me later.

[hearty violet]

hi, what must I choose to have a fast GPU to SD? Some of this that I pick starting SG with 15 minutes, other with 45... What must I observe?

[wispy nest]

DLPerf

and try to get DLPerf at least 30

[hearty violet]

thanks Neon

what is this links above?

[wispy nest]

bots

[hearty violet]

I pick one with DLP 30, take 8 mints to start 👍

[wispy nest]

for cloud you need fancy storage really

some thing like backblaze b2

[hearty violet]

I din't know about this

I can build my files on blackblaze and just start a instance on vast when I need to use SD? Is for something like it?

[hearty violet]

I'm tryng the diffetents preprocessor of lineart controlnet. I know in theory the differentes between then, but I don't se difference in the images generated

Img2Img for Controlnet Lineart

Preprocessor Realistic

Preprocessor Coarse

Preprocessor Anime Denoise

Preprocessor Anime

they are pratically randon, I don't see the patterns according to the preprocessors

[wispy nest]

yeah its fast download

[vagrant dust]

The preprocessing is used to determine the lines from the input image. So if you want to see a difference compare a realistic image and a anime image as source with different preprocessing models. Not your accident on a white paper 🙂

[hearty violet]

Now I understand, I will experiment this! 📚

[iron solar]

https://media.discordapp.net/attachments/1190323386828738761/1355226678154235934/maxresdefault.png?ex=67e828b2&is=67e6d732&hm=8997a757835c66bb1fa2fe678ad4432ed7ac9ee41bb4f3afd56e5858f4d2baae&=&format=webp&quality=lossless&width=1423&height=800

[copper matrix]

Flux dev or sd3.5 most likely considering the coherence

[iron solar]

tried sd3.5 but it doesnt really generaet what i want

[copper matrix]

I mean that could be a matter of the prompt

Maybe controlnet

[iron solar]

probably

[copper matrix]

Yeah probably controlnet was involved

[hearty violet]

@vagrant dust I made a study

Original images

Prompts:
1° - exterior architecture of a school by walter gropius, wood, glass, steel, realistic image
2° External scene of a biomechanical castle in a landscape of trees, with a lady in front of it

Realistic

Coarse

Anime Denoise

Anime

I understand the the preprocessor must be matched with the quality of the line image. Do must be more aproximaty from the controlmap, maybe?
But in pratical I don't see a real difference, everything looks randon
What I'm not understanding?

[sterile kiln]

hey guys,
Hasn't anyone noticed a big change in Comfy's performance lately?

I have the impression that it consumes more ram/Vram than before, or that it doesn't organize the vram><CPU/Ram cache as it used to.

It's having problems now, and makes me do virtual ram (pagefile.sys) whereas before it didn't... I use the same workflow, the same config, the same models... Nothing has changed on my end.
3080Ti, 32gigs, 3700x, 970 Evo+

@grand walrus ?

[sleek ore]

oh bro i use aniverse too im just reminiscin'

[undone dock]

I understand completely. I still think about F111, F222, the Abyss Orange Mixes, etc.

[delicate solstice]

"Turn this image into a Studio Ghibli-style animated portrait. Use the soft color palette, whimsical background, and facial features inspired by Ghibli characters. Style it like a scene from 'My Neighbor Totoro' or 'Spirited Away'

[quick yacht]

"Turn this image into a Studio Ghibli-style animated portrait. Use the soft color palette, whimsical background, and facial features inspired by Ghibli characters. Style it like a scene from 'My Neighbor Totoro' or 'Spirited Away'

[neon saddle]

I understand the the preprocessor must be matched with the quality of the line image. Do must be more aproximaty from the controlmap, maybe? But in pratical I don't see a real difference, everything looks randon What I'm not understanding?
No data source is currently selected. Please choose a data source from the dashboard and try again.

[still rose]

Guys u think it’s a good idea getting a 5070?

[timber tree]

Draw peaceful village with ghibli style

[copper matrix]

i think your looking for chatGPT's image converter

[timber tree]

Yeah!! Can you help me ?

[copper matrix]

probaby? its on chatgpt

i dont have an openAI account though

[timber tree]

Is there any option in dc to generate a pic to ghibli style ?

[cinder wasp]

heyy everyone where can i make request? i cannot see any channel for it

[reef shoal]

Hey guys. I'm making some static RPG backgrounds and I've generated a lot of images like:
I want to use outpainting (with stablediffusion webui) to extend these backgrounds and make full scenes.
Is this possible? Have any of you done something like this before? Any model recommendations?

[hearty violet]

yes, it's possible. I have did it a long time, I don't remember but the other users can help you 😉

Obs: you took the consistency compositions of characters very well

[sterile kiln]

PM

[halcyon burrow]

Image on the right is with inpaint using the settings you posted.

The teeth are kinda better, the eyes are fine, just a little looking up too much, but the skin color is horrible.

[dry crow]

yea its to bright in this case

[halcyon burrow]

Happens to me a lot with inpaint

[dry crow]

have you tried to lower the denois value?

to like 0.2 or 0.3 ?

[halcyon burrow]

I'll try now.

It's not as bad, but still not the right color

That was with 3 denois

[dry crow]

yea much better

[halcyon burrow]

With 2 it's better, but you can see a clear color change around the mouth

I guess I could try to paint only the mouth and eyes

Here's with adetailer

also bad

[dry crow]

you can select the eye only model in adetailer

[halcyon burrow]

I'll post the eye only, sec

Oof, eye only messed up the whole face

[dry crow]

are you in txt2img or img2img?

[halcyon burrow]

img2img

[dry crow]

yea there adetailer isnt as good to use

[halcyon burrow]

I'll try in text2img

[dry crow]

i tried a test with Adetailer in txt2img with its default settings

[halcyon burrow]

Yeaaa... that's like, not really an improvement

Looks about the same.

Little sharper?

But I kinda like the first image more

[dry crow]

yep mostly sharper

it behaves different with the models

[halcyon burrow]

It's odd, when I look at the two image you posted I like the right one better, but when I click on them and zoom in I like the left one better

This was eye only in text to image

But I don't know what it looked like before, so I don't know if it's better or not

[dry crow]

but the eye looks good

[halcyon burrow]

Yeah, not bad.

Gonna try something, I have a hunch, maybe if I remove "freckles" from my prompt it will work better. I think it might have an issue with red freckles on green skin. Probably confusing the AI with those colors mixed on a face.

[dry crow]

yea adetailer aswell as inpainting mostly smooth out freckles or blush, and they sometimes struggle with colored skin

clean faces work good

[halcyon burrow]

Yeah, I've been having a heck of a time with this goblin girl I've been working on for weeks lol. I had much better success with my last character.

[dry crow]

your making OCs ?

SCAM

[halcyon burrow]

yuuuup, there we go, when I removed freakles it was a LOT b etter

Yes, working on OC's

Playing a goblin alchemist in a pathfinder game and I wanted art

Phew, thanks for helping me through that, it's great to have someone to talk to about this stuff. I think the freckles might have been the issue. ^_^

[dry crow]

nice no problem 🙂

and yea dont bother with codeformer of gfpgan, they are outdated, and codeformer is for realism anyway

[halcyon burrow]

Maybe you can help me with another issue, that might not be an issue, but a limitation? I wanna generate images larger than 1024x but I think Illustrious doesn't like that. When I try to make an image larger than 1024 it destorts my characters legs and makes them lumpy

[dry crow]

illustrious is like pony models, based on SDXL.
SDXL models are trained on a 1024x1024 resolution.
So staying near that gives the best results but you can also go a bit larger like most people use illustrious at 832x1216.
Going to far above a trained resolution will result in duplicates and deformations.
So most users use upscaling to enlarge the image and also get a better quality. You can do that with the Hires fix for example

[halcyon burrow]

Yeah, I can hires 2x, but if I try to do more I get the lumps again

[dry crow]

yep, make sure the denois is not higher than 0.5

[halcyon burrow]

Yeah, I try to keep denois around 3-5

It's frustrating cus when I make an image that is say, 1280x1280, I can see how much better the face is rendered, it looks soooo good, but the body turns to lumps. Makes me sad >_<

[dry crow]

yea best is then to limit the width to a smaller value

so the person dont have space to deform xD

[halcyon burrow]

lol, yeah, I've noticed that in images where the character is closer to the screen and less of the body is rendered, like an image from the waist up, the image looks a lot more detailed. But this makes me sad too cus I like the whole body look.

I want my cake and I want to eat it too 😠

[dry crow]

yep exactly! full body is the hardest as you have less pixels for the face and hands

[halcyon burrow]

So basically, I need to give up on my dreams is what your saying? XD

[dry crow]

if you need perfect hands then maybe or use "hands behind back" as tag xD

for the face we have adetailer

[halcyon burrow]

I don't have too much issue with the hands, mostly the face.

Is it advicable to double or triple or even quadupal up on the adetailer? I see you can set it to do up to 4 masks at once.

Like say, one for the body, one for the face, and one for the eyes?

[dry crow]

thats useful yea if you want specific stuff upscaled:
there are for example a lot of community models for adetailer, for like glasses, furry faces, eyes, etc:
https://civitai.com/search/models?sortBy=models_v9&query=adetailer

[halcyon burrow]

Oh, here's another question, prompt related, drives me nuts, but any idea how to stop stable from generating symetrical backgrounds? Both sides always seem to be more or less a mirror of each other. I've tried prompts like "Asymetrical background" and stuff, but doesn't help much.

Ooh, interesting

[dry crow]

prompt for a background would be the easiest

instead of letting the ai decide

[halcyon burrow]

It's just hard, cus my prompts keep getting longer and longer and that's not good, right? XD

[dry crow]

yea to long is not that good

[halcyon burrow]

I;m tyring to keep it down to like, 240 tokens,

I often get up to 300 ish

It's so hard to keep it short, there's sooo much to describe just for the character's outfit

But thanks for the heads up on the adetailer models, that's awesome!

[dry crow]

no problem yea there are some nice ones 🙂

[halcyon burrow]

Hmm, do I have to use ones that say they're specifically for illustrious?

Like how you have to with loras?

[dry crow]

nope they work with every model

its just inpainting

the models are only for detecting stuff in the image

[halcyon burrow]

Oh cool, good. So they add onto adetailer and you just select it as a mask when you wanna generate in text to image?

nice nice.

This is great, making some more head way! Every week I grow stronger and stronger. ^_^

Thank you thank you. ^_^

Have you had much luck with things like mosaic or is that not worth the hassle?

[dry crow]

here is a good example of full body with adetailer off and on

left is on and right if off

thanks discord for switching them... xD

[halcyon burrow]

lol

Yeah, huh, that was the full body one? It looks like it only changed the face, ya?

It did a good job on the face though

[dry crow]

noe that was only the face

you dont need a body switcher for fully body

[halcyon burrow]

Ooh, you meant full body as in a full body image.

[dry crow]

yea sry xD

[halcyon burrow]

Gotcha, I thought you meant the whole body thingy in adetailer. gotcha gotcha

any cool extentions you recommend? I'm gonna look more into this adetailer model stuff when I have a moment.

I got the booru tag thing, that's super helpful.

Is the refiner worth using? I'm not sure if I can tell a difference with it on vs off.

[dry crow]

there is one very important extension. Its called Booru tag autocompletion.
As you use Illustrious models you need to know that these were like pony models trained on Booru Image board tags.
The extension suggest your these tags like a autocorrect and shows how common a tag got used. so you can prompt much better and more effectivly

[dry crow]

yea i should read before typing

[halcyon burrow]

Yeah, I've got that one, it's a big help. totally!

[dry crow]

nope its not recommended to use

it was made for sdxl in the early days

but hries fix is much less resource heave and works better

[halcyon burrow]

Okay, what about kohya hrfix integrated?

[dry crow]

that can be used like hires fix

[halcyon burrow]

I've tried it on vs off and I;m not sure if it's good or not.

Any scripts I should be using in the script box at the bottom?

[dry crow]

X/Y/Z script if you want to endless compare stuff with each other ^^ like samplers, steps, loras, checkpoins etc

[halcyon burrow]

I"ve heard about that one, not sure I need it though

Oh, here's a question that I have been struggling with for a while. Most of the illustrious models I've seen recommend using eular A for sampling method, but when I talk to chat gpt for advice and ask it what sampling method is best for fantasy character art it says hands down dpm++3m sde karras. I've tried both and it seems that dmp does look a little better, but I'm not 100% sure there's that huge of a difference and I'm curious why eular a is always recommended on the model description pages.

[dry crow]

euler a is cleaner

dpm++ samplers add more noise into the image

that can cause wierd issues with some models

but more noise means also more details

[halcyon burrow]

Hmm hmm. I see, I see. gpt says stuff like, dpm is sharper, has better lighting, and produces higher quality images?

[dry crow]

xD

nope

[halcyon burrow]

OKay, I'll stop trying to use it then XD, thanks.

gpt is often wrong. ^_^

[dry crow]

yea it doesnt have much clue about SD

[halcyon burrow]

What schedule type is best with euler a? I often use karras with it.

[dry crow]

karras is the best to use it with.
if you leave scheduler on auto it will always use the recommended ones for each samplers

[halcyon burrow]

I like the r-esrgan 4x anime6b for hires, that's solid, right?

For normal sampling steps, should I go crazy and do 100? XD I usually do 30, but I notice things do kinda look better with more steps, like 50 or 60 but not sure it's all that better.

Sorry, just basically asking a million questions lol. Sorry sorry.

[dry crow]

no problem yea the upscalers are okay to start but there are also some great community ones:
you can find them here:
https://openmodeldb.info/
For SDXL and illus i also use 30 steps its a perfect balance between speed and a good output

more steps doesnt mean better quality, the sampler then just tries to add something more, can be for example a new hair strain or a 6 finger

[halcyon burrow]

ya lol, that makes sense.

clicks link

Wow, a whole new rabit hole to get lost in lol

I have no idea where to begin

[dry crow]

UltraSharp for realism is good

[halcyon burrow]

Thanks, I'll look for that one.

I think I've heard of that one, is that for upscaling in the extra tab, or for hires?

[dry crow]

or UltraMix Balanced

[dry crow]

boath

[halcyon burrow]

takes notes

Ooh, I see. Okay, cool. So I can get something that works better then the anime one I currently use.

[dry crow]

Most Anime Users use Fatal_Anime

Remacri is also good and lolypop

[halcyon burrow]

Thanks. I'll look into that

[dry crow]

you can test with x/y/z and compare 😛

[halcyon burrow]

Sometimes I get some fairly good stuff, but often it doesn't render well for me. the struggle is real.

I've been struggling to figure out why sometimes I get a good image and most the time I get a bad image.

[dry crow]

looks awesome!

[halcyon burrow]

Thanks! I just cant seem to replicate the style consistently.

Like this.

It's not nearly as good

My best pictures seem to be with bad prompts that shouldn't work but for some reason do. Then when I try to fine tune my prompt, cut out bad tags, shorten things to be more concise and to the point, I get a worse image. I don't get it XD

I think one of the main differences between those two pictures is the second image is with the "full body" prompt to include the whole character, from foot to head, and so it had less pixels to render detail with for the face and such.

[dry crow]

yep thats it

[halcyon burrow]

Does that really make that big of a difference?

It's crazy

Like, even the lighting is way better in the first picture

Anyway, thanks so much. I need to go touch grass and get off the computer lol. You've been a massive help. Really appreciate it. Thanks again.

I have a lot of good new info to research now and a lot of helpful advice.

[dry crow]

no problem! yea there is a lot to learn and test with SD ^^ have fun!

[jaunty prawn]

cute goblin girl!

[halcyon burrow]

Thanks. ^_^

[mighty wren]

epic transformation

[naive cloud]

Can anyone please explain or tell me where to find a "how-to guide" for creating images and polishing my existing photo for use as a Facebook profile pic?

Noob here

[rain gazelle]

@halcyon burrow - consistency with your character looks good there. Which method are you using?

[halcyon burrow]

consistency with outfit and landscape, just not quality and lighting. I don't think I know what you are refering to when you ask for method? I'm still very new.

[rain gazelle]

Consistency, meaning your character looks the same across multiple images. Just detailed prompting? Or using ipadater, a Lora, or something else?

[halcyon burrow]

I have 6 loras in the prompt and the prompt is almost 300 tokens long.

idk what ipadater is.

But thanks! yeah, I'm happy that the character consistently looks the same. ^_^

I also told the prompt to base the character off of Tifa Lockhart, so that helps with consistency as well.

[halcyon burrow]

I use illustrious as well, if that's helpful to know. I really like the checkpoint model illustrij v8.25. The creater has newer versions but I like 8 the best.

[rain gazelle]

LOL... Tifa huh? I wonder if anyone around here ever generates Tifa images.... 😆

The illustrious models are tough for me with their severe reliance on artist tags. And you gotta be careful of 'ai face'.

[halcyon burrow]

what's "ai face"?

But yeah, Tifa is a common one for sure XD

The booru tag extention helps a lot when prompting for illustrious.

It is kinda funny though, my image is based on Tifa Lockhart, but it honestly doesn't look like her much. This is in part because in the prompt I stated "based on" instead of saying that it is her. Seems to make a big difference.

You can see it clearly in my characters bracers on her arm though

[rain gazelle]

It's the phenomenon where there is a very common (or a few common) faces that all look the same for characters people generate. 🙂

Now with anime-style art, this is sort of its own thing, as there tends to be a lot of similarities in the characters anyways.

But for some of the more "realistic" or "uncanny valley" models (thinking back to like the RevAnimated model, when that got big), you run more into the AI Face.

[copper matrix]

I find that most models are really prone to that ai skin look

Even in anime ones

But because people don't prompt around it etc it's the most common too

[rain gazelle]

Checking to see if I have an example. Problem is I have like 10k images saved and no way to easily search anything 😄

AI Face....

Versus images like this that are far more interesting. (Though this could have used a face segementation...)

Anyways, here's a Tifa. I'm going to go get a snack. 😛

[halcyon burrow]

Aw, yes, I see what you mean. I have noticed that a lot of faces look similar nod-nod

[supple zenith]

@dry crow Thanks to you 😄

[dry crow]

nice one! np 🙂

[jaunty prawn]

I love the owls lil hat :3

[supple zenith]

Just the lil hat ? 😭

[jaunty prawn]

I like the whole picture ofc!!
The little hat just made me giggle~

[dry rover]

I find that some of the tags that are supposedly bad prompts are not actually that bad or rarely even exactly what I desire

[indigo scroll]

Hey guys i struggling here how can i make ghibli kind of images from the potrats in stable diffusion ?like chatgpt

Whith out changing the facial structures

[mighty wren]

في هذا العيد: أمنية كل مسلم ومسلمة 🇵🇸🤲

[mighty wren]

Anyone here didn't fall in the trap of OpenAI Ghbli style? That's not even Ghbli style despite the adherence to face features. this is Ghbli style

Not this

Or this

[copper matrix]

Its a combination of the different Ghibli styles (movies)

So its very similar but different

[supple zenith]

Haha we all love little hats

[supple zenith]

@jaunty prawn Now giggle.

[weak wigeon]

how do i create ghibli images?

[jaunty prawn]

Whats with the obsession with Ghibli style?

[mighty wren]

seems very different to me.
Even the most powerful AI they say of Sora - OpenAI cannot draw "kissing forehead" correctly

[rain gazelle]

My friend totally had this controller... his dad worked for Nintendo.

[analog mortar]

Create Ghibli style image

[halcyon burrow]

Millions of people had that controller, it's the standard Nintendo controller. Nothing special.

[astral robin]

Turn in info Ghibli art

[rain gazelle]

Not sure if serious... 🤔

[halcyon burrow]

I'm mean, it's true. Lol idk

I played Nintendo all the time as a kid

[mighty wren]

Just made epic profile picture for me lol far beyond Ghibli style drama

[halcyon burrow]

nice! Love it

[mighty wren]

Let's spread real Ghibli

Back to Ghibli quickly 😂

[jaunty prawn]

😴

[mighty wren]

Science has been launched by Muslims

[ripe cedar]

Anyone know a flux workflow where I can have the exact same background and same character, but just with each image showing the character doing different poses?

I make an image with just a background and generate a seperate image of the character with a transparent background and paste the character into the background. But the shadow and lighting are not correct when you do that. So I don't know what to do

[copper matrix]

Spam

[copper matrix]

If you joined them, leave its spam

[ebon steeple]

To troll the original artist with… which I’m not even going to touch it… it’s not even the actual style.. just particals

[mighty wren]

Let's spread real Ghibli

[fathom fjord]

ghibli this picture

[manic vale]

.

[prisma epoch]

@copper matrix

[copper matrix]

Ahh you didn't change the image resolution

If you were to match it in "aspect" ratio you'd get better results

Hmm though your also using SD1.5 i see

On an 1024*1024 image

[copper matrix]

What gpu do you have?

[prisma epoch]

nvidia rtx 4070

[prisma epoch]

The model didn't really turn the Chinese family into robotic humans. For example, they still look like regular people. What is the main issue here? Also, my original resolution is 768x1408 if I generate an image at 512 resolution, wouldn't that result in low quality? What's the logic behind choosing 512x512 and then upscaling afterward? Is that the correct workflow?

And regarding increasing the resolution, which tab under the 'img2img' section should I use? I couldn't find it, and even though I watched tutorials so I still don't fully understand how it works. 😦

[copper matrix]

With a 4070, i recommend looking for SDXL models

Hmm ai gets confusing because old content gets promoted on youtube

The logic on that one though is that SD1.5 was trained and made for 512x512 images

While sdxl is made for 1024x1024

[prisma epoch]

I think control net is not supported 1024x1024

For example I created this it didnt exist before. text2image

My final goal is to convert an existing image of mine into anime style using img2img, but I just can’t get it to work. That’s my only real problem. 😦

When using txt2img, I can enable the 'Hires. fix' feature it generates the image at 512x512 and then upscales it to 1024x1024.
But in img2img, I can’t find the 'Hires. fix' button at all.

So how am I supposed to upscale after generating the image at 512x512 using img2img?
I just can’t figure out the logic behind this!"

I think it can not generate hands and limbs anatomy 😄

[flat sand]

if you want higher resolution, use a newer model. SDXL might be sufficient for your purpose

controlnets can deal with any resolution. Its just important that the control image and the generated image have same resolution

btw. "highres-fix" is nothing else than generating an image in low resolution, upscale it, then do img2img to improve details

[prisma epoch]

ı downloaded this model

[flat sand]

in your case you want to generate an anime version of yourself. img2img is not the best way of doing that

probably something like IPAdapter is better suited

you can try img2img and controlnet (I would use depth controlnet instead of canny though), but you will need a lot of noise to get your image changed into anime style and then you will loose the similarity between your face and the generated anime face

for IPAdapter see https://civitai.com/models/157698/ip-adapter-models-for-sdxl-image-based-prompting

this example for ipadapter should be exactly your usecase:

[prisma epoch]

thanks I also tried something and it's my result but I couldnt understand the logic inpaint to fix fingers and feet

[rain gazelle]

Some Aeriths...

and a Tifa for good measure

[ashen field]

how r u creating those

mine just tries to make it realistic and it looks bad

like here i tried to create a car

this shit happens when i type something like catgirl

like it looks pretty b ad

[quiet current]

looks like sd 1.5

I m guessing that s your problem

(Although you can do pretty amazing things with just sd1.5)

[smoky vigil]

Since it is AI we dismiss it as bad. but if this was hand drawn by a person it would be very disturbing.

Like that would be the series entitled "Gertrude"

[loud sail]

#✨｜sdxl BB King playing guitar on a stage

[turbid scaffold]

crie uma umagem de um urso com roupa de super man o urso da Burr Bear

Create an image of a bear in a Burr Bear superman the bear outfit

[manic vale]

(original) deepseek coder 33b instruct / flux unchained by SCG hyfu 8-step / flux hands 000016 lora / flux detailer v3 000007 / fluxdetail fluxhands unchained RAW photo, showcasing an awe-inspiring evening view in a stylish loft apartment within a futuristic cityscape teeming with life. The room features floor-to-ceiling windows extending at an angle across the full length of the image overlooking a bustling urban backdrop. Soft, golden hues of sunset light filter through these expansive windows, casting dramatic shadows on the luxurious furnishings and the man sitting by the window. High-definition resolution highlights intricate details in the modern architecture and design of the cityscape outside. The television is positioned on a table to complement this panoramic view, creating a well-equipped leisure space for entertainment.

[signal crag]

Fusion of future aesthetics and natural themes, flowing glass texture, refraction effect

[copper matrix]

Made with flux?

These are dope

[rain gazelle]

dall3, of course. 😄

[copper matrix]

Ahh

Shame, would love to run it local

[rain gazelle]

some day? 🙂

[copper matrix]

Well by the time they release it im sure better stuff would be around

[rain gazelle]

yep, keeps changing. dall3 is still my favorite for a while, though.

[mighty wren]

Here is a very noble woman who recently owned Microsoft a big tech company in the US and they say Ps will not change US. just wait in the upcoming years to see a new US.

[still rose]

Chat what u think the “best” flux loras are for portrait fashion photography?

[copper matrix]

@scarlet lodge

to find them

[scarlet lodge]

They look Hella plastic

Although I will try out the Taylor Swift one myself

[copper matrix]

admittedly SD 3.5 is better at skin

[scarlet lodge]

I don't know anyone who uses 3.5

It was instantly overshadowed by flux

[ashen field]

get novaanimexl if ur looking for anime

its good :3

[mighty wren]

Ibtehal Abu Al-Saad: One Woman Against 10^6 Men. Here are more details about the brave stance of Ibtehal Abu Al-Saad against the use of artificial intelligence in Gaza 🇵🇸✊

After being expelled from the hall, she went to a press interview and said that she fears the consequences of her uprising against Microsoft less than she fears unwittingly contributing to writing code that would be used for the genocide in Gaza. 💔

And the issue of using artificial intelligence in warfare is not new; it is documented and has been reported by independent newspapers before. 📰 Mohamed Mahdi Al-Tabar even received a warning when he exposed the matter in the OpenAI Discord, causing him to leave the heavily censored server. 🤐

But this virtuous activist, this fearless exposer—“a witness from among their own” (but she turned to be among Palestinians)—rose up even more forcefully against one of the AI syndicates on the company’s 50th-anniversary celebration day, making an honorable stand that even the Arab rulers themselves could not record, despite their awareness of this crime. 👏👑

And it is not only Microsoft that uses AI for genocide in Gaza; there is also a report accusing Google of being complicit in this heinous crime against humanity. ⚠️ You can watch the video “How AI Tells Israel Who to Bomb – Vox,” which documents in detail this new sadism and fascism. 🎬🔍

#Expose_Technology_Crimes #AI_for_Killing #Ibtehal_Abu_Saad_Rises #Palestine_Will_Not_Die #Boycott_Complicit_Tech_Companies
Designed with artificial intelligence by Mohamed Mahdi Al-Tabar 🤖🎨

You may use this image (but please credit the source and link to the designer’s personal page) 📷📢

[craggy grotto]

why dont work inpaint in FLUX model?

[wispy nest]

because flux aint no inpaint model

[ornate reef]

generate augment code vs cursor image

[hearty violet]

Study that I'm doing put a 3d image in a photo

3d model from blender

image where where the 3d model will be inserted (like context)

I did this results with inpaint

Applyed lineart on this and got a result more realistic

I can play with the form lowing the controlnet weight

[dry blaze]

This will be interesting. Gonna try making a lora that uses 5 seconds, up to minutes worth of videos.

Though, i don't know how to make it train batches of frame sets yet.

[hearty violet]

why on some Vast.AI that I rent, I install the extension but it dont appear on the UI?
I install HUB and Civit.AI extension

I already reaload the UI

[dry blaze]

Oh boi, thank heck for batch scripts lol.

Had GPT make me a script for every range of scene changes and happenings to describe everything at proper intervals for this one 11 sec video clip. And that's one out of 10's of clips i have plans for lol

[strong lance]

#🆕｜sd3 cat

[snow shore]

How do I download Flux to webui version of Stable

[wheat siren]

generate website for procurement system

#🏞｜general-with-images Webiste

[hearty violet]

do this is a error message? My SD don't open

ok, thanks. Was the second isntante that I tryed

hey, this guy is a bot??

he deleted the message

[viral frost]

Has anyone tried HiDream I1? It’s a new 17B t2i model released under the MIT license and supposedly has done well on some benchmarks and arenas. It uses T5, CLIP, and llama 3.1 8B. https://huggingface.co/HiDream-ai/HiDream-I1-Full

[rain gazelle]

ChatGPT image gen... pretty good.

love the order of the names.. haha

[pine sapphire]

#🏞｜general-with-images Three East Asian children sitting on lush green grass, chatting and laughing. One cheerful boy holding a smartphone, gazing at the screen with bright eyes and a radiant smile, sunlight creating soft bokeh in the background, vibrant summer atmosphere, anime-inspired illustration style with soft shading and warm color palette.

[viral frost]

Another (relatively) new model, apparently autoregressive: Lumina-mGPT-2.0. https://github.com/Alpha-VLLM/Lumina-mGPT-2.0

Just, saw this, too: someone has attempted to distill T5-XXL to keep just the visually-meaningful stuff. Could potentially be useful to reduce memory use with Flux and SD3, but I haven't tried it yet. https://huggingface.co/LifuWang/DistillT5

[lament anvil]

generate a picture with a dog in

[dawn seal]

That living room would be noisy as hell. Interesting image, though.

[viral frost]

Unfortunately, this doesn't seem to work with Flux in ComfyUI on MPS. It just generates noise.

[scenic sand]

Any Stable Diffusion/AI in general badasses can tell me how does one generate more images in different poses of the same character wearing the same clothes/armor?
This is as close as I could get, I think. Idk though
I tried using ControlNet and all but it didnt seem to work very well

[rain gazelle]

try 'character sheet' in your prompt.

[scenic sand]

I did, thats how I got the lil sprites, but I need to make separate ones too

I just need to know if theres a trick in general to keep the same looks/armor

[rain gazelle]

You can use a Lora for your character, and/or IPAdapter.

[scenic sand]

Lora?

Like create one?

[rain gazelle]

Yep

Or use an existing one if one exists for your character.

[scenic sand]

definitely doesnt

but ill look into it

whats an IPAdapter?

[rain gazelle]

Or decide Wonder Woman and Mega Man should duel.

Fun fact: "Wonder Woman" was not in my prompt anywhere. Thanks AI.

[rain gazelle]

It uses image inputs along with promtps to copy composition, style, characters, etc.

I have not personally used it, but many have.

[scenic sand]

Isnt that the same thing ControlNet does?

[rain gazelle]

Or, well, hmm... maybe I did many many many months ago. 😄

[scenic sand]

Is it an addon on Stable Diffusion webui or is it a separate thing?

[rain gazelle]

It is different than ControlNet, but can be combined with Controlnet.

[rain gazelle]

I know it is supported with comfy, I'm not sure what else.

[scenic sand]

Im on Automatic11111 or whatever its called

[rain gazelle]

I'm sure someone has published a tutorial on the Y'tube.

In fact, I think the creator publishes videos. They might be a bit heady.

I think it is limited to 1.5 and XL models.

[scenic sand]

Im using ChatGPT as my guide up to now lmao

[candid lake]

gggg

[flat sand]

no. A controlnet is "layed over your image", so it influences the image directly. An IPAdapter is more like an "inspiration". The model takes style or content from your conditioning image but it does not copy it 1:1

[last silo]

help me, I'll pay, how do I make sure that I throw photos into stable diffusion and the model automatically finds clothes and puts them away? I'm creating a telegram bot with this functionality, please help me, I'll pay!

[pseudo scroll]

Hello

[mighty wren]

This style > Ghibli

[mighty wren]

Again AI showcasing bias? This ruined Rockman integrity. Please delete 🥹

[rain gazelle]

You never know what it'll do next!

[copper matrix]

You seem really upset about ai bias when it's bias exists of the images in its dataset

I've had random characters show up sometimes due the scene, no big deal

[mighty wren]

Best wallpaper. I wonder if you guys can help me to enhance it further with your inspiration like by removing some artifices introducing something else

[copper matrix]

Seems like a depressing wallpaper, why not have one that celebrates the culture of Gaza instead?

You seem to advocate a lot for it so some positivity could be nice

But if that's the vibe your going for id make the right half of the street a bit darker to match the left

[mighty wren]

Can't ignore the darkness for celebrating the culture. It is not a subject matter of the moment but I hope to do it soon. Unfortunately we don't live in a pink world and yeah I will edit the right side now to see though It seems to me like a flashback scene from an Anime

[brisk spade]

You could have the subject look into the camera, Mohammed.

[mighty wren]

that would be destructive unless I use Sora

[copper matrix]

hmm otherwise, she seems to be a bit "clean" considering the devesation in her surroundings

maybe some grime or dust

[mighty wren]

seems darker now

[copper matrix]

looks better yeah

hmm

the airplanes below the text are also odd ish

[mighty wren]

It is symbolic rather than realsitic but the presentation itself should be enhanced

I would like to make it more realistic that It will not sound like AI

I would it have like visual illusion 3D overlap. It did do it but It is very flat ant nod smooth

[brisk spade]

Get a non cartoon model

[mighty wren]

seems nicer but the anotmy can still be enhanced further

[brisk spade]

What are you using?

[rain gazelle]

Here, have a Tifa.

[brisk spade]

an Tifa?

[sterile kiln]

Hey Omnia, what model is it?

[rain gazelle]

That was 4o.

[rain gazelle]

New Google video generation model.

[still rose]

RayFlux FP8 (4 min a gen on my m2 pro MacBook with 16gb ram but I can’t complain it works)