🏞｜general-with-images

[steep sleet]

There should be an IF_Image-to-Prompt node

[dire lion]

ah need to download at ollama.com

[steep sleet]

cmd yes

[dire lion]

ya that is the node

[steep sleet]

Either direct d/load from ollama

Or open a command prompt and type ollama run llava2:latest

[dire lion]

don't know if command prompt works for me

[steep sleet]

ok

[dire lion]

where do I put the model if I download it from ollama?

[steep sleet]

It knows where it s going automatically

[dire lion]

ah it's an executable

[steep sleet]

Yes, double-click when it has finished d/loading

[dire lion]

still gives the invalid model selected error

there's no model to select

no model when I type ollama list either

finally figured out

llava2:latest dows not exist for some reason

had to pull llava

then run it in command

and now I have the model in the module

no idea what's going on here

how is it pulling the text string?

I provided this image

and gave it this prompt in the IF image2prompt : What is this an image of? Make it look like a Heroes of Might and Magic VII artwork.

and then I get this prompt: Question: What is this an image of? Make it look like a Heroes of Might and Magic VII artwork.
Response: Best quality, professional, Product photo shoot, This image features a character that resembles the style of a Heroes of Might and Magic (HoMM) VII artwork. The character is a woman with red hair, styled in a way that might be reminiscent of fantasy or role-playing game aesthetics. She has a confident expression and is holding what appears to be a staff or wand. The background is dark and moody, which could suggest an epic or adventurous setting often associated with the HoMM series.

In the style of HoMM VII artwork, this character might represent a hero that the player would control in their quests within the game. The art style includes detailed shading and lighting effects to give depth and dimension to the character and the environment. Mythical, surreal, magical, otherworldly, imaginative,

and the result is this

oh forgot to lower the strenght on the LORA

[steep sleet]

Well, it needs careful prompting - some'll work - and others it'll miss by a mile!!!

[dire lion]

what's a good low strength?

0.20?

also, insetad of what is this image of, I should describe the source image right?

I just need to apply a painting effect to the 3D model, that's all 😅

[steep sleet]

Make the image look like a painting by van Gogh or Dali or Picasso or Geiger or xxxxxxxxxxx

The source image prompt follows on from the initial question - what is this an image of?

[dire lion]

ya I descibed the source image, but results are similar tot he first one

it shows a closeup of the character, not the whole character

I nned the same pose and same details

just a digital painting / concept art effect applied to the source image

[steep sleet]

Mebbe you actually need Controlnet to go with that ...

[steep sleet]

https://comfyui-wiki.com/tutorial/advanced/how-to-install-and-use-controlnet-models-in-comfyui

[vagrant birch]

https://civitai.com/models/793598

[vagrant dust]

Well search for ipadapter style and composition, this should do the trick. But you input image without background has not much noise to create nice painting backgrounds

[stark bay]

here's a raw output from my newly installed Flux Dev installation through ComfyUI. Isn't it incredibly grainy? Am I stupid? This is 10 steps.

[wispy nest]

yeah that's pretty grainy

[stark bay]

I can't believe I got it running on an 8gb RTX 3070, and only 16gb system memory. Apparently, 20 to 30 steps is normal when using Flux [dev], this is 26 steps:

[steep sleet]

Vaguely abstracted modernist oil painting in an expressionist painterly style. A young child stands by the glass wall of a zoo, a gorilla sitting on the other side in its leafy green enclosure. They both hold up a hand to sign a kind "I love you". Subtle imperfections and splattery effect. Bold textures."

[stark bay]

Say less

[dire lion]

I tried with control net and IPAdapters, problem is I need to keep denoise to 0.2 - 0.3 in order to maintain maximum detail, everything over that distorts details or hallucinates new unwanted ones. Then there's the balance between detail and painting effect: if I keep denoise to 0.3, the painting effect is almost non-existent.

[dire lion]

I don't need the backgrounds, the result will be cut-out

[steep sleet]

Yes it is a delicate balance - keep trying until you find the sweetspot

Openpose Controlnet will give you the subject's pose

Canny Controlnet will give you the contours and outline of your subject

[dire lion]

ya I used canny and depth

[steep sleet]

I added 'cubism' to the text box

(Slightly cubist)

[vagrant birch]

is there any ipadapter for flux that works in forge?

[steep sleet]

https://www.youtube.com/watch?v=XomzICHIXrA

[vagrant birch]

thanks

[lavish garden]

Prompt: https://civitai.com/images/30615175

"The Keeper" - OC (Glory Day)

[eager siren]

#🏞｜general-with-images

[languid pebble]

Good morning!

[stark bay]

Just had Flux crash my entire PC. GG

Don't try to make fp16 work with 8gb VRAM and 16gb system memory... lol

[languid pebble]

Yes, that's a little low ....

[stark bay]

good morning to you too sir! :^)

[languid pebble]

You also have something to eat 😄

[stark bay]

the whole thing instantly locked up and both my monitors turned off with a "no signal" message. my PC stayed on, but it was unresponsive. had to restart the whole thing

[languid pebble]

There are lower versions that should work ...

[languid pebble]

https://www.youtube.com/watch?v=PmfPuv4-Www&t=1176s

[stark bay]

This is Flux1[dev] running in ComfyUI. I can understand that NF4 is easier on a lower VRAM/system RAM system, but is the image quality significantly different or are the two indistinguishable?

[wispy nest]

small differences

its definitely fine

not indistinguishable though

[languid pebble]

TBH I didn't try lower configs ...

[stark bay]

Thanks man, appreciate the help. If I set the weight_dtype to fp8_e4m3fn with clip_name1 set to t5xxl_fp16.safetensors, I can make it work - it does slow down my system severely though. Would it be worth it to try NF4 or should I stick to my current setup?

I'm gonna clarify that I'm not doing any type of productivity though. I just like to gen for fun.

[wispy nest]

its important to keep the entire diffusion model in VRAM

personally if I had low VRAM, I would cache text embeddings

i.e. make my text embeddings earlier and save them

then use a workflow that loads existing text embeddings instead of using an encoder

I believe there is smaller than NF4 by the way

the GGUF quants go really low

[stark bay]

What resources would you recommend I look at to understand what this means? I'm pretty new in this space - I'm just a gamer trying to make some cool pictures with my RTX 3070. I'm not shy to try to learn about how these things work, though. It just seems a little confusing and overwhelming. Thanks in advance.

[wispy nest]

I think I can explain each bit

so

you have DRAM on your motherboard, which is what is called RAM most of the time

and super fast VRAM on your graphics card

you want things that require speed to use the VRAM

the text encoder is t5 and clip

and the diffusion model is the main Flux bit

[stark bay]

I'm with you so far... :D

[wispy nest]

overall Flux is 4 models
clip, t5, vae and the main diffusion bit

so

if you put clip and t5 on your DRAM that is not so bad

but if you put the main diffusion bit on DRAM its really slow

I am actually not sure about the VAE. Its small anyway so I guess put it on VRAM

[stark bay]

Because I need to use the tensor cores in my GPU to handle the diffusion, yeah?

[wispy nest]

yeah

there is a thing called CPU-offloading
where the model is mostly on GPU but some layers are on CPU
I don't like it but some people do that

[stark bay]

Do you know how I would specifically allocate the clip_l and t5xxl_fp16 processes to be handled by the DRAM rather than the VRAM?

[wispy nest]

by far the easiest thing would be to just use ComfyUI
and it does all this for you

it can even do it dynamically
so it can load and offload parts during the workflow, automatically

[stark bay]

I'm already using ComfyUI, and I actually have it all loaded up as we speak.

[wispy nest]

like offload clip after clip has finished

okay there should be comfy settings that get what you want, as far as I know

would recommend asking on the comfy discord

I don't do low vram stuff personally so I don't know

[stark bay]

I'll do that. Thank you so much dude.

I always feel bad asking for help with these things as it's unpaid labor LOL

[wispy nest]

its okay people like talking about AI and hardware stuff

[languid pebble]

If we can help many like to do ...

[wispy nest]

I personally put everything FP8 cos there is a speedup just for FP8
on recent GPUs

this is the --fast flag in comfy

you do lose a little bit of lora following with anything other than FP16 though

[stark bay]

I really appreciate it!

So, it'd be better for me to use t5xxl_fp8 rather than t5xxl_fp16?

[wispy nest]

your GPU wouldn't get the FP8 speed boost so there is not special advantage for FP8

its not huge anyway its like 30%

[stark bay]

ah, because I'm not on more recent architecture?

[wispy nest]

yeah its only a handful of very recent GPUs

and then the next ones I guess

[stark bay]

I see.

I'll try to follow your advice.

[wispy nest]

I use cloud so I jump around GPU a lot

[stark bay]

That makes sense. I've tried cloud but I find local rendering really interesting. I don't know if my power bill does.

[languid pebble]

Working local with A.I. for 2 years now and don't really have an issue ...

[wispy nest]

its mostly upscaling without tiling that requires the crazy high VRAM

I've started using tiling more these days anyway

[stark bay]

Another expensive hobby for me... oh well.

[wispy nest]

ye its a costly hobby

[low whale]

video is taking ages to get build from an image on my macbook.. any suggestions?

[wispy nest]

video?

[languid pebble]

Wait for NVIDIA 5090 😄

[stark bay]

I gotta wait to hit the jackpot before anything else LOL

[languid pebble]

Thank god I have a 5000.- Coupon for the ASUS store ^^

[stark bay]

Is that dollars or another currency?

[languid pebble]

€ ...

[stark bay]

Jealousy sets in...

You'll be able to buy exactly three quarters of the RTX 5090 with 5000 EUR. ;)

with how GPU prices are evolving... lol

[languid pebble]

Bad luck AMD missed the A.I. train ....

[stark bay]

At least the AI boom has resulted in more companies looking into fabricating semiconductors, which hopefully will drive down the prices considering there's gonna be more competition.

We'll see about that last part though.

[low whale]

image to video

[stark bay]

... on a Macbook?

[low whale]

yes

[languid pebble]

Stable Video Diffusion?

[low whale]

yes

[stark bay]

Remember to bring the fire extinguisher!

[languid pebble]

It has to calculate 25 pictures ... that takes some time ... even on a 4090

[low whale]

ok

[languid pebble]

(I'm cheating)

[stark bay]

must be some serious computing power behind something like this

meanwhile my system is at a grinding halt to make something like this:

[languid pebble]

I use SVD to generate at 3 FPS and boost it to 25 FPS afterwards ...

[steep sleet]

"Hawaii 5-oh-9-0!!!"

[steep sleet]

Plotagraph does a lot like this - or the Motion Timeline in Photoshop! It's pretty cool!

[languid pebble]

Sometimes it helps if you are an oldie with A.I. 😄

[steep sleet]

Yes, most video is as yet underwhelming. But I am trying Klingai, and RunwayML Gen-3 for a better, more joined-up experience! 😄

[steep sleet]

Suggested price-point $1,600 - https://www.dexerto.com/tech/nvidia-rtx-5090-2301788/

[languid pebble]

That would be fair ...

[stark bay]

9 years ago, a top-of-the-line GPU was 999 USD.

(NVIDIA GeForce GTX Titan X)

[languid pebble]

But I think 4090 has been around 2k for some time

[steep sleet]

When 5090 is released - buy some 4090's like there's no tomorrow!!! What with the price-drop ...

[languid pebble]

Not sure about that. Rumors say they already clear the storage ...

But the price on used market should drop ...

[languid pebble]

Banana! 🤪

[steep sleet]

A close-up, intense view of a baseball catcher signaling a low curveball, with the focus on the right hand giving the sign. The catcher’s fingers are clearly extended downward, hidden behind his legs, showing two fingers as he discreetly calls for the pitch. His left hand holds the mitt low to the ground, but the real attention is on the precise and subtle movement of the fingers, communicating strategy in a tense, pressure-filled moment. The dirt-covered ground and beads of sweat on his hand add to the intensity, while the shadowy atmosphere heightens the focus on the sign itself.

[stark bay]

Did you generate this prompt with ChatGPT?

[languid pebble]

I'd think so ... A.I. won't know how to signal a curveball

[steep sleet]

DallE Theme of the Day Discord - I borrowed it!!! 😄

GGUF Flux + LoRAs

[clever oar]

At first the video card progressed for the sake of games, now for the sake of AI

[kindred creek]

xd ^

been on dis screem for a bit now xd

yay it moved

[hard quartz]

Does anybody knows of an AI that could turn Oil Painting to realisitic 3d model?

[viral python]

logo chinese style

[steep sleet]

A charming chalk drawing of a futuristic spacescape, featuring a campsite with tents, sleeping bags, and outdoor essentials, the sky is a glimpse of outer space with stars and comets. The landscape radiates warmth and comfort, bathed in a golden glow that entices viewers to explore its hidden secrets. Looming over the campsite is a sleek, modern space station, connecting to the lunar surface via a shimmering energy bridge that glows with life.

[royal monolith]

with the lora

without the lora

@covert locust

[covert locust]

look at this shit lool

[royal monolith]

the boy with the raven almost looks like one of his screen shots though. what's your prompt?

[covert locust]

the same as hugginface a girl holding a large black raven close to her face, eerie and mysterious, in the style of BSstyle004....trying using 1024 x 1024

[scenic yew]

nice, detail is crazy. what are you using?

[crisp stream]

Stable Cascade with Invictus Redmond v1.1 checkpoint for stage C (most important), for sharpness the CSBW refined lite stage b and HQ vae with 21 steps along dmpp SDE karras on Stage C and euler_a karras at 31 steps on stage B. Can link you the models if you like

here is as well the workflow (it´s using an input image in this case)

[scenic yew]

This one? https://civitai.com/models/316681/invictusredmond-stable-cascade-stage-c-finetune-generalist-model
will try it out later

[crisp stream]

yep

[crisp stream]

there are the additional models:

https://huggingface.co/ClownsharkBatwing/CSBW_Style/blob/main/cascade_B-lite_refined_CSBW_v1.1.safetensors

https://huggingface.co/madebyollin/stage-a-ft-hq

[crisp stream]

@scenic yew here is the workflow

[steep sleet]

Q8_GGUF Flux

[crisp stream]

🤭

[solar pivot]

guy

[brave niche]

|

[outer sandal]

Im trying to install these files, what folder do i put them in when i get them?>

[clever oar]

you keep posting psychedelic pictures again😃

[fossil mason]

User hasn't listened to enough albums (0 of required 9) for a chart this size.
Please try a smaller chart or a bigger time period (weekly/monthly/quarterly/half/yearly/alltime).

[vital sand]

wow the new checkpoint illustrious-xl is really good

[vital sand]

like novelai alternative

[steep sleet]

A vintage Polaroid camera with a picture emerging from it. The photograph distorts reality, showing trees, buildings, and faces that twist into impossible shapes. Skyscrapers spiral upwards, melting into the clouds, while the horizon folds in on itself. A human figure stretches like elastic, caught between dimensions. The once-clear landscape morphs into a kaleidoscope of colors and surreal shapes, defying logic. The image emerging from the Polaroid frame contains this surreal world where reality and the impossible blur together.

[pallid ruin]

so cute ❤️

[soft cairn]

Намалюй Логотип Pavlo Ruban School Діагностика ходової частини та Встановлення кутів коліс . Елементи автомобіля: Включи силует автомобіля або його деталей (наприклад, колеса або підвіску), щоб одразу зрозуміти, що моя школа пов'язана з автомобільною діагностикою.

Інструменти: Додай іконки інструментів, які використовуються при діагностиці авто (наприклад, гайкові ключі або рівні), щоб підкреслити практичний аспект навчання.


Текст: Використовуй сучасний шрифт, щоб написати назву "Pavlo Ruban School". Можна виділити слово "School", щоб акцентувати на освіті, або "Ruban" для індивідуальності.


Кольори: Використовуй кольори, що асоціюються з автомобілями, такі як синій, чорний або червоний.  також можеш додати металеві відтінки, щоб підкреслити технічну тематику.


Символи: Можеш додати символи, такі як компас або шестерня, що може символізувати точність та якість навчання.

Draw Pavlo Ruban School Logo Diagnostics of the running gear and setting of wheel angles. Car elements: Include a silhouette of a car or its parts (such as wheels or suspension) to immediately understand that my school is related to automotive diagnostics.

Tools: Add icons of tools used in car diagnostics (such as wrenches or levels) to emphasize the practical aspect of learning.

Text: Use a modern font to write the name "Pavlo Ruban School". You can highlight the word "School" to emphasize education, or "Ruban" for individuality.

Colors: Use colors associated with cars, such as blue, black or red. you can also add metallic shades to emphasize the technical theme.

Symbols: You can add symbols, such as a compass or a gear, which can symbolize accuracy and quality of learning.

[rain gazelle]

Trying the "vintage travel poster" look. It's tough.

[steep sleet]

A photorealistic image of a lighthouse standing on a cliff by the sea, with its beacon shining through turbulent weather. The light pierces through stormy clouds, symbolizing hope and guidance. The image captures the contrast between the dark sky and the bright, unwavering lighthouse light, representing direction in times of uncertainty.

[languid pebble]

Spaceballs? 😄

[languid pebble]

Cheesus Christ!

[dense vigil]

flux is something else man

[rain gazelle]

dang cropping

[languid pebble]

Just post 1

[languid pebble]

People will see full version clicking it ...

[rain gazelle]

yeah I know.. I just don't like to flood the chat 🙂

[languid pebble]

I'd think if you post 2 pictures seperated it should be OK

[clever oar]

Mojo

hello

[languid pebble]

Hi dicordos

[clever oar]

how are you

have news?

[languid pebble]

Fine here and you? Sounds like there's a new Flux-Controlnet upscaler

[clever oar]

you use flux local&

[languid pebble]

Yes ... but AFAIK there's no downloadable workflow yet ...

I tried online and that's the result ...

[clever oar]

dont understand its work on yor pc &

or online only

a cant run xl models...if i start flux/...

its kill my pc

[languid pebble]

There's an online version of the upscaler at the moment but to many requests: https://huggingface.co/spaces/jasperai/Flux.1-dev-Controlnet-Upscaler Willing to use it local when I find a workflow ....

[languid pebble]

Here you get a few credits for free every 6 hours. Better only produce 1 picture per trial: https://fluxpro.art/create

[clever oar]

i use it online

on hug face

sd 1.5 with lora

[languid pebble]

SD1.5 is still good for some things!

[clever oar]

yep!

i like freedom in sd 15

[languid pebble]

The only bad thing is to upscale it in a good way ...

[clever oar]

you know what is it?😃

[languid pebble]

Kinda mask?

[clever oar]

yes

from horror game

[clever oar]

same game

[wispy nest]

So one second I'm generating a cute girl or something and then I add an innocent seeming word and suddenly stable diffusion feels like generating an eldritch abomination
Why is this?

Like tf

[viral frost]

I spend every Saturday night hanging out at the Turex Naterfed Hat Lime.

[viral frost]

This just looks like a very computationally expensive bicubic scale.

[languid pebble]

Not perfect but if you need a bigger print OK

[crisp stream]

would you mind posting the prior upscale version, would like to compare to upscaling locally

[languid pebble]

Here

Some details gone, some added ... if you need a biger print it's OK

[steep sleet]

Florence2/GGUF_Flux img2img

[deft walrus]

I'd like to make other images like this but do an inpainting with a new image. What's the best way of doing that for a large number of different images? I'd like the cards to have different images, but the same background.

[shut pivot]

Has anyone tried high CFG scale (30 CFG), using the Integrated Dynamic Thresholding (CFG fix) in forge UI for Flux models, in order to create details? The image is from SD 1.5, automatic 1111 for reference (!)

[clever oar]

I haven’t used SD 1.5 for a long time and forgot that you need to be an promt engineer to get simple things😃

[grave copper]

Hi all, does anyone here understand why Stable Diffusion doesn't understand prompts and makes siamese and many inconsistencies?

[vagrant dust]

Hi there, would go with attention mask and the ipadapter. Create a workflow in comfy use the developer tools to make an api out of it and then use a simple python script to send the large number of card prompts to your api.

[viral frost]

Please don't post NSFW images in this Discord. You're liable to get banned. Also, I don't know what type of model that is, but if it's based off of SD1.5 then its base resolution is 512x512. If you go much higher in terms of pixel count, you will start to get repeating features.

[deft walrus]

I'll experiment with some of that but it seems beyond my level of knowledge. I think I found something to help with implementing that tho. https://github.com/SaladTechnologies/comfyui-api

[grave copper]

Sorry I wasn't thinking when I shared I forgot to blur 😬

[royal monolith]

https://techcrunch.com/2024/09/29/gov-newsom-vetoes-californias-controversial-ai-bill-sb-1047/

[wispy nest]

amazing news

[nimble mason]

wow

[nimble mason]

that's gorgeous

[languid pebble]

"My treasure..."

[nimble mason]

pretty sure these are safe to eat

[languid pebble]

Wanted for murder ...

[lavish garden]

Prompt: https://civitai.com/images/32038635

"I'm everywhere I'm so..." - Luigi (Nintendo)

[steep sleet]

"Minimalist rugged oil painting in faded earthy green hues, capturing delicate details in vast solid patches. A post-apocalyptic city lies in smoking ruins, beginning to be overtaken by nature. Vines crawl up the sides of half-crumbled skyscrapers and leaves form in mats in the streets. Intense bluish mist slightly obscure the setting sun in warm salmon tones. Low angle cityscape from the ground on the other side of the river near the city."

[heavy mist]

First posting of mine.
I am "Testing the Waters" as it were

[clever oar]

must be done with the Prime Minister of Israel

[pallid ruin]

Nice 🤗

[tepid fjord]

Does anyone know any model which can create anime screenshots / anime screencap style images

[nimble mason]

pretty sure that hasn't been done

[tepid fjord]

Ohh damnn :(

But i see people making pic that are straight up from anime but ai generated

[steep sleet]

70+ text prompts from today's img2img using Florence2

[vagrant birch]

lora?

[wheat pecan]

good

[wispy nest]

okay

i mean

[cinder coral]

Still easier than that mess of cobwebs in "comfy" UI!

[nimble mason]

comfyui is 100% the way to go

[cinder coral]

I believe it, but I'm not very intelligent lol. Other people's workflows hardly ever work, there are 10s to 20s of dependency plugins for every workflow. I couldn't even figure out how to get a lora to work in it without downloading a workflow and it was still not right.

[cinder coral]

In Forge you put your lora in and prompt it and you're gold.

[wispy nest]

that is a shrog

frark

[wispy nest]

my main advice would just be to not use other people's workflows

if you build it yourself out of the nodes its easier

it takes a while but if you learn what each node does

you don't rly need to look up workflows after that cos it fits like lego blocks anyway

[wet trench]

Character:
Anna, a friendly ghost, is floating gently, with a soft, translucent glow surrounding her. She has long, wavy blonde hair that drifts slightly around her face. Her facial features are warm and kind, with gentle eyes that convey wisdom and care. She’s wearing a faint outline of a factory worker’s uniform – a simple, ghostly version of workwear, with a button-up shirt and utility pants that are slightly faded to fit her ghostly appearance.

Background:
In the background, you can see a whimsical, old-fashioned factory floor. It’s a light and airy space, with soft, muted colors that give it a dreamlike feel. There are conveyor belts, machines with rounded edges, and the sense of light work being done. Ghostly workers in the background are faint and almost translucent like Anna, giving a sense of her past life in the factory.

Style:
The overall illustration is playful and charming, with bright, soft colors. Anna, though a ghost, feels approachable and kind, making her a great character for children to relate to.

#🏞｜general-with-images Character:
Anna, a friendly ghost, is floating gently, with a soft, translucent glow surrounding her. She has long, wavy blonde hair that drifts slightly around her face. Her facial features are warm and kind, with gentle eyes that convey wisdom and care. She’s wearing a faint outline of a factory worker’s uniform – a simple, ghostly version of workwear, with a button-up shirt and utility pants that are slightly faded to fit her ghostly appearance.

Background:
In the background, you can see a whimsical, old-fashioned factory floor. It’s a light and airy space, with soft, muted colors that give it a dreamlike feel. There are conveyor belts, machines with rounded edges, and the sense of light work being done. Ghostly workers in the background are faint and almost translucent like Anna, giving a sense of her past life in the factory.

Style:
The overall illustration is playful and charming, with bright, soft colors. Anna, though a ghost, feels approachable and kind, making her a great character for children to relate to.

[pallid ruin]

with flux schnell

[tepid fjord]

any model like niji journey?

[crisp stream]

Just for comparison when upscaled locally with realESRGANplus x4

[sterile kiln]

My new model guys 🙂
https://civitai.com/models/818292?modelVersionId=915018

[languid pebble]

Thanks Moofi

[mighty wren]

Flux Pro is just another level. Overtaking DALL3 as well in realism

[west cypress]

does anyone know how to prompt an output like this? i love the style

[still rose]

I’m loving flux with Loras

(It’s also my pfp Ik I like it a lot)

[cinder coral]

which front-end are you using?

[wild sorrel]

anyone knows good lora for icons or maybe prompt that limits main subject to like 80% of image size?
Like this thingy

well, that contradicts my explanation...bare minimum - it needs to be completely inside image, not to be cut out , but if it wuld be like 80-100% that's even better

like...there's example of what I'm trying to avoid:

it often gets worse...that's somewhat usable, but you get the idea

[viral frost]

Stole your prompt and changed it to be photographic.

[still rose]

I used on macbook draw things as the application

[plain minnow]

What exactly is this, anyone can tell me please?

[vagrant dust]

Trying to be as exact as possible with the tons of information you added to your question.
It is a screenshot with black background white letters which shows python code. It checks for some variables and wether they are true or false. As the variables are named nsfw_preview and nsfw I would say it is code to prevent showing images with nsfw content in the preview (maybe if an specific option is not set or set).

[plain minnow]

And how do i change this, can u tell me please?

[vagrant dust]

First by adding some more information, and second by maybe using a channel like #🤝｜tech-support
What information you might ask yourself. Well start of by what do you want to change, what are you trying to do, what tool are you trying to get running, maybe the operation system, how could anyone try to recreate the problem,…

[plain minnow]

Ok thanks i ll ask there

[steep sleet]

Searge SDXL "Trans-y"

[deep bane]

ok guys what will i have to do to get tis image colored in stable diffusion? New to this so please go easy on me lol

[viral frost]

Hmm, Flux gives me quite different results. I don't have your custom nodes, so maybe I mistook your negative for a positive in the metadata.

[steep sleet]

This is Searge SDXL Nodes

https://github.com/SeargeDP/SeargeSDXL/releases

[reef mantle]

#artisan-1 people running on the beach

[wraith plank]

my first generated image

hell ye got it to work

[languid pebble]

Congrats!

[sterile kiln]

in an alternate reality, Frieza is actually Barack Obama...

[sterile kiln]

The entire fate of humanity rested on a young Sayian, who had grown up on Earth, and was called “Trump”...

[vagrant dust]

With colorize you mean like this?

[vagrant dust]

Why an online service, you could use CogX img2vid Workflow or?

As usual the last 5 Frames are out of order and my prompt was kind of simplistic: a blond man in front of a mechanical device get burned by an explosion

[languid pebble]

Always trying different ways ...

And I didn't install Cog5 local yet 😄

[vagrant dust]

Understandable because it has some limitations for now. For example the resolution is fixed with 720 x 480 so the cropping was necessary. Pretty sure if time goes by it will be easier to use out of the box

[languid pebble]

I've used it once online and it took me 1200 seconds 😄

[weak apex]

llustration of a 5 year old boy at school.

[lucid cradle]

[steep sleet]

Illustration of Picasso at McDonalds 🥳

[languid pebble]

https://youtu.be/oTO6SeYuAcM?si=ni9JS2OVcNOjnhFh 😄

[clever oar]

Mojo

[languid pebble]

dicordos 🙂

[clever oar]

where did you make your video?

[languid pebble]

The one with podcast just showing a picture?

[clever oar]

last your video

[languid pebble]

Audio is from here: https://notebooklm.google.com/

[clever oar]

no audio

with liquid terminator

[languid pebble]

It should have a watermark ...

Pika.art

[clever oar]

you tried haluoai?

i think best for video

[languid pebble]

I have an open tab with it but don't think I've tried it ...

[clever oar]

😃

[languid pebble]

Have you seen this? 😄

#🏞｜general-with-images message

[clever oar]

its ai dictors?

[languid pebble]

All A.I. with my twitter as only input 😄

[clever oar]

ai takes world😃

pig bomb😃

[languid pebble]

poor pigs ^^

[clever oar]

don't worry the fried meat went to kebab

[celest sigil]

1

[clever oar]

good morning

[pallid ruin]

Good morning 🤗

[clever oar]

i not sleep need to go😎

[pallid ruin]

x2 but im falling asleep

Time to sleep xd

[steep sleet]

Searge SDXL + Clownshark and Future Cubism LoRAs

[still rose]

Guys what would u say the best model rn is for realistic images on people?

[sterile kiln]

Flux

[still rose]

Figured XD just checking on what people answer with

[sterile kiln]

Not only are the basic model(s) incredible, but the lora works 10 times better than on any SD version.

[still rose]

Question what loras you recommend for realism

I mean I think it’s known also now that Flux tends to also make plasticity skin

[sterile kiln]

All this realism, 4k and others lora like this are just dogshit.
There are a lot of people who don't really understand what they're doing, and give a biased view of their experience. Take Flux1dev, or flux1devQ8, then the smaller models, there's already a world between the 2.
In any case, Flux is currently the most realistic.

[still rose]

Fair enough I just struggle with doing prompts

[sterile kiln]

Personally, I don't think the skin looks plasticky...

[still rose]

ayo prompt needed

[sterile kiln]

`a woman, sexy, black tanktop with a mini skirt, on a futuristic Yamaha motorcycle, a pistol in her right hand, futuristic cyberpunk city, night, neon lights, shadow, sharp, low angle view, dark colors, lora:le@hFluxv1:0.6

Steps: 20,
Sampler: Euler, Schedule type: Simple,
CFG scale: 1, Distilled CFG Scale: 3.5,
Seed: 4266710719,
Size: 1366x768,
Model hash: 52cfce60d7,
Model: flux1-dev-Q8_0, Lora hashes: "le@hFluxv1: e3b0c44298fc", Version: f2.0.1v1.10.1-previous-488-g79b25a82, Diffusion in Low Bits: Automatic (fp16 LoRA), Module 1: ae, Module 2: clip_l, Module 3: t5-v1_1-xxl-encoder-Q8_0`

(hires fix x1.5, 40 steps 4xultrasharp, EXTRA tab x2 4xultrasharp upscale)

[still rose]

What’s that lora u had on?

[sterile kiln]

A lora I made

[still rose]

Ah ok nice my friend

Ty for the information also

Appreciate it I remember u also tryna help me with getting flux in MacBook with Forge

[tepid hornet]

I am trying to use controlnet openpose and it seems to be ignoring the poses I gave it entirely.

is it my model?

[clever oar]

its works with many poses?

[tepid hornet]

It doesn't work with any pose at all as far as I can tell

[clever oar]

i check my model

you try it with preprocessor?

[tepid hornet]

yes, it just ignores it entirely

[clever oar]

what in cmd when you start it?

[tepid hornet]

I only added it today, but the model is from almost a year ago so it might simply be incompatible with something

I'm trying that next

[clever oar]

what your contol weight

[tepid hornet]

1

[clever oar]

work for me 😃

[tepid hornet]

isnt that reference? Not poses?

hmmmmm

[clever oar]

rerefence in left

[tepid hornet]

werid, I put in openpose_full and suddenly I cant pick a control model anymore

[clever oar]

very strange

its must show

maybe your model broken

redownload

[tepid hornet]

which one do you use?

soo weird, I have the exact same one

but I'll try reinstalling it

[clever oar]

you know where put it?

[tepid hornet]

in the controlnet folder, right?

[clever oar]

yep

\models\ControlNet

[tepid hornet]

did you get it from civitai or hugging face?

[clever oar]

i thing hugging

canny i get from civitai

[tepid hornet]

Isn't there supposed to be a ph or y-whatever file?

[clever oar]

https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_openpose.pth

[clever oar]

all new?

[tepid hornet]

yes

[clever oar]

lets try!

[tepid hornet]

it

it's taking a while to pre process for some reason

[clever oar]

it works now

[tepid hornet]

just finished, trying it now

[clever oar]

first try it with preprocessor and picture

[clever oar]

👍

[tepid hornet]

do you have controlnet 1.1? I have 1.1.455

[clever oar]

dont remember how check

[tepid hornet]

at the top left

[clever oar]

now all works?

[tepid hornet]

I don't think so

[clever oar]

why

its read pose from your picture

[tepid hornet]

it doesnt apply it

maybe a less complex pose

[clever oar]

ControlNet v1.1.455

use simple pose

if not work check cmd

[tepid hornet]

and now it doesn't work at all 💀

[clever oar]

what hewhat it says when a picture with a pose is generated saying

in cmd

[tepid hornet]

hmmm

[clever oar]

model conflict

[tepid hornet]

it says it's not compatible with my sd version

so, I have to get a different version of SD altogether?

[clever oar]

update your version maybe

[tepid hornet]

idk how to to that lol

[tepid hornet]

ah

at the bottom

[clever oar]

hm

[tepid hornet]

do I have to completley reinstall?

to downgrade a version

[clever oar]

i think maybe you need new cn model

for your sd version

[tepid hornet]

neq cn model? I assume that's 'neq controlnet model'?

[clever oar]

is not comp with sd model(version)

maybe model ?

checkpoint

[tepid hornet]

'

this?

[clever oar]

if i was you i will try different cn model or different checkpoint

can you change

[tepid hornet]

idk

[clever oar]

Stable Diffusion checkpoint

[tepid hornet]

oh, I would have to grab another, I only have the one

gimmie a sec

[clever oar]

you have 1.5 based model?

[tepid hornet]

where do I check that

[clever oar]

name

give me name

[clever oar]

I'm not sure that my model that I gave you is compatible

[tepid hornet]

Do I need to get SD 1.5?

or update to 1.7 like you?

[clever oar]

first try download 1.5 based model an try

[tepid hornet]

sd model or cn model?

https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5

this one work?

[clever oar]

you have xl model now

[tepid hornet]

oh

[clever oar]

you need cn model for xl

[tepid hornet]

could I just grab xl cn model then

oh okay

lemme do that

first, delete the old files

[clever oar]

you need use only right cn model

[tepid hornet]

I have a big ass list now

https://huggingface.co/lllyasviel/sd_control_collection/tree/main

do I just grab the open pose for now?

[clever oar]

dont know 100% what model for you

[clever oar]

maybe there

[clever oar]

works now?

[tepid hornet]

eldtritch abomination

[clever oar]

what is it)

[tepid hornet]

I wanted megumin to do that pose

it didnt work

[tepid hornet]

trying that one next

I dont understand

it gets it

it gets the pose

but it doesn't seem to apply it

[wispy nest]

the thing about control net is

the difference in quality between different ones is huge

so you've gotta do a bunch of research each time

to make sure you have the best one for the task you want to do

[tepid hornet]

how do I even check if it's following my poseing and not my prompt?

[wispy nest]

oh that's easy

take the pose out of the prompt

[tepid hornet]

just did that, trying again

I mean,

it's smilling, but it seems to match

yeah nvm

it's not lol

[wispy nest]

not sure open pose was ever a good idea TBH

compared to just using canny and depth at low strengths and durations

[tepid hornet]

the main reason I wanted poses above everything else is that I want to train a lora trainer for a chracter I made so I can make a comic with them

And I CANT because this damn poser doesn't want to listen to me

[clever oar]

what now in cmd?

[steep sleet]

i2i Searge SDXL

[nimble mason]

wild

[clever oar]

titanic?

[nimble mason]

yep, historic photo actually

[pallid ruin]

https://tenor.com/view/hi-babybeans-babybeansdesigns-cat-hi-cat-hello-gif-11694844958654797346

@nimble mason @clever oar

I could fix flux out of memory error, I forgot adding swap memory to windows using the SSDs 😁

[clever oar]

the current staircase looks older than the one from the past😃

[clever oar]

its message fom me ?😃

[pallid ruin]

Idk xd

[clever oar]

how are you

I bought a bag of sweets, I'll go make tea

☕

[pallid ruin]

Im good, and you?

[pallid ruin]

Ooh nice

[clever oar]

first monster make with flux

second sd 3

I think flux has less brightness in pictures

[pallid ruin]

Yeah probably, in my opinion flux is the best local model but sd3 has better textures most of the time (by default)

[clever oar]

but at the same time they are an order of magnitude better quality

hands and geometry

[clever oar]

i want mix with thix models

all good from flux and from sd3

[pallid ruin]

Idk who asked for the image but here it is

[pallid ruin]

Yeah it can do a lot of things better

But its not perfect

DMD2 is another model that is one of my favourites, the effects or lighting of it is epic

[clever oar]

too heavy for the system

[pallid ruin]

This is DMD2

I could not generate something better on flux

[clever oar]

xl based?

[pallid ruin]

Yep

I used LeoSam´sHelloworld as secondary model or something like that, I dont remember how it worked

It generated that in 4 steps, fast af

[clever oar]

i cant run xl models😎

only 1.5

4gb vram -)

[languid pebble]

Hello dicordos 🙂

Hi Leo 🙂

[steep sleet]

Searge i2i SDXL + LoRAs

[languid pebble]

Maybe we could build kinda snowman ....

[still rose]

Would u lot say flux is also good with anime or I rather just use a dedicated anime model?

[viral frost]

Wow, Adidas AND Nike.

[patent snow]

tensor.art is good

[main portal]

"old washed up cosplayer sick of posing"

[steep sleet]

i2i Searge SDXL + LoRAs

[scenic yew]

It’s actually pretty decent but if you don’t care about prompt following or text rendering, anime-specific models might be better.

There are a few anime specific flux models too, you can try them out.

[wispy nest]

yeah that's the thing, Flux has very high image quality but
its not a specialist model

and the anime community have some specialist ones

[still rose]

Yeah I’ve seen some Flux Loras for anime like the flat anime colour lora and a few others

Idk about anime Flux models I have not seen them

[clever oar]

i ask chat gpt what gpu is needed for run flux local
answer:
GeForce GTX 1660 Super
GeForce RTX 2060
GeForce RTX 3070

😃

[steep sleet]

I'm running Flux.Dev on RTX2070 8Gb VRAM, but I also have 64Gb RAM

[clever oar]

these are not the minimum, these are recommended 🧐

[sterile kiln]

GTX 1660, possible, but with low size models like Q2

[clever oar]

maybe I didn’t specify the model, that’s why this is the result

Intel Core i7 AMD Ryzen 7
NVIDIA GeForce RTX 2070 AMD Radeon RX 6800 XT
chat gpt said is recommended

strange

[sterile kiln]

don't use an AMD GPU, RocM and Zluda are much slower than CUDA, and require separate installation of branches and forks.

[clever oar]

now i dont buy amd gpu -)

[sterile kiln]

if you were in france, I would have suggested to sell you my old 2070 Super x)

[clever oar]

what's the price? -)

[scenic yew]

ChatGPT doesn’t know much about flux, flux is pretty recent and chatgpt has a knowledge cutoff. It can use web search but is not very accurate for things like this.

I believe 3090 is usually the best bang for the buck for sd, and just general ai, plenty of vram and pretty fast while being not too expensive.

[clever oar]

good choice

[scenic yew]

Amd gpus are pretty amazing too, better then most nvidia gpus I believe but it’s just that most libraries don’t really support it.

[clever oar]

i think about new gpu /

first i want 3060 becouse is cheap and 12 gb

now i think need something better

[pallid ruin]

Used 3060 are good for its price (I got mine for $150), but a used 3090 would be a better option for AI because it has 24gb of vram

[languid pebble]

Let it grow ... let it grow ... let it grow ...

[languid pebble]

Using Flux 1.1 Pro API is afaik 0.04$ ...might also be an idea

[languid pebble]

AFAIK you can use it for free @ gliff

[sterile kiln]

120€, but originally it was a PNY dual fan, I put 2 Noctua chromax black (92x14 mm), because with Stable Diffusion, I couldn't stand the noise of the fans anymore.
As a result, it's probably the quietest of the dual fans.

[clever oar]

nice price in my country more expensive used gpu

[rapid mauve]

does anyone know what software is used to make an image like this?

[nimble mason]

diffusers

[rapid mauve]

stable diffusion?

[rapid mauve]

@nimble mason my glorious king

[brazen sand]

do you notice something off with dimensions or something ?

[brazen sand]

or in this is there something off?

[woven token]

Image / Create an eerie, gothic world filled with whimsical, exaggerated characters who inhabit twisted, shadowy landscapes. The scene should blend dark, muted colors with vibrant accents, capturing a sense of fantasy and isolation, while evoking both beauty and unease.

[nimble mason]

Here is the image you requested.

[pallid ruin]

Yeah, that rock in the middle make the man in the right seem like a giant or something if i look fast haha

[pallid ruin]

cool gen! Is this flux?

[woven token]

This is eery and creepy and beautiful 🙏🏿

[brazen sand]

yes that seems to be the case, ty. Do you think that changing the rock will fix it ? Still the position of the man is weird maybe

[brazen sand]

It is not flux, but I don't quite remember if I used cascade and then a lot of inpaint and upscale with SD XL

or jus SD XL

the first one was an img2img with flux

and the inpaint and upscale with SDXL

sometimes I really hate SD as it seems to fix some foot is a nightmare

like why it do it so bad

I got the feeling no one here ever inpaint or correct image

[brazen sand]

Amazing quality, how was it done?

[nimble mason]

Using my RES4LYF nodes

[brazen sand]

what model, flux, SD3?

[nimble mason]

Flux

[brazen sand]

flux dev? do you use some kind of img2img there?

regular flux dev took like 7 minutes to generate here

with 1 megapixel dimension

[nimble mason]

Flux dev, no img2img or cnet etc

[brazen sand]

there is always somee kind of pattern repetition (pardon the redundancy) in your generations, it may have to do with the workflows you made? usually there is carbon copy of stuff

[brazen sand]

It made me remember this

[nimble mason]

What do you mean by carbon copy

[brazen sand]

like the pink trees

[nimble mason]

That's prolly the prompt

The screen cap above is pretty much the entire WF

All that's missing is the prompt and the load model and save image stuff

[brazen sand]

I mean like these ones https://cdn.discordapp.com/attachments/1230206273451069540/1289527448433725543/image.png?ex=6705abc5&is=67045a45&hm=fa4a24a6aceb7fa147231a0b743674305ea98c1040fe4eca4295291c0ee4b04f&

or better this one https://cdn.discordapp.com/attachments/1230206273451069540/1285137281744633911/image.png?ex=6705851c&is=6704339c&hm=732c60f8e5426345b8e580b5561bd427476ab3fc40d6f58e4c24ee508e2c8193&

the way the flowers appear with the same design

it is something that happens in many generation but in yours I noticed specially, I though it has to do with the weird wolkflow stuff

[wispy nest]

likely due to clownsampler

putting a lot of ancestral noise

[steep sleet]

The originals are produced by Searge SDXL i2i; then fed into GGUF_Flux/Florence2 (i2i/t2i) to produce the final image.

[wispy nest]

flux colours so nice

[steep sleet]

Have you produced a Flux version for RES4LYF?

[wispy nest]

yeah he did, it works on flux now 🙂

[steep sleet]

Anyone got a w/f or embedded png at all?

Cascade RES4LYF produces extraordinarily unique looks - and 11 minutes/image on my 8Gb VRAM PC 😄

[wispy nest]

here you go
just delete other stuff in the workflow if needed

[steep sleet]

Thank you

[wispy nest]

probably should delete Mega Modifier node its weird

[steep sleet]

Florence2/GGUF_Flux is so cool!

[wispy nest]

yeah I've been using Florence2 since the day it came to Comfy

as well as Moondream and WDtagger

[steep sleet]

Clownsampler and Sharksampler fail to load ...

[wispy nest]

yeah they aren't in the manager
but there are install instructions here https://github.com/ClownsharkBatwing/RES4LYF

[nimble mason]

that's a lora

[steep sleet]

i2i Florence2/GGUF_Flux

[still rose]

FLUX is my fav model

[steep sleet]

i2i Florence2/GGUF_Flux

[dry blaze]

Getting somewhat hang of kohya for SDXL, and training just isn't quite there yet

This was the result using ai-toolkit and flux. Instant perfect result lol

[jagged badger]

made with leonardoAI's anime XL, still can't find anything not commercial close in terms of style and flexibility

[pliant prism]

Ive tried i think everything and i just cant find a way to do it.

[vagrant dust]

Something like this?

[pliant prism]

YEAH this is wayyyy better than anything ive ever gotten

how did you manage to do it

[vagrant dust]

Used the ipadapter and a simple juggernaut sdxl model.

[pliant prism]

hmm odd ive tried it before but I cant seem to make it work

maybe im missing some models?

[vagrant dust]

So the whole workflow to create it was:

1. used a LLM to describe me your drawing in a prompt. Result:
   "A cute anime girl with long, wavy orange hair tied into two large pigtails. She has light skin and is wearing glasses with round lenses, giving her a studious yet playful vibe. Her outfit consists of a lavender sailor-style top with dark stripes on the sleeves, paired with a matching pleated skirt in a muted purple tone. She has an expressive, confident stance, hands clenched into fists at her sides. She also wears knee-high purple leg warmers over dark shoes. Her eyes are large and sparkling, in shades of light purple, adding to her charming appearance. The art style is clean and simple with bold outlines, in a vibrant and light color palette."

2. Used a comfyui workflow with a simple JuggernautSDXL Model. (changed the prompt to old woman instead of girl). Added the ipadapter and used your drawing as Composition Input and one image from the Lora you linked (the one with the organge Hair) as source for the style.

If you would take for example the black haired style image (more glossy 3d render look) you might need to change the hair color in advance because it might come out black instead of orange otherwise.

[pliant prism]

would it be better if I used comfyui instead of stable diffusion?

[vagrant dust]

Should work with webui and for many people webui frontend is easier to handle. I prefer comfyui because of the structured way of the node editor.

[pliant prism]

I see thank you for the help and info btw :D

[vagrant dust]

You're welcome

[weak lotus]

hey, if I try to upload an image to cloud storage such as Edgestore using nextjs and if I send it to the image generation AI endpoint for (image-to-image) generation, I get error: input not instance of Blob

[smoky vigil]

London.

[steep sleet]

Flux1.Dev.fp8 in PortraitMaster

[formal wadi]

these are so good, do you have any tips on making almost very precise anime character fan arts? I'm new to ai

[steep sleet]

PortraitMaster+Flux

[little dagger]

The details here are amazing!

[jagged badger]

sorry, I don't usually do fan arts, but you should consider loras etc trained on specific character

[formal wadi]

Oh okay, thank you

[scenic yew]

You can try flux, it’s pretty good at prompt following and text rendering. For example Naruto in dragon ball style.

[steep sleet]

Yes, Flux is like that!

[formal wadi]

Isn't flux a paid one. Can it do multiple art styles? Like I'm going for a bit of a anime realistic style which is clean. like this

[sterile kiln]

Try with this one 😄
https://civitai.com/models/818292?modelVersionId=915018

[vestal vortex]

❤️

[orchid elbow]

some images made with scribble controlnet on SDXL, still tuning the parameters

[scenic yew]

Flux(dev and schnell) is open source like stable diffusion, you can run it locally. There is also a "pro" one but its pretty similar to flux dev and closed source. The image I created was from the base model(not for anime but general) and still performs pretty good, you can try out an anime finetune as well.

I got this from quantized Flux.1 dev, you could probably get better results from some anime finetune of flux.1 dev

[languid pebble]

Quickie ...

[formal wadi]

Oh dam that really looks pretty good for general without any anime art style checkpoints. How did you manage to make luffi exactly like the one in anime with sll the details like scars and stuff. Lora?