🏞｜general-with-images

[languid pebble]

😄

[keen sphinx]

/generate

[opal portal]

Create a beautiful fertility hospital logo realistic

[fleet goblet]

me over here on a a770 generating hunyuan video gens

they must've fixed hunyuan's vae decode tiled for native

they added temporal size to the native node

[dry blaze]

Oh right, was meaning to try that one as well. Could you share the workflow you use? Or the git to the nodes needed?

[fleet goblet]

The nodes are native.

[dry blaze]

Also, need to one day start messing with comfyui on deck as well

[fleet goblet]

it's a unet model so you load it with diffusion model loader

that alongside emptyhunyuanlatentvideo

[royal monolith]

that's in the workflow on the blogpost that comfy just made

[fleet goblet]

and modelsamplingsd3

[dry blaze]

Could you share the workflow and link to the model you use? As i can then go from there

[fleet goblet]

You could also just go to the comfyanonymous examples and get it

because it's there

https://comfyanonymous.github.io/ComfyUI_examples/

[dry blaze]

Oh,. there, ghotcha xD Was looking in the announcement channel for some reason lol

[royal monolith]

here's the workflow. you'll need to open this in the browser, then just save it and load it into comfy - dont' try to save it right from this post - click on it, open in browser, then save it. https://blog.comfy.org/p/running-hunyuan-with-8gb-vram-and and follow all the instructions in the blog post for models, vae, etc

[fleet goblet]

If you are running on consumer hardware, utilizing hunyuanfastvideo's lora alongside a q4 gguf model is recommended imo

https://huggingface.co/city96/HunyuanVideo-gguf/tree/main

https://huggingface.co/Kijai/HunyuanVideo_comfy/blob/main/hyvideo_FastVideo_LoRA-fp8.safetensors

[royal monolith]

i'm on a 4060, and the models used in the workflow work... okay... but are fairly slow. so gguf might be a better way to go

[dry blaze]

thanks :) Cause i've been trying to test the new longer 10 second 1.1 model of gguf, but still can't for the life of me get it to make videos for longer than 2-3 sec

[royal monolith]

you running out of vram?

[dry blaze]

Nope. Just turns out to be that short

[royal monolith]

i can't do longer than that with 16g vram as i run out

[dry blaze]

currently interpolating a tv show episode, so gonna try in roughly 30 min

[dry blaze]

what parameters do i change to make it longer?

Just total frames?

[fleet goblet]

total frames

[dry blaze]

Gotcha. Gonna test afterwards and report back :P

[fleet goblet]

The model was trained on a framerate of 24fps

so it must be 24+1fps

an endframe

max i've done is 384x256 at 121fps

[royal monolith]

also look at the blog post and try changing the various settings if you do run out of memory

you can always change the speed/FPS in davinci resolve if you need to

[dry blaze]

Fps i change with RIFE node. Or just pull it through flowframes

[royal monolith]

it also doesn't work well at an AR that isn't what the comfy workflow is using

at all

[dry blaze]

Same thing i'm doing here actually. Just either do within workflow, or separate with flowframes

[fleet goblet]

Comfyui has RIFE and Film VFI as nodes.

I combined that alongside a 2x fast compact upscaler (nomosuni compact 2x)

[dry blaze]

Oh man, gif dithering xD

[fleet goblet]

That isn't a gif

also

the original video is 512x384, so yeah

dithering gauranteed when both interpolated and upscaled

[royal monolith]

that happens when you change the aspect ratio. it doesn't like most AR

[dry blaze]

Huh, send me that workflow, wanna see if i can make it less artifacty out of curiosity :P

[fleet goblet]

this isnt true, like LTXVideo it comes down to total prompting.
Hunyuan video has a spatial decoder that allows it to handle much lower resolutions

AR is fine as long as it is nearing 16:9 or letterbox ratio

idk about 9:16

[royal monolith]

it's not? that's odd, since that's what happens with EVERY SINGLE TEST i ran yesterday with an AR that wsn't what is in the comfy demo workflow

[fleet goblet]

And then here's me at low resolutions generating coherent walking animations

[royal monolith]

but you know best, as always

[fleet goblet]

even at 384x256

[royal monolith]

@dry blaze i'll be on the L3 discord if you need me

[fleet goblet]

Lol he can't defend himself at all

Every time, man.

Here's me generating zombie knights

[dry blaze]

Is there a tool out there similar to tensorRT, but for models that isn't SVD to speed up generations for images and videos? Like animatediff, gguf and hunyan?

As at least with images, other faster processing nodes exists, such as AItemplate, but it seems to be broken on every new confyui version for some reason. but curious if aitemplate can be used with older video models

[dry blaze]

No idea what the L3 discord is

[fleet goblet]

Don't bother.

Literally every time I've conversed with that man it becomes an argument.

I'm not the only one either.

Here he comes.

[royal monolith]

mostly because you're wrong every time.

[fleet goblet]

You leave every time without affronting your argument in any way

[royal monolith]

love how you argue with the programmers

[fleet goblet]

I don't care that you can code. I clearly am getting outputs that are operational at lower resolutions

You can't disprove that, so leave and go to L3.

[dry blaze]

Oh no.. what did i start lol

[fleet goblet]

You didn't start anything.

This guy just has ego.

[wispy nest]

torch.compile

[dry blaze]

Oh well. Well, i'm not knowing of the discord he mentioned, so i'm stuck here anyways

[fleet goblet]

He's generally disliked anyways 🤷‍♂️

[dry blaze]

For aitemplate?

[fleet goblet]

I've had DMs from other people regarding him

[royal monolith]

you started nothing. i tried to help you, the resident troll lord jumped in to start a fight and arguement per normal

[fleet goblet]

I'm a troll supposedly

If you actually had any argument maybe you'd stay a little longer.

But you don't.

https://tenor.com/view/baby-seal-sad-cry-meme-seal-gif-19608082

[wispy nest]

no its an alternative

if you are able to, using deepcompressor for 4-bit quantisation is either the current best or close to it
torchao int8 AoT compile is also very strong

[fleet goblet]

Time flies.

Fish flies.

Dragon Flies.

[dry blaze]

holy crap, i need docs on how to do all of that, and read a "explain like i don't code" article of what each is/does, as i appear to be quite out of the loop xD

[wispy nest]

so long as you have pytorch 2.5 cuda 12.1 or 12.4
and the model is a DiT not a Unet
then torch.compile is easiest and fastest to do

[dry blaze]

Quant part i know at least. Making lower bits of it to save space and time by making it "less "lossy", and more "sacrifice slight visual for speed and size"

[fleet goblet]

cuda moment

i cry inside

[wispy nest]

oh yeah without cuda things are rough

[dry blaze]

Crap, outdated pytorch. 2.3.1+cu121

[dry blaze]

Especially when you try to run comfy on steam deck via distrobox because arch xD

[dry blaze]

What can you quantisize? Any model? And how small can you quant them to? As if i can quant to any smaller, i wanna have some fun xD

Also, just realized, sorry for mad multiping

[wispy nest]

flux 1.58 bit just came out like an hour ago lmao

you can quantise most models in fact

[dry blaze]

Oh sweet! I guess i know what i'm gonna have fun doing xD

[dry blaze]

What's the main difference between these though?

[wispy nest]

deepcompressor has a special new method for dealing with the large amounts of outliers you get if you quantise to 4 bits

[dry blaze]

Oh neat! Though, it failed. Host key verification failed. upon trying to install with poetry

[south temple]

I appreciate you guys working on video 🙂

They have released a vid -> vid but its not great from what I see

[light tusk]

zebra

[south temple]

that is a car

[fleet goblet]

384x256

so low a resolution

yet still partially coherent

[south temple]

bruh these files sets are HUGE lol

for LTX

[still rose]

Pretty cool

[unborn gorge]

test - @fresh venture (fast)
-# Create, explore, and organize on midjourney.com

Hey @worldly quest
Use the buttons below to subscribe, or to link an account that already has a subscription.
[Please check the bottom of the channel for more information.](#🏞｜general-with-images message)

[robust sparrow]

My pizza emoji is so hot it became lava 🔥
-# flux.1-dev Q8, open-genmoji lora

[verbal lagoon]

Total chaos. Scattered data, messy inventory, delayed orders, and stressed logistics

[clever oar]

new year cake

[spring mist]

So Long, and Thanks for All the Fish

[wispy nest]

shuttle3

[leaden charm]

Dragon flies cinematic effect colour splash

[candid kindle]

Create a similar image

[hidden tusk]

New technique / AI Flipbook style animation - https://www.youtube.com/watch?v=e-F7rtctxHs

[stuck nimbus]

New year tree

[crimson dawn]

@night crescent found this mentioned in the project's readme. maybe this offers insight.

[night crescent]

oooh, thanks, that was it

[eternal sand]

Photo via Kodak portra 160,young pretty girl, 20 years old. She has blonde hair, blue eyes, pale skin. Split into four images, Shot of different angeles, white background --style raw --v 6.0

[spring kelp]

this isnt midjourney lol

[glacial cedar]

If I wanted to draw this image, can someone sugest a prompt?

[copper matrix]

realistic or anime

[copper matrix]

but yeah the prompt i used:
masterpiece,best quality,amazing quality,very aesthetic,absurdres,newest, 1girl, short black cheongsam, short sleeves, red with gold branch floral pattern, red floral pattern black mask, black gloves, high heels, sitting on table, crossed legs, black hair, hand near head, hand on knee, side view, cherry_blossoms,

didnt use the tag latex or plastic that your image has because that would generate a whole nother image

could modify the dress length and other details as you see fit as i feel like a normal one was a bit too long and short a bit too short

[glacial cedar]

wow!

[stoic lotus]

Hello boyz und girlz
CC-Super is now available on CIVITAI
https://civitai.com/models/1102583

[topaz moss]

Crie uma imagem Mulher poderosa de 27 anos, loira, cabelo médio em ondas suaves, vestindo vestido de seda azul-marinho ou preto. Postura confiante, com uma mão cruzada ou segurando bolsa minimalista. Ambiente luxuoso, pisos de madeira polida, paredes cinza escuro com detalhes dourados, luz suave entrando por grandes janelas. Móveis elegantes, mesa de vidro, pintura abstrata na parede, transmitindo sofisticação e exclusividade.

[winter lotus]

it took me 27s to generate a picture like this in 7900xt, i think it's pretty good

[heady dune]

help me to make a cartoon cute image of the parrot in the picture

[feral portal]

Create a picture of elephant

[celest sigil]

Here is your requested image...

[vernal plaza]

create a picture of messi

[copper matrix]

@night crescent something like this?

some lora's dont have a trigger word they just work

but if your using illustrious

try using: toriyama_akira_(style) (see msg below)

[night crescent]

i know, i was just saying one of them did use one

[copper matrix]

toriyama_akira_\(style\)

[night crescent]

it works with akira toriyama on illustrious, but it's not as close as the images i see on civitai

[copper matrix]

could be the model

used mistoon_anime for that one

used WAI-NSFW-Illustrious one for this one (same prompt)

[night crescent]

the DBSbroly model looks nothing like any of the pictures i see

[copper matrix]

Dragon Ball Super Broly Movie (Series Style) [Illustrious] was made on the wai nsfw one and you gotta copy their settings

[night crescent]

since it's not exactly toriyama there, it's a specific style, and the lora just doesn't replicate it, so i dunnno what i'm doing wrong

tried using the same prompt as the xemple images too

[copper matrix]

like euler A or

[copper matrix]

remade this image but instead of Eular A its now DPM++ 2m + Karras

so sampler does influence it a lot

[night crescent]

yeah but did you try the DBSbroly model?

[copper matrix]

ah didnt download it. i get the style using their artist style name (changing up my main prompt for it entirely) you get pretty close

[night crescent]

yeha without a lora i get something ressembling the style a bit

but the DBSbroly style is more unique, and i can't replicate it with the lora

like i'm trying to do cloudstrife with it, and it looks like this :

[copper matrix]

just asking

how is your prompt

[night crescent]

i used the same prompt as preview image (positive and negative), and just removed the character description to replace it with cloud strife

[copper matrix]

i think its just too syleized on the sayans in the description any image in the gallery below isnt even close to their main examples if it isnt a sayan

[night crescent]

yeah you're right, i got fooled by the special effects

i'm trying the other ones right now

[copper matrix]

this sounds like bad advice but generally if any lora is below a few hunderd downloads i lower my expectations quite a bit unless its extremely niche

[night crescent]

i've got good results up until now, but dragon ball, no deal, lol

thanks anyway!

managed to get something good with the manga style art lora, but the clothes are nowhere near what cloud should wear, even with the correct prompts, mmmh...

[night crescent]

managed something good finally

[rain gazelle]

nice, love it

I do a lot of the FF7R characters. 🙂

[restive cloak]

i made a model for emoji generation but its based on sd 1.5 not flux or sdxl or anythign so its not quite as good but im trying to optimize it to run locally on mobile devices (like apple genmoji but for noto emojis)

unfortunately android development is hard and i cant figure out how to quantize the weights

[shrewd sequoia]

/analytic visible:3d风格一张图以红色为基调，上面祝福语是大吉大利周围装饰图标金元宝

3d风格一张图以红色为基调，上面祝福语是大吉大利周围装饰图标金元宝、红包、烟花、梅花

[merry iron]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[copper matrix]

Also huge, starting price on a 5070 is very reasonable vs release of the 3070 price

[viral frost]

My taters. Deez taters is mine.

[merry iron]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[slim timber]

is there any model that generate images like these

and what type of images are these anyways? like what kind of paintings

i've seen them being used quite alot in philosophical reels and things. I find them quite soothing to see

[rain gazelle]

love these birds 😄 and the fish is cool

[rain gazelle]

scalpers are going to decimate that supply 😦

[copper matrix]

Im gunning for a founders edition

[rain gazelle]

yay Tifa. 😄 This image is actually pretty good because the background has a very midgar-feel to it with the industrial look. And her character face came out pretty well. I didn't think flux knew her very well.

[rain gazelle]

Good luck! You might have to camp-out somewhere a week in advance. 😛

[copper matrix]

Stores in the netherlands dont got such a fierce competition luckly

[rain gazelle]

ah right on

[restive cloak]

thank you 😊

[rain gazelle]

nice 🙂

I still like that painting style.

[rain gazelle]

these are neat too, whimsical...

[past pelican]

I swear it looked like it was moving.

[fleet goblet]

Hmm.

I think I have one major issue.

I need an upscale model that doesn't cartoonize it.

nope this one https://openmodeldb.info/models/4x-Nomos2-hq-atd

[wispy nest]

the Helaman ones are good

[crimson dawn]

@glacial tinsel if you want results that are somehow "more realistic" than this, you're likely just looking for a particular photograph aesthetic. not realism

[mortal raptor]

You could try prompting using words like: unsaturated colors, sepia, impressionistic, post-victorian, sketch, painting, brushstrokes, etc. Give it a try!

[viral frost]

Assuming you are doing a second denoising pass after upscaling, I recommend trying 4xUltrasharp. I’ve done comparison tests on a number of different upscaling models with Flux and always keep coming back to Ultrasharp even though it is old and not as fancy as the others. It just seems to strike the right balance between adding details without unpleasant patterning, suppressing oversharpening of high contrast edges, and avoiding both boosting film grain and suppressing it so much that detail is lost.

I also recommend not using the tiled VAE encode/decode nodes that are in core — they have a problem with inconsistency between the tiles that varies from run to run. The custom node package “Tiled Diffusion & VAE for ComfyUI” contains what appears to me to be a better implementation, at least with fast mode turned off. The quality also degrades much less when you decrease the tile size than the core implementation.

[fleet goblet]

VAE tile decode nodes have an issue with consistency between detiling

[wild condor]

what model were you using here?

[fleet goblet]

it causes visual artifacts

I've noticed that when running both hunyuan and ltxvideo

A curious ermine weasel with sleek brown fur walks gracefully along a winding rocky pathway. The path is lined with moss-covered stones and small boulders in varying shades of gray and earth tones. Dappled sunlight filters through overhead tree branches, creating dynamic shadows on the ground. The weasel moves with characteristic quick, fluid movements, occasionally pausing to investigate its surroundings. Its long, slender body weaves between the rocks while its alert black eyes scan the environment. The scene has a natural, photorealistic quality with soft depth of field. The lighting is warm and natural, suggesting late afternoon golden hour. Small ferns and arctic wildflowers peek through the rocks, adding touches of green and color to the scene. The weasel's fur appears slightly glossy and detailed, moving realistically with each step. Camera angle: low to the ground, following the weasel at its eye level, creating an intimate perspective of its journey.

LTXVideo 0.9.1

I need to find a way to pull better prompt comprehension and motion coherence out of the model

[royal monolith]

Stable Diffusion 3.5 large

[viral frost]

The custom node I mentioned does much better than the core node. I tested both by fixing the seeds of both first and second stage and just changing the tile size of the tiled VAE decode at the end. Varying between 512 and 1536 in increments of 256 and running an image comparison routine on the results (subtracting one image from the other and multiplying the results by 10), I found the custom node to give almost identical results, only varying slightly on various details spread around the image — the differences are difficult to see even with a flicker test between two images. The core node gives much greater differences, both in the details and in large scale aspects like tone and brightness of each tile. Using tiled encode and decode is important to avoid memory issues when upscaling to large sizes and doing tiled denoise, so I recommend the custom node for highest quality results.

[fleet goblet]

I'm not using tiled decode nodes anyways

🤷‍♂️

I'm doing a full decode

and then running it through rife

[viral frost]

How are you denoising the second stage after upscaling? Flux doesn’t like being run above 1 MP.

[fleet goblet]

Ultimate SD Upscale

That tiles it by itself

[viral frost]

I don’t know what rife is. Ultimate SD Upscale is tiled denoising.

[fleet goblet]

RIFE is a frame interpolation software

[viral frost]

But a worse implementation, because the tiles don’t communicate. Using Mixture of Experts, the tiles can communicate during denoising, which helps consistency and hallucination.

[fleet goblet]

This simply doesn't matter with redux clip vision being used alongside an upscale controlnet

[viral frost]

Eh, I haven’t bothered with the controlnets. They seem to take up too much memory, slow down inference, and my workflow works fine without them. Flux is generally a lot more resistant to tiling artifacts than other models. MoE just helps it that last little bit.

[fleet goblet]

That's why I use the controlnet and redux on Schnell.

The main image is made with dev alongside 4 anti-burn nodes

3.5 cfg, negative prompt setup.

[viral frost]

If it works for you, that’s fine. I don’t mess with cfg or negatives.

[fleet goblet]

Having a higher CFG with flux helps a ton, It allows for better overall prompt comprehension, and combined with other anti-burn methods really makes the image shine

[wispy nest]

I've used CFG with flux since day one yeah

[viral frost]

I haven’t had much trouble with prompt comprehension. Sometimes Flux is a little too opinionated about image aspects that are associated with, but not stated in, the prompt, which gets annoying. But most of the time tweaking the prompt to be better helps a lot.

[fleet goblet]

Tokenization bias that comes from relative tokens

Where you have dog, so you have "tail" but it isn't mentioned

Negative prompts address that issue

even moreso when you realize having a negative prompt with flux fill

allows ridiculous levels of control

[crimson dawn]

lol so much misinformation when you do this server search query. i remember cw being one of the loudest mouths spreading "flux can't do negative prompts" ever since i got here

when used effectively, negative prompting is a massive boon and a good tool in the kit

i would fully agree that "bad anatomy" and "bad hands" are probably not going to work. they're probably relics of sd15 negative embed prompting that got popular. like there were embeds trained on those tokens specifically

[fleet goblet]

🤷‍♂️ We need a net specialist who can pull out Flux's bias tokens from its trained dataset

Either that or a hell of a lot of trial and error

[crimson dawn]

has anyone done embeddings tests withh flux? i only ever find loras on civit

[fleet goblet]

🤷‍♂️ I'm only now just trying out pytorch 2.7 dev on arc

I've tried out InstantIR and got very disappointed by it

redux from flux with an upscale controlnet works better

redux and controlnet upscale kinda beat everything imo

nothing seems to be as good at mainting coherency and semblance to the original image while also massively improving details

[wispy nest]

SD 1.5 embeddings work on Flux lol

[crimson dawn]

that might be seeming to work with confirmation bias. it shouldn't . sd15 embeddings would use clip g and flux only has clip l

[wispy nest]

sd15 is clip g?

maybe it was SDXL then

[crimson dawn]

sdxl is clip g and clip l yeah. sd2 embeddings might work on flux since it's clip l

sd3 is clip l g and t5

i think i'm getting that right

[wispy nest]

maybe yeah
I always use T5 Ella with SD 1.5 so I forget what the normal way is

[crimson dawn]

i miight have dysliexic'd it. sd15 is clip l and sd2 is clip g. But the same point should apply. flux uses open clip

[crimson dawn]

we're all long past the normal ways 😄

[wispy nest]

I hope companies don't do multiple text encoders again

yeah you can use Qwen with flux now lol

https://huggingface.co/Djrango/Qwen2vl-Flux

its the vision version too

[crimson dawn]

LLMs are losing tokens if i seen the research right. i probably haven't. But maybe .. maybe the next gen of image models won't have locked in tokens either

just an abstract notion that i have no specific idea about

[wispy nest]

patchwise or continuous tokenizers are a thing yeah

[crimson dawn]

i saw a youtube, where a guy was talking about the current limitations of image generation models and he brought up an interesting example. all clock faces have 10 and 2 hands, no matter what time you prompt it to show.

[wispy nest]

hmm not sure I had a look on Civit, searching "clocks" for Flux images and there was a range

[crimson dawn]

i'm seeing that too. i did the same look see. tehre's a ton of 10 an 2 though

https://civitai.green/search/images?sortBy=images_v6&query=clocks

[wispy nest]

if you put the flux filter on you see improvement

[wild condor]

hey can anyone here point me in the right direction for making images like this? https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcQQ8lJN_IGxOPJl1iHT2wa_X3ON1QeJgTjbTA&s

I just wanna use imgpaint or whatever it is to do this for my friends and myself, i think is funny

[crimson dawn]

good call

[tacit wigeon]

Hi all, does comfyui create images with simple prompts like Fooocus? I’ve been trying AUTOMATIC1111 but I can’t get consistent results or even anatomically correct images. I just started using Fooocus but it seems a bit slower to me although it’s very easy to create good and consistent images. Also, the stop or skip buttons are unresponsive for me.

Thank you in advance!

[fleet goblet]

Comfyui itself does not enhance prompts on it's own. You'd need a node that gives it that function.

In this case you could use flux prompt enhance node, llm_party, ollama prompt text encode, etc...

https://github.com/marduk191/ComfyUI-Fluxpromptenhancer

[past flame]

Does anyone know what’s wrong with me using flux?

It only generates blank images

[crimson dawn]

what gpu do you have? using forge? you might need special launch tags. or just use swarmui instead

[smoky oak]

@crimson dawn

oop, he kicked me lmao

[crimson dawn]

familiar looking lil companion screen thur

https://www.youtube.com/watch?v=jEcQqhcAtOI you scrubed 4 of the other pc build vids . how soon till this one?

[smoky oak]

@crimson dawn

[crimson dawn]

because computers cant change ever

[smoky oak]

Bro, is there any winning this?

This guy? Hes not me lmfao

@crimson dawn

[crimson dawn]

that was a stolen vid that they were pretending was them. crafted alt persona. but you know this

they never did a real face reveal

https://www.youtube.com/watch?v=NimnGpBEEvY og video of that guy. not the dice guy

[crimson dawn]

don't blame me for whatever reputation your employers catch. thats what you should say

[smoky oak]

@crimson dawnOk, so he seems to be just as confused as I am. Anyways, dude seems to be sketchy, and I don't blame you for not liking him. He kicked me from his server, and when I joined back he told me "We don't do drama here" and then banned me before blocking me. I don't blame him tho, I must look like a fucking weirdo messaging him about something so stupid

Anyways. I am not gonna let this bother me anymore. I am not whoever you think I am, and its kinda ridiculous that you can't see that. Please just leave me out of this, cause I have employers in this server, and I don't need them thinking I am some jackass who keeps harassing you with messages for some malware ass game.

This whole thing is wild, and I wish you a good one, but stop slandering me, thanks man.

[crimson dawn]

i never brought it up before until today so i won't again. just don't poke the bear.

[smoky oak]

Not planning on it, cause I am literally not that guy 😭

[crimson dawn]

true story

[smoky oak]

oh my god lmfao

[visual hemlock]

Currently working on some stories and making characters.. I love how this turned out but I struggle to keep her facial stuff .. How can I keep her facial to another prompt with different outfits etc? I have copied the Seed number and keep the look prompt but I only change the clothing ..

[copper matrix]

Lora, you need a lora

Or photoshop the face in

[visual hemlock]

A lora for the face?

[copper matrix]

Yeah, character lora

[visual hemlock]

This is an O'C Character so only doing it in prompts.. So I cannot use the Controlnet to keep the face or something then.

I have to make an Lora just for that character's face? But if I can't keep the same face when doing the character .. I can barely do a Lora with it as I need couple images that are the same?

[copper matrix]

Hmmm there is something. Its Not optimal and i haven't used it befkre

[visual hemlock]

So if I make 20 characters (high number for now) I need a lora for each?

[copper matrix]

Havent watched the video but basically here

https://www.youtube.com/watch?v=TTk00FLc2kY faceswapping

But hair is still a thing

[copper matrix]

Theres a reason visual novels mostly have only the main cast and the rest as faceless males/females tbh

[visual hemlock]

So I should just try to make like 50 images and take as many as close as possible with the face and do a lora? 🙂

[copper matrix]

I would recommend trying ipadapter but training many loras takes a lot of time

And if you do it on civit, some money too

[visual hemlock]

Yeah, I am trying to learn how to do it on my PC but keep getting errors etc, trying to follow a video but i fail haha 😄 I manage to make 1 lora but don't understand koyas_ss yet. But I have not done much more, but maybe time to do it 😄

[copper matrix]

Its definitely a learning curve but projects like these always take time

[halcyon trellis]

I'm experiencing the issue (ComfyUI). Any help?

GroundingDinoSAMSegment (segment anything)
Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument mat2 in method wrapper_CUDA_mm)

[cloud adder]

You guys like hoomans more so here...

Some SwarmUI result

[undone torrent]

Oscuro panorama

[tawny forge]

”Let The Damnation That Is Your Failure Die With Your Blood”

[viral frost]

Surprise!

[strange crater]

image of a horror woman

[smoky vigil]

we also like kats.

[cloud adder]

I mean the alternative would be furries in my case :D

[proven citrus]

Hi everyone! I found https://deepai.org/ that allows me to select a model and generate good images (example attached), but I have little control and I cannot customize. Every time I generate an image, it resets and the whole style changes.
I tried running Sogni, a mac app, and select some model. But then the quality is not high.
Any recommendations of an app or free online service you know I should try?

like the images @viral frost produced - how was this done? online? some service? locally? what tech?

[copper matrix]

Ah probably locally but you need a pretty good pc for that

[copper matrix]

Before i recommend anything, what are your computer specs?

[proven citrus]

M2 pro with 32GB

and a large SSD

[copper matrix]

Hmm we had people here running it locally on a m2

[proven citrus]

it?

what were they running?

[copper matrix]

Yeah stable diffusion models

Theres a whole bunch

[proven citrus]

Oh yeah, this app I tested uses SD models

[copper matrix]

But i dont think you can run the newest stuff on it like flux

[proven citrus]

3.5G

I can try

[copper matrix]

Cs1o made a great guide for apple
https://github.com/viking1304/a1111-setup

[proven citrus]

Oh yeah, I heard of Automatic1111

[copper matrix]

But if you get stuck i recommend throwing a post into the help channel as i dont know much about apple

[proven citrus]

I'll give it a try in a bit and if there's anything interesting to share, I'll send an update

thanks for sharing!

[copper matrix]

You do need certain models though to run but you can find those on civit ai under the tab models > filter

Sdxl is okay for realism. Illustrious is getting really popular for anime

Checkpoints are what you need for to start

Once i wake from my nap i can show some stuff in the ui. Just apple stuff i do not kno

[proven citrus]

Nice!

A quick question about working with a model.
I noticed that the generated image inherits trends from the model. So, if the model has dropshadow below the character, you'll get it in the generated image regardless of instructions. Am I right?

And if the character is partly cut (like you see half the figure), the result will be similar. You cannot ask it to produce a full character

[copper matrix]

Hmm style can be heavyly influenced by the checkpoint yes but producing a full character should be possible

But even then styles can be changed with lora's

[proven citrus]

what's lora's?

[copper matrix]

Hmm lets say a checkpoint does a lot of stuff pretty ok but not great

A lora does a little stuff but wayy better

[proven citrus]

total newb here. I don't even know what a checkpoint is 🙂

[copper matrix]

Like a checkpoint knows that emma watson is a girl but a lora can make her likeness really well

Ah the model sorry

[proven citrus]

Oh ok 😄

[copper matrix]

Iike sdxl juggernautXL

[proven citrus]

gotcha

[copper matrix]

Or sdxl illustrious

[proven citrus]

and lora is what then? settings? another type of model?

[copper matrix]

Basically a mini model you throw ontop of the main model

[proven citrus]

Ahhh. I'm really learning here

[copper matrix]

So lets say juggernautXL (main) + Emma Watson (lora)

To get emma watson in that style

[proven citrus]

Oh cool

[copper matrix]

Or a lora can be a concept

Such as a pose, event or item

And even a style

[proven citrus]

Oh I see the loras on civitai

[copper matrix]

If you disable the pg filter its mostly gonna be what you expect its gonna be so for beginners sake keep it on lmao

Or use civitai.green

Its the sfw variant

Lmk if you managed to install a1111 I'll rest my eyes a bit. After im free to hop on a quick call to tour the thing

Been up working since 5 am lol. (Almost 9pm rn)

[proven citrus]

Thank you! I've been up the whole night with the kids, so I feel ya

I'll give it a try and see where I get

really appreciate it

[fleet goblet]

is there not any other repo that isn't a1111

from what i know its behind the other ui options now

its been a while

[spring kelp]

maybe a concept lora lol

[copper matrix]

For mac its what i saw in the pinned msg in tech-helpdesk

[copper matrix]

my nap accidentally took too long

it was meant to be 30mins

i even set a alarm

[viral frost]

I’m running Flux dev with ComfyUI on an M3 Max 64GB.

You can find my workflow in my image metadata, but you’ll have trouble running it with just 32 GB.

[fleet goblet]

I wonder how good the GGUF quantization methods would do for mac

Since you're already on CPU

"CPU"

[viral frost]

I think some people do quantized, but I’ve never tried. MPS doesn’t support fp8, so I don’t know how it’s being done. I’ve heard that T5 always needs to be run as fp16, no matter what you’re doing. Even though bf16 is kind of slow, I’d rather run the full model and not have to think about whether or not I’m suffering from quality loss.

[fleet goblet]

Well, comfyui-gguf has a t5xxl GGUF model

If that is operable then that'd be an option.

[boreal ingot]

Hello

Goodbye

[viral frost]

What am I doing? Just, you know...lying on the street, sucking up beans through a straw.

[clever oar]

lol

sd3 always fun)

[jagged badger]

old prompts (like sd1.4 old) with xl models...

[wise turret]

Anyone know how to get the AI to actually listen to me?

Hair isn't even brown, and it doesn't have the tiny devil horns as I stated in prompt

This is what I have btw

[fleet goblet]

Have you looked at Nomos8k_atd_jpg?

That is an insanely good restoration upscale model

great at denoise

[rain gazelle]

I love running my old prompts. 🙂 I feel like my prompts were more creative back then.

[viral frost]

These are the upscaling models that I've tried. I haven't found any of the HAT, DAT, or ADT models to be better for the purposes of an interim upscale, with denoising afterward. I think this is because what makes the image good for denoising must be different than what makes it good for viewing. I was hoping that the _hq models, since they were trained on high quality images, would produce better results, but in actuality they took the subtle film grain from the initial generation and amplified it, producing very noisy final images after second stage denoising.

[spring kelp]

yell at it

[fleet goblet]

I'd use the nomos8k for a pre-inference pass into ksampler

and the ksampler will handle the details

in this case tho

its just schnell with redux and controlnet upscale

and thresholding etc

origina

[viral frost]

Well, all I know is that I tried all of them (with frozen seeds) and looked at the details using flicker tests and I always came back to 4xUltraSharp.

[fleet goblet]

Ultrasharp barely changes the image. It doesn't seem to add details.

Which is good for inference of course, given that the image itself isn't noisy

[fleet goblet]

I've used a couple that aren't on this. LDSR, InstantIR, SUPIR, AuraSR, and Latent Interpolated upscale

Generative-based upscaling methods.

InstantIR and Supir are heavily prompt dependant.

LDSR is slow but it's a good clean upscaler

in terms of just normal upscalers, I've tried nomos8k, nomosuni compact otf medium, remacri, esrgan, bsrgan, and ultrasharp

Not gonna lie, with how controllable redux and the controlnet tile is (strength-wise) it is probably the best upscaler right now just because of image retention

[wispy nest]

for flux do you use the jasper tile or the union one

[fleet goblet]

I think I'm currently using the jasper tile one.

at 0.5 strength

[wispy nest]

ah okay yeah I need to try it
been using the union one all this time
think I picked the wrong one lol

[fleet goblet]

If you are of age, simply go to Civitai and look for them there.

This isn't a place for discussing stuff like this (since i think it's very clear you're looking for nsfw content generation)

[neon valve]

Can someone assist me with training SDXL (RealVisXL v5). I've tried alsorts, but get no result, at epoch/step 0 is almost identical to 5000 steps.
Real character lora.
here's my settings in OneTrainer

I've tried learning rate 0.003, 0.0003, 0.00003, 0.000003
Masked training on/off, image repeats 10-50, batch size 1,2, 4. train text encoder 1/2 on/off. EMA cpu/off, Train UNET on/off. Change the keyword, remake the tags with different generator (WD1.4, joytag, Florence-2-large)
I'm completely stumped.

I've tried other SDXL checkpoints also, nada

4070 ti super, 20+ hours of training, nada

hair left on my head from scrathing = nil

[rain gazelle]

Who's getting the RTX 5099? It's 9 more.

Text edition, too.

[copper matrix]

im getting the 9090 TI SUPER

[copper matrix]

the side says 50990 now thats one futuristic card

[rain gazelle]

dang, early access!

[rain gazelle]

I would have rented this game from Blockbuster.

[neon valve]

I'm particularly interested in the **RTX 5099 Text **version

[copper matrix]

Cant see who the mods are but watchout chat another scam on the loose userid 1325847535612334102

[heavy python]

Dont engage em. You can spot them by DM wording, usually also links with url shorteners.

[copper matrix]

oh yeah. i dont care about engagement. in the offchance its a real person typing (i had a few that being the case) any moment wasted on me is a moment not spend on another person

[heavy python]

Fair logic

Stay safe guys

[copper matrix]

one time i got one to even block me after i kept asking for feet lmao

[visual hemlock]

Do anyone know what I need to use in the Script to be able to try different Lora strengths? Like 1, 1.1, 1.2 and so on?

[sterile kiln]

could you send your profile .json?

[neon valve]

Someone has sent me a setup that seems to work quite effectively and I don't have a specific profile, I was just using a preset one - #sdxl 1.0 LoRa

[copper matrix]

Oh no not the watermark with controlnet

I got to say ai images being watermarked is peak irony

[sterile kiln]

Dragon Ball "Super" 🤮

[spring kelp]

same reaction i had seeing that pfp lol

[sterile kiln]

THIS is Dragon Ball :
https://civitai.com/models/818292/akira-toriyama-freezer-era-style
https://civitai.com/models/1031988/akira-toriyama-cell-era-style

[viral frost]

Wat dis? Imma poke it.

[copper matrix]

juggernautXL or Flux but id delete that image homie before a mod comes and 🔨

[mild arrow]

Hi everyone,
I’m trying to place a real heat pump into a real photo via Stable Diffusion’s Inpaint.
I need it to look perfectly real, including the logos and details, but those keep getting distorted.
Any tips or tutorials for achieving fully accurate branding?

Thanks!

[copper matrix]

If you can wait till 6-7pm my time (after work) i can show you how i do it

Pic i did for lolz, ai fixed the monitor impaint edges only

[pastel saffron]

For example, I need these camera angles (photo 1) And this is the promt: 2.5D, 4k, realism, UHD, Cartoon style, white background,maximum detail, Image in the center: One-story house, overgrown with grass, abandoned, ((Side view, 40 degrees altitude 3.0)) the result of the generation is this (Photo 2)
I also tried to make a top view like here is an example (Photo 3) For promt: 2D, 4k, realism, UHD, Cartoon style, white background,maximum detail, Image in the center: One-story house, overgrown with grass, abandoned, ((View from above 3.0)) And here is the result (photo 4)
Please help me figure it out

[pastel saffron]

in general, the result of the latter can be achieved from the latter, but how to make a position that would generate a clear front or side view?
In fact, I need to make images in such a way that it would be possible to clearly set the viewing angle.

[mild arrow]

Thank you! Thats very kind.

[twin jolt]

Should I be impressed that the "stars" out the window are perfectly reflected in the mirror?

[copper matrix]

@cosmic void

its a button

[royal verge]

this chat is really awesome, love you everyones ideas and help.

[rain gazelle]

Hmm I think these were flux dev out of the box. But this is a pretty common generation (and these aren't even really that good.) likely can be done with some xl models, and certainly with loras.

Video games, boxes, consoles, book covers, Lego set boxes, etc, all come out pretty fun

[kindred raptor]

I've purchased a Macbook Pro M4 Max w/ 36GB RAM and 1TB drive (already filled up more than half). Will Automatic1111 run well on that? I wanted to check before spending a bunch of time configuring it to find out it can't handle it.

No I havent' tried anything yet. That sounds pretty cool, especially for a noob like me with big dreams! Would you have a link by chance?

[clear whale]

https://youtu.be/Tk61auVeF0s?si=Xm8bv_U-4odQsraU
had some fun inserting myself on back to the future stage in comfyui. Enjoy!

[copper matrix]

Upscale and then impaint

And it depends on the checkpoint on how the eyes come out

[worldly estuary]

vshshbsjss

[wispy nest]

i found this on web how to make this i have promps but dont work is there a checkpoint for it?

[copper matrix]

Lora probably

Or their specific checkpoint

[wispy nest]

I want it but need to know what is the checkpoint

[copper matrix]

thats like asking me what oven my neighborg used for his pizza

but if you just seek something similar:
https://civitai.com/models/716982?modelVersionId=801785

[crimson pivot]

is that a policy i.e. strictly only stability.ai images are to be posted in here. (I am running flux through stablediffusion.cpp)

[copper matrix]

Ah no, i said that comment because it was hella nsfw

[fleet goblet]

Flux de-distilled is pretty powerful.

Doesn't need flux guidance nodes at all. Can be set to a 3.5 in dynamicthresholding and 100 CFG in ksampler

hands a lil messy buit I assume I have settings that aren't yet optimal for this model.

I need to look more info up

[crimson dawn]

is it based on dev or schnell?

[fleet goblet]

Schnell.

[crimson dawn]

neat!

[fleet goblet]

It seems the one I'm using is non-commercial.

I am wrong.

There are two schnell de-distilled flux models

but it isn't this one.

[crimson dawn]

ahh okay. the license is why i ask. thanks for clarifying

[fleet goblet]

👍

honestly im probs gonna still stick with base dev since i already have a well-operational workflow for it

dynamicthresholding allows me to use negative prompts eitherway

At least until there is a definitively better model

that is open-source

[fleet goblet]

took your image, ran it through perplexity claude 3.5 sonnet to get a prompt

threw that into flux prompt enhance node to get varieties

It's generating rn

white monster with da red berries

An intricate 3D render of a small alien-like creature with pale white skin and large black eyes, its fur a kaleidoscope of red berries scattered across a reflective surface. The mysterious creature, with its elongated fingers and claws, is the focal point, while the gray background adds depth. The image is a stunning, captivating visual experience, capturing the creature's emotional connection with its natural surroundings.

only 50% done

furry alien berry grabber

Flux.1 Dev and SDXL's base resolution are both 1024x1024.

Flux being better at prompt comprehension than any current open-source model

Also

for the details

it's the sampler, dpmpp_2m and scheduler sgm_uniform

for flux.1 dev

👍

oh uh i also use like 4 cfg anti-burn nodes

to allow myself to use negative prompts with base flux.1 dev

https://www.reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/

[viral frost]

Original prompt: bizarre unbelievable action photograph of a white Buttwinded+thirsty! grinbending on a alien meatball pile diagram meal artist product, intricate, chaotic, smoky, realistic lighting, octane render, unreal engine, dreamlike resolution, best resolution isdom gumballs STERY0LDE, 4k, fisheye, 8k, cinematic lighting

[fleet goblet]

GRINBENDING

airbending but for smiles

[viral frost]

I bet you smiled when you read it!

[fleet goblet]

https://tenor.com/view/cheese-food-gif-5113834195819307982

[karmic tundra]

Thank you

[celest sigil]

Does that pencil have a pen nib? 😄

[celest sigil]

@proud dagger Can you point me towards the downloads for Cosmos clip and vae please?

I've got the model, and I think I have the correct clip (not certain), but really struggling to find the VAE.

[proud dagger]

if you use SwarmUI it'll autodownload, if you need to manually download for some reason they're here https://huggingface.co/comfyanonymous/cosmos_1.0_text_encoder_and_VAE_ComfyUI/tree/main

[celest sigil]

Thank you!! I'm using ComfyUI 🙂

[fiery ravine]

Hi everyone! I’m Valtteri Hirvonen, a professional photographer based in Finland, and I’ve been working worldwide for the past 20 years. Lately, I’ve been diving headfirst into AI art, blending my photography background with the endless creative possibilities AI offers.

These images are part of my personal experiments with surrealism and new techniques. I love exploring the space where reality and imagination collide—creating scenes that feel strange and unexpected. For me, it’s all about pushing creative boundaries, seeing where the process takes me, and having fun along the way.

If you’re interested, you can find me under my name on most social media platforms or visit my website: www.valtterihirvonen.com.

And please excuse me if this isn’t the right place for this kind of introduction—I’m totally green with Discord… 🙂

Thanks for checking out my work!

[outer turtle]

https://tenor.com/bojEf.gif

[wispy nest]

https://stablediffusionweb.com/image/25622118-robot-woman-with-removed-face-plate any idea what checpoint this is?

[copper matrix]

again lol, https://stablediffusionweb.com/#faq

so tbh prbbably a lucky result as the prompt + neg is a terrible prompt

[wispy nest]

Ok

[karmic tundra]

Quick question and I don’t know if this is a tech-support question so I’m asking it here. When I use Lora’s in forge I usually just click on it, and it post the name of the Lora in my prompt do I still have to use trigger words in order to activate it?

[rain gazelle]

If the model instructions tell you to use the trigger word, you should.

[karmic tundra]

What if it doesn’t have trigger word the Lora is called Add Detail. Do I just click it?

It’s full name is add-detail-xl

[rain gazelle]

Just click it, yep

[karmic tundra]

Also out of curiosity, have u done anime characters before, and if you have how did you do the eyes or do you just use photoshop?

I don’t know why but for some reason to me 3d eyes are easy, but anime eyes are really hard.

[rain gazelle]

Thousands of times

Not sure what you mean by "do the eyes"

You can use the adetailer plugin or generate a higher res to get more eye detail

[karmic tundra]

Prompt, Lora, or do you use Photoshop

[rain gazelle]

(or segmentation if you're using SwarmUI like me)

[spring kelp]

prompting

you can do allot with just promptig, just gotta know how to word and place it

[rain gazelle]

Yeah just prompt. Most anime models use tag based prompts, so "yellow eyes" or whatever will give you that

[karmic tundra]

Are you stuff like high definition, eyes, perfect eyes, high detail, eyes, red eyes. Stuff like that

I use

[rain gazelle]

Nah you don't need that junk

[spring kelp]

no

[rain gazelle]

Check out the #🍥｜anime channel

[spring kelp]

maybe in 1.5 sd days

that was needed

[rain gazelle]

Find an image you like and check the meta

[spring kelp]

but for xl models not rly

discord strips meta doesnt it?

thought it did

[spring mist]

if you open in bowser and save from there meta is good, the preview here is compressed with no meta

[spring kelp]

oh neat

beeen neglecting png info cos though dsicord stripped all metadata, cool

[karmic tundra]

K thank u

[spring kelp]

just simple purple eyes prompt

then adding free-u

[karmic tundra]

What does free u do

[karmic tundra]

Oh so is it already in stable diffusion or do I have to download it?

[spring kelp]

i use swarm

and comfy

its an available extension for forge i imagine

[karmic tundra]

Do I have to install it?

[spring kelp]

yep

[karmic tundra]

Yeah I had to. Thank you

[spring kelp]

yw

[karmic tundra]

? Am I doing something wrong

[spring kelp]

use prnt screen

its 2025

you dont need to send phone images lol

[karmic tundra]

😂

[karmic tundra]

😅But what am I doing wrong

Not my prompt got it from tutorial

[spring kelp]

i dont see anything wrong other than youre over prompting the negative

i think your chasing perfection from an imperfect tool

as advanced ai image generation and ai in general as come, we're still very early in the infancy of it all

take the time to learn the tools aswell will be beneficial in the long run

[karmic tundra]

I have been watching tutorials and that what it said to do. Thank you. 😆 Any suggestions

[spring kelp]

suggestions for what

[karmic tundra]

In you character eyes you have extreme details in it

[spring kelp]

my prompt for that was extremely simple
positive:(masterpiece, absuredres:1.2), 1girl, purple eyes, delicate pointy ears, upper body, close up face
negative: 3d, monochrome, text, (female child, loli:1.2)

[karmic tundra]

What model or Lora

[spring kelp]

Illustrious_miaomiao

and i use swarm

[wide tinsel]

Hi! I've been trying to use my friends who know more about image generating than me (i'm such a tech noob I don't know where to start). They've tried to generate a picture I'm looking for that I want to put on a t-shirt for a friend of mine that I know would make her so happy. But we've been unable to get close to the result I'm hoping for. Is there anyone who could help me? Preferable if you can try to generate the picture I want for me if possible, cause my friends have already tried several times and they seem to lack the skill in getting what I want. I'd owe you a favour and it would make me and my friend so happy 🙂 If you can help, please DM me and I can tell you the picture we've been trying to create. It's totally safe for work, it's just very specific and don't want to bore everybody with the details.

[rain gazelle]

Did you figure out what you were trying to do with eyes?

[languid pebble]

Transformer 😄

[potent sandal]

heh heh heh

[karmic tundra]

Quick question I want to post a few ai arts of where the AI messed up, but in a way created a masterpiece, the problem is it shows legs is it OK? There’s no nudity or anything.

I just want to show how cool the art is. Especially since it was a AI mess up.

[karmic tundra]

I put it on civiltai and it said it was PG so I’m posting it

[viral frost]

The Budgie of Madness

[karmic tundra]

Anybody know what model they used for this? I really like the eyes.

[copper matrix]

Probably used illustrious or flux

Specific model is kinda iffy to ask

[karmic tundra]

I think I found it “ Animagine XL V3.1” did a compared analysis with this image and they look similar in terms of eye shape, nose shape, and jaw shape.

[copper matrix]

Yeah but the pupils are hella off

Probs a merge?

[karmic tundra]

Between what

[copper matrix]

What brand flour did your neighbor use for their cake. It could be a merge, could be a lora

I am no ai or got the metadata 😭

[karmic tundra]

Lol I see your point😅

Do you know of any good anime models to merge with this one?

[copper matrix]

Animagine is a solid one but illustrious is the current popular model

Wainsfw or NTRmix

But different architecture but you get similar results

And it knows like most characters on danbooru

[tender sapphire]

general-with-imagesThe magic magic carpet is like in the fairy tales about Shahrizada

general-with-images

[abstract fog]

hi there
which model can draw this kind of image?
I tried several models, and even though I entered no eyes, no face, no hair, the final image still had eyes and hair. Or do I need a model or Lora specifically designed to generate this kind of character?

[karmic tundra]

Does anybody know where I can download the anything v4 model. It is on hugging face but there are two options one is download test model, and other is download the inference config.

[copper matrix]

Yes.

[idle agate]

随便生成一张女性图片

[karmic tundra]

Found it 😆

[karmic tundra]

Does anybody know what dynamic thresholding does?

[cunning hearth]

/Style: Aardman-inspired claymation style, reminiscent of Wallace & Gromit or Creature Comforts.

Scene 1: A Day Unlike Any Other

Mood/Color/Style: Soft morning light with warm, golden hues casting long shadows. Style: Aardman-inspired claymation with visible tool marks and rounded forms, in the style of Wallace and Gromit.

1.1 Establishing Shot (Bird's-Eye View) + Slow pan across the zoo + The modern zoo at sunrise + (empty action) + A high, wide view of the zoo. Curved pathways, glass enclosures, lush foliage. The Matriarch's untouched trail to the waterhole is visible. The Fox is subtly seen watching the Matriarch's enclosure from behind a bush. Squirrel scurries across the path + Soft morning light, golden hues, long shadows + Aardman-inspired claymation, visible tool marks, rounded forms, in the style of Wallace and Gromit.

[copper matrix]

What did he mean by this. You can't generate images here

[wispy nest]

https://stablediffusionweb.com/image/25622118-robot-woman-with-removed-face-plate what for checkpoint is this on Stable diffusion

[copper matrix]

Buddy i told you the last time 😭 its base SDXL. Look on their FAQ

Literally basic sdxl 1.0

No merges, no extra trained

[wispy nest]

Okay so checkpoint is sdxl

[sterile kiln]

Illustrious model?

[sterile kiln]

I did a workflow to use Illustrious img and upscale it with Flux and upscaler. Really nice

[wispy nest]

i have issue running flux to upscale with

had a bad system crash by running comfyui direnctly on windows

had to reinstall windows, and like people were talking about using comfyui within container for security reasons, i had to use windows WSL that lets you run ubuntu natively inside windows but that has a drawback with memory, so flux throws OOM when i try to refine with 2 different models

[sterile kiln]

Comfy in a container. My God, why go to all this trouble?
How much ram/Vram do you have? Your specs?

[wispy nest]

you need container for security reason, cause custom scripts can mess up your system badly

vram is 12gb, ram 32gb

[sterile kiln]

The whole world works more or less like that, I'd say. And unless you install a dubious extension, maybe...
Comfy is open source, and so are its extensions. It's all very closely monitored. Using WSL 2, a container, will just waste your time.

[sterile kiln]

I have the same. It's more than enough for Flux gguf Q8.

[spring kelp]

12gb is enough for dev

[wispy nest]

i ran flux before the incident

[spring kelp]

no need to run the crappy distilled ones

[sterile kiln]

yes it is

[wispy nest]

now not taking chances again after system crash

[spring kelp]

also container on comfy?

why

[wispy nest]

and incidentally now that i get oom for running comfyi within ubuntu i dont play with flux much now

[spring kelp]

get comfy, normally, get wavespeed node, use the first block cache at 0.12

[sterile kiln]

Crashes with this type of AI can mostly be explained by a lack of RAM, as your system uses pagefile.sys for large loads. A Dramless SSD, for example, can easily be knocked out.

[wispy nest]

not a ram situation

what happened was my windows file got corrupted apparently from malicious code

[wispy nest]

the thing is i can run flux within container too but when it comes to fixing faces with refiner which needs me to load 2 models thats when i get oom now

[past flame]

Hey guys how can I fix this problem

It appears only when I enabled Hires fix

[spring kelp]

download more vram

in all seriousness, turn down the highres

if you have it at 2

turn it 1.5

[copper matrix]

Even lower if you still get that error

[past flame]

How do I do that

[spring kelp]

it was a joke

[copper matrix]

.. you can't it's a joke

[past flame]

….oh

[spring kelp]

lmao

[past flame]

Cuz I was able to generate images with Hires fix a while ago but today it’s starting to become problematic

[spring kelp]

howm much vram does your card have

what gpu do you have

[past flame]

8GB

[spring kelp]

highres at 1.5 or 1.25

highres fix takes up a good amount of vram cause its upscaling mid generation

[past flame]

Okay I’m trying this at 1.25

[copper matrix]

||in swarm i regularly do it at 2 and when i impaint i forget to disable it so it does it again at 2||

Also 8gb gang here

[past flame]

Oh 1.25 is good for me

[spring kelp]

try 1.5

then 1.75

till you oom

trying to find your limit

[past flame]

Oh boy not even withstanding 1.4

[spring kelp]

alrighty well found your limit. try to stay within in to not oom

if you want additional upscaling at that point

youll wanna upscale it through img2img

[past flame]

Any magnification algorithms besides latent to make things easier?

[heavy python]

Found this on HF, tarot card styles

Mighta been up till 6AM saving HF URLs 🥲🥴

[pliant pumice]

QUIERO UNA CHICA DE ANIME DE CUERPO COMPLETO EN TRAJE DE BAÑO

[pliant pumice]

WHO CAN TELL ME HOW I CREATE THE IMAGES HERE, I'M NEW

[spring kelp]

go to #🤝｜tech-support and check the pins for install guides, youll need use your own hardware, importance is gpu and how much vram it has, and what brand (nvidia, amd), the pins will cover all the questions you will have for installation, anything else after that you can ask in tech support or general chat for advice on how to generate good images,

[pliant pumice]

MY MEMORY IS 16GB AND RAM 8

[lofty ivy]

has anyone figured out how to make AI comic books? im an artist working on a comic, and would love to use AI art to speed up the process but its been churning garbage for me for a year now. i even tried making a lora with my characters

like this is my hand drawn panel. step by step. i give it to stable D hoping itll complete a step for me

im using img2img with prompts calling on my character lora + comic book panel loras. these are the results:

[visual condor]

Multi-scene generation in SD has always been janky, just do it panel-by-panel, and then put the images back in place.

It will also help SD better understand your prompts too.

Tho I do recall seeing a comic panels LoRA on civitai, so maybe your LoRA + something like that can work

[lofty ivy]

jea ive got 3 different comic loras that just dont work with img2img. they can make a gorgeous txt2img but then it wont have the vision/characters i wanna use. the time i spend fixing it could be done drawing the perfect vision instead

you right let me try individual panels instead jea.

[lofty ivy]

so the panels make some AWESOME facial expressions for me to use. but the body is still horrendous lol. so im just taking the faces XD

[sage fractal]

Hi, why is the model hash different? I downloaded the model fresh from civitai, but it creates very different results and looking at the model hash it is also different. What am I doing wrong?

[warm umbra]

Anyone use this with Fooocus?

[spring kelp]

nothing. also 1.5 is extremly outdated. dont uss that if you can. use pony or illustratious sdxl models

[sage fractal]

I like the style of 1.5 models. It's more realistic than most pony models.
For SDXL I got too less VRAM.

But why does the hash change? I try to reproduce the same picture, as I did ~1 year ago with the same model, but I get totally different results

[spring kelp]

more realistic? how lol, 1.5 is the oldest base model trained at 512px, logically speaking, thats impossible lol cause xl is trained at 1024px,

hash is just version identifier

[spring kelp]

try ruined fooocus, more updated fork of foocus

[sage fractal]

Yeah, resolution-wise the XL models are better, sure. But my GPU can't handle it, that's why I am using SD1.5. For me it has good enough results and it does a good job in photorealistic images

[spring kelp]

but whats up, i use it

ah okay, make sense

cos once you get a taste of xl quality, hard to downgrade

same thing when youll play with flux

but flux makes my 12gb cry

[sage fractal]

I am not convinced of flux images. Most images I am seeing online are too clean. They probably need more fine-tuning, but nevertheless I can't use that reliable anyway

[spring kelp]

flux is amazing lol

allot of stuff wont look good if you dont prompt it properly 🙂

[sage fractal]

So how comes it that the version identifier has changed? Even the sample pictures have a different hash, than the has in the description of the model on civitai

[sage fractal]

I am talking about photorealstic pictures

[spring kelp]

cause hash is a specific version of a model, if any changes are made to it like additional training etc, it updates the hash this explains it better rthan me

https://en.wikipedia.org/wiki/Cryptographic_hash_function

in the grand scheme of things, hash doesnt play a role

in the 2 yrs ive been here answering questions thisis the first time someone's mentioned the hash to me

[heavy python]

🤓 hashings fun

[sage fractal]

I am not that deep into those topics, but hashing is nothing new for me.
I was just wondering, why I am getting two very different results, when using the same configuration and same model. The only difference I noticed was the hash.

Shouldn't be the model hash the same?

[fleet goblet]

You'd know that flux is more than good at photorealism with finetunes and loras

sampler and scheduler choice also are heavily output-dependant for flux, which dpmpp_2m sgm uniform being generally recommended for photorealism

rough example, 2nd pass would easily fix artifacts at 0.4 denoise

[warm umbra]

Somehow managed to get a VERY NSFW model to generate this.

[viral frost]

This makes me uncomfortable.

[viral frost]

Dog hands!

[karmic quarry]

what is the prompt to generate image

like / imagine

[sterile kiln]

What model is it?

[viral frost]

Yum. How appetizing. 🤮

[warm umbra]

Let me guess - "A ham shaped like a rhino beetle."

[fiery ravine]

Some astronaut fun. Created today.

[fleet goblet]

Top left image

Three legged chicken

three legged chickens

Actually yknow what

I should set up an upscaling workflow

see if i can fix it

You added graininess into your prompting, or a lora i assume

that realistic analog photo feel

uh nope i fluxified it

I need a defluxing lora

[viral frost]

Ooh, so close. It's actually a bizarre unbelievable action photograph of a large delicious detailed steamed ham on an operations couch, goblins and thelikebecomeing instructions labels to the left and using gorilla oil technology (final fantasy)ram like technology gogurt device, jj abrams et al, professional photography

[fleet goblet]

the chickens got 2 feet now

[fiery ravine]

I'm not trying to achieve realism anyway 😅

I added grain in upscale process. I like grain...

[fleet goblet]

Grain is good.

Grain gives the illusion of realism

I know you're not, though. It's just a little pinpointer.

[fiery ravine]

Yep, well noticed anyway. And easy fix.

[viral frost]

Lol, the little hearts!

[fleet goblet]

It seems the reason why my outputs were so clean was due to the upscale model I used

nomos8k atd jpg has JPG noise removal 🤷

Imma do an output with just 4x ultrasharp

[fiery ravine]

I prefer 4xffhqdat. It gives more realistic and a bit softer look.

[fleet goblet]

i wouldve expected helaman to release a decent overall upscaler in over a year by now

surely there is a better one that is more recent accuracy-wise

though tbh guess it doesnt matter if you upscale then inference

[fleet goblet]

Using that upscale model with a film grain node

on dpmpp_2m sgm uniform schnell ultimate sd upscale + flux redux + flux controlnet upscale

My result

Only problem

WE GOT MORE THREE-LEGGED CHICKENS

I might have

to bring up mty film grain by .1

just to get the same effect you're going for

had to also decrease guidance down to 2.2, where i had 3.5 as default

[viral frost]

A man always has a first love.