✨｜sdxl

[shy kelp]

ah yeah I think I incorrectly said detail when what I meant was complexity

[copper kraken]

yeah there's all diff kinds... all are valuable in some contexts, bad in others

there's apparent detail... which can be enhanced by shit like film grain, sharpening filters etc that's not really hallucinating anything real

real detail... adding complexity on a fine level that should be there... then compositional detail, which is what we're usually thinking of when we mean "add detail with an upscael" etc

[boreal bough]

"real detail"
train model on only last 100 timesteps. Then use it as a refiner model, which only takes over the last 10% of steps.
works 10/10

poor mans refiner model 🤣

[boreal bough]

looks like a legit photo of a gold leaf artwork

[vast galleon]

Thanks, i'm working on a specialized Mangled Merge for styles. In the process of merging a bunch of TBH style loras to Output Block 1 and found some new techniques that are working out really well.

[stray warren]

@copper kraken your stuff is amazing it keeps getting better and better. Makes me wish I could run comfyui (which we'll probrably get at one point in DT). Are you using your own model, if so are you sharing it somewhere?

[fierce heron]

is a 6.7gb sdxl model fp16 unpruned?

[stray warren]

Kolors

[copper kraken]

thanks! last few have been base cascade with ultrapixel. i need to overhaul these nodes completely from the ground up so this can actually work in comfyui properly and support things like loras, alternative checkpoints, different samplers, etc

this is really the floor of what we can prolly do with ultrapixel

[stray warren]

Loooks like it's getting the best out of Cascade. Does it need a lot of vram?

[copper kraken]

nope, these are all being run in something like 10-11gb vram at the peak

it requires far less vram than sdxl at the bigger resolutions

[stray warren]

Still too much for my poor 8gb, but I think our dev can make it work for the smaller devices.

[copper kraken]

oh you can do stuff on that too for sure

that's for these massive resolutions

[stray warren]

oh great!

We have tiled decoding and tiled diffusion 🙂

[copper kraken]

cascade is really memory efficient if you're using the "lite" checkpoints

usually the peak vram is with the vae decode (stage a) step but you can tile that too

def look at the full version of that one^

[golden ether]

anyone knows what it does mean black and red dot in segment anything?

[vast galleon]

If I remember correctly, the red dot is what you want to select. It's been a while since I've used it.

[golden ether]

thanks buddy

TypeError: 'NoneType' object is not iterable using controlnet openpose in SD WebUI Forge
anyone would help me, please

[vast galleon]

are you using comfy?

[spiral bridge]

Right skeleton doing sum quite odd to that frog

[peak dove]

Fooocus

[hasty smelt]

Hi there, I've been out of the loop for a while. My models are still SDXL, using SDForge. Is there anything more up-to-date or better than this? thanks

[night vessel]

Generate an image of AJ

/deram Generate an image of condo

[copper kraken]

cascade is great, espec with ultrapixel... working on that myself lately

[hasty smelt]

I'll check it out on YouTube, thanks for updating me.

[copper kraken]

Here is the image you requested.

[hasty smelt]

Can you recommend any tutorial that teaches how to install this?

[copper kraken]

can't think of any

comfyui is where you wanan start though if you don't have it

[hasty smelt]

Yes, I have ComfyUI, but I liked using Forge. Is it available there?

[copper kraken]

ultrapixel isn't

[hasty smelt]

I have 8gb vram, is possibile to rum?

[copper kraken]

you could prolly gen something smaller than i'm doing here

[hasty smelt]

Jesus Crist, this is crazy

[copper kraken]

cascade is pretty mem efficient

def more efficient than sdxl if you use the lite checkpoints

[hasty smelt]

I'm watching this video, and it looks like it will install UltraPixel directly in ComfyUI. Can I do that? https://youtu.be/4Gf9lkgZ0SE?si=sfgAxOn54kGwki04

[copper kraken]

no idea with the video but if you git clone the 2kpr repo then swap in the stuff i addressed in a pull request you should be set

https://github.com/2kpr/ComfyUI-UltraPixel

https://github.com/2kpr/ComfyUI-UltraPixel/issues/28

download what i converted, those two safetensors, into models/ultrapixel

[copper kraken]

then replace the python files in the custom_nodes ultrapixel folder with these

the workflows that use these are embedded in the images above

@willow stone

[hasty smelt]

@copper krakenIs this delay normal the first time I'm rendering something? Or is this a mistake, something wrong?

[copper kraken]

yeah it's downloading the checkpoints

again just make sure you used the .py files i gave ya and the safetensors in models/ultrapixel

mitigates potential security risks cuz the ones they provided weren't really safetensors

[limber kindle]

Numpy 1.23 eek, will that mess up other stuff?

[copper kraken]

no idea

it didn't fuck up my install

i don't have a ton of shit on here though

[unique field]

Can you use SD 1.5 Loras with SDXL models or even sd3?

[limber kindle]

Ok, well will try it out in my dirty comfy that is spammed anyway

[copper kraken]

i'd just duplicate your comfyui venv, and if it gets fucked then just revert

lol nice

[limber kindle]

I have a clean version too, so will see : D

[tawdry current]

iirc, numpy 2.0 has been causing a lot of headaches lately. I wanna say even comfyanon rolled back a req on it to like 1.26, but I'd have to double check

[limber kindle]

Yeah true, numpy 2.0 single handedly killed my entire VoC env lol

[tawdry current]

I know a lot of other things have been rolling back, so it's possible I'm getting it mixed up. But all I know is I've seen it in quite a few repos lately for various AI related things

[limber kindle]

Ye, but numpy 1.23 in this case should be fine

[limber kindle]

UltraPixel raw renders, Idk if it'S worth it to pixelf***k it by downscaling and upsampling with a detailer again? : D But so far it doesn't like complex prompts

[shy kelp]

prompt adherence is not gonna be cascade's strongest area

I mostly use jugger cos it has one of the best prompt adherences

but sometimes losing some prompt adherence for looks is okay

I would use kolors as an example cos it looks great but can't follow prompts that well

[tawdry current]

Kolors follows prompts really well if you prompt correctly. The problem a lot of people have with the newer style models is they still, prompt, like, this, rather than using correct grammar with them. They are using datasets where large portions of the data are captioned with natural wording. Long story short, 80% of the time, it's poor prompting

Using an LLM to take some 1.5 style tag prompt and expand it into a real prompt, can help alleviate a lot of the issues people have with adherence in the newer llm style models

[shy kelp]

for context I am one of the worst prompters around
I've just never paid attention to learning prompting properly
so yeah good prompting may get way better results than I did

[shy kelp]

I do use natural language rather than tags though
for the encoders that take natural language

so at least I avoid that mistake

[vast galleon]

So I'm at the 2750 lora mark of Mangled Merge. I've spent the last 450 merges into merging only style loras focusing purely on 2D going 100% into Output block 1, and 50% into input block 8 and output block 0. The model works fine between 2 and 8 normally. It's lost some of it's photorealism (on purpose), but something I just found out in the attempt to find where the best photorealism CFG mark is, that CFG .5 with FreeU and PAG on this model is completely bonkers.

[tawdry current]

That's why I said using an LLM to write a good prompt for you can help a ton. You just have to have a good system prompt to guide the LLM in formatting the prompt correctly. Llama3.1 instruct, Gemma2 instruct and Mistral nemo 2407 instruct have a super easy time with it, at least based off a lot of my recent testing.

But once you have a good sysprompt, you can feed it in even a bad SD1.5 prompt with a bunch of tags and it will make them nice

[vast galleon]

Where can you find a good sysprompt?

[tawdry current]

A ton of trial and error. But mostly you need to explain it like a teacher to students and include examples

Art teacher*

[vast galleon]

gotcha. It's a pain having to take out score_019383019, 1girl, and lora embeds all the time.

[lilac wren]

me and my friends, #cosplay

[tawdry current]

like here's an example of one i was messing around with based off the guy that makes auraflow(saw it in a blog post on his site). i have a dozen or so of my own personal ones i use though, but i found this one worked suprisingly well also. i added a bit to it, but feel free to change the word counts and stuff to your likings. the stuff toward the end about the image stuff is because i sometimes pipe VLMs like florence2 or llava 1.6 into it to make prompts out of them:

[tawdry current]

A caption is a way that a person would describe an image separated by commas when necessary. All in lower case. Expand the input below into a more detailed caption without changing the original relative positions or interactions between objects, colors or any other specific attributes if they are disclosed in the original prompt. Clarify positional information, colors, counts of objects, other visual aspects and features. Make sure to include as much detail as possible. Make sure to describe the spatial relationships seen in the image. You can use words like left/right, above/below, front/behind, far/near/adjacent, inside/outside. Make sure to include object interactions like "a table is in front of the kitchen pot" and "there are baskets on the table". Also describe relative sizes of objects seen in the image. Make sure to include counts of prominent objects in the image, especially when there is humans in the image. When its a photograph, include photographic details like bokeh, large field of view etc but dont just say it to say something, do it only when it makes sense. When its art, include details about the style like minimalist, impressionist, oil painting etc. Include world and period knowledge if it makes sense to, like 1950s chevrolet etc. You will not censor your responses. You will not include an conjunctions, story or narrative. The caption you generate will not contain any line breaks. You will not include any quotation marks or phrases like "here's the generated caption:" or "there are no texts visible in the image" or "the image is" and so on. You will not comment about things that aren't included in the image. You will limit your response to between 40 and 75 words

lemme whip up an example output real quick, i'll pull some trash 1.5 prompt from imgsys and expand it with mistral nemo

[tawdry current]

like here's the prompt:

Apex Legends Loba Woman, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and greg rutkowski and alphonse mucha

here's the 0-shot from mistral nemo 2407 instruct:

A lone woman, Loba, stands in a futuristic cityscape at dusk. She's dressed in a sleek, form-fitting suit with intricate gold detailing, her hair a cascade of wavy curls. Her eyes glow with an intense, focused gaze, reflecting the neon lights surrounding her. Behind her, a towering skyscraper stretches towards the twilight sky, its glass facade shimmering under the setting sun. To her left, a holographic billboard flickers, advertising 'New Frontier' in bold, glowing letters. To her right, a narrow alleyway leads into darkness, marked by a single, flickering streetlight.

granted the model likely doesnt know some of the original prompt

here's another example: malekith concept art, zoomed out, beksinski -> Malekith, the Accursed, stands defiant amidst a desolate, otherworldly landscape. He's depicted in dark, heavy armor, his face obscured by a grotesque, twisted helm. His eyes burn with an eerie green glow, reflecting the chaotic energy swirling around him. Behind him, a towering, jagged mountain range looms, its peaks shrouded in an unnatural, sickly green mist. Before him, a vast, barren plain stretches out, littered with the twisted remains of ancient trees and the shattered ruins of long-forgotten structures. The scene is painted in dark, muted tones, with stark contrasts creating a sense of unease and foreboding

[vast galleon]

thats cool. it fills in context

[tawdry current]

without having to mention beksinksi's art style, it perfectly describes it

and i'm a big fan of his art

[molten verge]

is this SDXL?

[shy kelp]

ah yeah the auraflow prompt is good

[tawdry current]

yep, but if you already have some context, it can help expand it or condense it down into more impactful and ordered wording

[shy kelp]

I need to find an LLM node for comfy

[tawdry current]

ollama node

[shy kelp]

ah ok yeah

I can chain florence 2 into it

[tawdry current]

yeah

thats what i do a lot

actually, sometimes if im patient, i'll combine the outputs of both florence and llava1.6mistral and then use all that context with a good llm to make a prompt

[vast galleon]

Here's something weirder I just found out about this model. I tried going to CFG 0 and noticed it was giving me my negative prompt. So I switch the negative and positive prompts around. Somehow, my model is working amazingly at CFG 0 with prompts switched? How?

[tawdry current]

zero means no guidance at all

dont use it

[vast galleon]

But here's the thing. If I put the positive prompt into the negative prompt, it's working.

[tawdry current]

run more seeds, it will likely flipflop btween them

[vast galleon]

nope. pretty consistent

[tawdry current]

comfy's cfg might cap to 1.0 though, i'd have to check the code

internally

[vast galleon]

I also have PAG and FreeU. but that's so weird that the prompts flip at 0

I'm gonna put some styles in the positive and see if it gives me any

[tawdry current]

yeah there's some kind of clamping going on with those nodes then, zero cfg with a default workflow will generate weird shit that doesn't follow anything

lemme flip the inputs though

ahh nvm, yeah it's working now with the neg in the pos and vice versa

at any rate, i wouldn't go that low with cfg

[vast galleon]

Tried the same prompt at different seeds, and putting illustration, 2D, artwork by sigeru miyamoto, into the positive, didn't give any illustration.

Looks good for randomness too. I'm gonna test more.

[shy kelp]

For v-pred

cfg_result = uncond_pred + (cond_pred - uncond_pred) * cond_scale

[tawdry current]

ahh good catch, so it just equals the uncond

because the other half zeros out

[vast galleon]

so whats controlling the uncond_pred?

[tawdry current]

negative prompt

neg prompt=uncond and pos=cond

[vast galleon]

ahhh. ok that makes sense now. weird.

[tawdry current]

that's why it's working though, but yeah, it should never be zero, if yout set it to even 0.01 it will allow the positive to work again

[shy kelp]

negative prompt yes
and stuff like control net, IP adapter, SAG etc can also go via negative conditioning

[tawdry current]

think they all get cfg multiplied before sampling though, so they'd end up being zeros as well, except their negative prompt portion if you use advanced nodes (for cnets i mean)

[vast galleon]

is it possible to go to .01 in comfy? Im trying, but it keeps rouding to 0

[tawdry current]

yeah in the settings you can change precision

[tawdry current]

cfg will snap to two decimal points though on stock ksamplers

so 0.01 will be as low as you can go

[vast galleon]

found it. at CFG .01 with the prompts normal it's going back and forth between the negative and positive as you described.

[tawdry current]

but yeah, use a normal cfg. honestly, don't go below 1.0. think of 1.0 like 0

[vast galleon]

I just like to experiment 🙂

[tawdry current]

yeah for sure. at 1.0 cfg, you have no negative prompt

[tawdry current]

it will also diffuse much faster too since it's not having to handle unconditioning

like stable cascade used 1.0 and 1.1 for cfgs i think

[vast galleon]

These are at CFG .01 with positive and negative equal.

[tawdry current]

right, but some of those other nodes as likely doing clamping stuff on the cfg

bypass them and you'll see what i mean

[vast galleon]

yeah, it probably wouldn't work without PAG and FreeU

[tawdry current]

so what i meant earlier was that you cfg isn't your actual raw cfg

like if you use automaticcfg, for instance, the cfg on the ksampler isn't your cfg anymore

and PAG is cfg+pag amount

so by default, PAG wants to be 3.0, so 0+3=3

[vast galleon]

makes sense

[tawdry current]

which is another reason why a lot of people dont like pag because they will use a cfg they normally use like 7, with pag 3, it makes an effective cfg of ~10

so your target cfg with pag is target=cfg+pag

so if you want 5.0, with pag of like 2, then cfg would be 3

and so on

[vast galleon]

might be better for me to just stick with FreeU then.

[shy kelp]

FreeU really is a free lunch yeah

[vast galleon]

These are CFG 1 without FreeU and PAG.

[tawdry current]

i personally dont care for freeu, automaticcfg and/or pag is pretty much all ill use anymore with sdxl. freeu just doesn't do the right things for me i've found.

[vast galleon]

These are CFG 1 with only FreeU

[shy kelp]

FreeU is just really subtle and weak

its not a very powerful tool

but sometimes it can help

[vast galleon]

what do you guys use?

[shy kelp]

FreeU can also harm

[tawdry current]

yeah look at the hand

turned to playdoe

[vast galleon]

I've found that with the right settings. This is how I usually keep mine.

[tawdry current]

if you want some real fun though, start messing around with beta noise schedule. shits cash

[vast galleon]

I do want real fun lol. I'm gonna look into that.

[shy kelp]

beta noise is nice yeah

in my experience it didn't beat GITS/AYS

[vast galleon]

where can I find info on that?

[shy kelp]

but it might with other images

[tawdry current]

helps a ton for the early steps and fine details at the end. it's in the default ksampler now

click scheduler and it's in the lsit

[shy kelp]

outside of test images I use adaptive samplers anyway

cos that just removes any sigma issues

[tawdry current]

yes and no, they still use sigma curves afaik

i dont really mess with adaptive samplers that much though

[shy kelp]

they take it as input but they ignore it

[vast galleon]

I guess I need to update my comfyui. gonna try this out.

[tawdry current]

yeah do that. obviously, don't expect magic or anything, i've just found that it works excellently for images in the 30-35 step range, which is all i do

[shy kelp]

oh yeah it might be good for 30 steps

I was testing it as like 12 steps

where AYS/GITS do better

[tawdry current]

oh yeah for sure

those sigmas are centered around that specific task

using that malekith prompt example from earlier, this is hunyuandit1.2. fucking nailed it(helm might be one of those face helms)

[shy kelp]

I kinda only do two kinds of generation
12 steps UniPC GITS in a few seconds
or adaptive implicit midpoint, adams or RK4 for nearly an hour

I don't really like the middle ground

hunyuandit is good sometimes yeah

[tawdry current]

ive been surprised with the 1.2 version of it

it's pretty good for art

[copper kraken]

i think that might be AI generated

[molten verge]

ok, so its definitely very well conceiled

[gray swan]

there is lot of lighting in the photo, do you suggest any keyword/ method to lower the lighting? I am trying to achieve iphone camera quality picture, but cant seem to work with lighting, I am using realvisXL 4.0 lighting (baked)

[shy kelp]

IP adapters with photos with a lighting level that you like

[hasty smelt]

@copper krakenHi, my knowledge is quite basic, I don't know how to apply that command you showed me. I still can't run Ultrapixel, but also, I'm not sure if my computer can handle it. I have 64GB of RAM 3200 2x32, 2070SuperFTW3, and a Ryzen 5700X. I really want to create images like yours, but if the 8GB of GPU limits me, do you think it's better to keep using Forge while I wait for a LOW GPU version of Ultrapixel? Forge benefits from low GPU, which is my case.

[gray swan]

oh, I will try

[spiral bridge]

Cool what model is this

[thorn pine]

cool

[vast galleon]

It's the next version of Mangled Merge I'm working on. Hoping to release sometime this week.

[spiral bridge]

W

Can't wait @ me when it comes please

[vast galleon]

Will do. ☺️

[spiral bridge]

[copper kraken]

such cute spiders!!

[lusty wolf]

Spider I'm your father.

[glacial star]

One message removed from a suspended account.

[surreal narwhal]

Lion

[deft goblet]

XXMix_9realisticSDXL Model + ADetailer (person_yolov8n-seg, face_yolov8n, mediapipe_face_mesh_eyes_only) + 4x-UltraSharp Upscaler

[dense sequoia]

NFT? That's still a thing?

[rose canopy]

is this also ultrapix?

[meager canopy]

No, this is SDXL

It's using my image2image detailer. It started with this image

[meager canopy]

Now that's a dirty girl!

[tawdry current]

As a depth map? Looks like what I'm always talking about when I tell people to just mock up a simple scene in blender and export a depth pass

[meager canopy]

No, just as a starting image

[tawdry current]

Ahh yeah looking at it now, it's an ambient occlusion map anyways

But if you have the 3d scene, you can export a depth map and even a basic render with flat colors as eell

[meager canopy]

I don't, and there's no need when I can create images like I have above.

[tawdry current]

I see

this is what i'm talking about btw. in ue4/5 you can make an extremely quick postprocess volume material and get perfect depth maps if you want and even adjust the cutoffpoint distance.

obviously, render it out to the resolution you want and jack up the antialising on it. shit works really well with the union cnet

[plain snow]

Looks nice, what checkpoint do you use for this?

[meager canopy]

Boltning

[tawdry current]

here's an example of using the technique with the union cnet (i like making fake cryptid/ghosty stuff)

UE5 is great for it because you can get a shitload of free assets off the quixel bridge and mock up scenes you want. props, materials, characters, etc. there are also meta humans you can make and drop in, pose and so on. so you can get a ton of control over the scene and then use SD to do its thing with cnets and/or just img2img at some denoise amount

(for people that want strong control and direction over a scene they want to make)

[meager canopy]

Sounds like too much effort for my liking 😄

[tawdry current]

well i've been the vfx world for decades, started doing 3d shit as a teen around 2000. i can tell you that this hybrid approach absolutely rocks compared to having to manually model/texture/tweak everything for a 3d scene

but i guess not everyone is down with it and that's fine too

just like to show options and workflows for people that might have similar skillsets

[limber kindle]

Reminds me of There Will Be Blood

[limber kindle]

Depth is probably more interesting for AD or so. Actually not even sure if comfy would clip 32bit depth renders to 8bit

BTW also coming from. VFX

[peak dove]

Fooocus using 4 x Image Prompt - Checkpoint = NightVisionXL

Styles Used

[peak dove]

Mr Originality - nice to see someone breaking barriers and exploring beyond the typical!

[tawdry current]

Yeah idk, but 256 levels should be enough for diffusion since it's a "fuzzy" process anyways. In theory, people could train a 24 or 32 bit grayscale depth cnet model.

I was actually talking about that with some peers a couple months back for terrain generation. Literally make a diffusion model that spits out height maps with good erosion and flows. The RGB channels would be depth, flow and deposit. But I don't have a handy cluster of a100s and thousands of man-hours for captioning height maps to train a model from scratch :/

Vs using apps like gaea, world machine and world creator

They become painfully slow once you get into the 8k range. I wanted to make some kind of model that would take the course output and then use diffusion to finish the rest

Obviously it could be expanded to other tasks like making other splatmaps for things like vegetation, rocks, debris, etc

[copper kraken]

yeah i really need to try this

[limber kindle]

Yeah kinda as you said. Currently on mobile so don't have example images, but we have a setup that is ultra old-school. Took hq stuff from https://gisgeography.com/usgs-earth-explorer-download-free-landsat-imagery/ fed it into https://www.bluemarblegeo.com/global-mapper/. It's 2d kinda since sat data from above. But when you camera project ai renders in nuke over a displaced point cloud you get really cool stuff. Workstation in repair atm, but can send some images if you're interested

[limber kindle]

Powerlines 😀

[boreal bough]

@uncut steeple @smoky patrol
There are two cryptocurrency messages above this one. Can one of you remove them? ❤️

And all their posts as well (they spammed the crypto thing in multiple channels)

[smoky patrol]

Thank you ❤️

[hasty smelt]

Hey guys, is there any model that performs well with typography?

[shy kelp]

yeah SD3

[hasty smelt]

Can I use this for personal projects where I'll be paid? From what I understand on the website, I can, as long as I'm not a big company. Is this correct? Thanks.

[shy kelp]

up to 1 million

[hasty smelt]

Thanks

[vapid hill]

how do i set clip skip to -2 in webui?

or is it just 2

because it with go in the negative

[shy kelp]

not sure but in comfy there is a node for it

[boreal bough]

yeah, its just 2 as you expected

[vapid hill]

Alr thanks

[copper mirage]

Does anyone know if there's a way to run adetailer on a batch of images in a folder?

[vast galleon]

@spiral bridge https://civitai.com/models/447902?modelVersionId=684665

[shy kelp]

cute faces, how well does that model work on half body or full body images?

[shy kelp]

i see the author released v1.0, user posts on it look good but they are mostly half body

[gloomy lark]

Cascade

[coarse kelp]

which one do u think is better?

[shy kelp]

both of the image textures look good each with unique difference in perspective

no such thing as better

better would be strictly subjective opinion

[coarse kelp]

that's true

what's your opinion though?

[shy kelp]

i dont offer that kind of opinions

those 2 images have good quality in tone and texture is all i will say

[coarse kelp]

it would really help me out because 1 of those is my attempt at an improvement

but thanks either way

[shy kelp]

its up to you how you wanna apply your idea in context of what your actual objective with those images are

[coarse kelp]

my objective was to get a more varied image similar to how i feel smaller resolutions of images offer a better look variation wise but is worse quality wise

the right image is my attempt at taking a 1024x1024 image and improving it by using controlnets and downscaling it to 512x512 before upscaling it to 1024x1024 and further upscaling it afterwards

the left is a standard 1024x1024 image

[shy kelp]

there you go

[coarse kelp]

imo the background on my version is better but that's only my opinion

[shy kelp]

your objective = your opinion that will matter

[coarse kelp]

thank you :)

[shy kelp]

np, so yeah both are good quality

[coarse kelp]

here's another try with an improved background

[shy kelp]

well the right one has more depth cause it incorporates the background and foreground

im not saying better between either of the two, if you want close up left is good too

[coarse kelp]

that's the thing because a 512x512 image is cray good at doing that for some reason but it's always very zoomed in compared to a 1024x1024 image. to combat this i used a 1024x1024 image and a controlnet to force the 512x512 generation to be zoomed out, then i used that to generate a 1024x1024 image

[shy kelp]

any reason why you are choosing square canvas?

you could mix the height and width like 768x1024

[coarse kelp]

now that you mention it i saw that faces became a bit deformed when generating the first 1mp image to use as a reference for the 0.3mp image

i'm not sure why that is

the top one is for the 1024x1024 image

[shy kelp]

technically choosing 512px with sdxl model is not a good idea

[coarse kelp]

i made it so it always ends up being 1mp regardless of the width and height

but it doesn't seem to like that for some reason

[coarse kelp]

the problem happened with the 1024x1024 image before going over to the 512x512 image which is what kinda confused me

[shy kelp]

sounds like you are attempting something that goes beyond image dimension

[coarse kelp]

because 1024x1024 or anything close to 1mp should be the best but it's somehow deformed

[coarse kelp]

just something like 1080x1920 messes up faces, let me try to show it

here is a 1024x1024 image for reference

[past cradle]

Hey guys I've got a question I feel like I might not be in the right place to be asking us and so you guys can just tell me to go fuck right off if I'm intruding on your space but can I pose my question?

[shy kelp]

SDXL
1024x1024 default

Variants:
640x1536
768x1152
768x1344
832x1216

• 896x1152

[coarse kelp]

now i'm generating the 1080x1920 image

[shy kelp]

use those image resolution for basic render

[coarse kelp]

it stretches the features to fit the frame

[coarse kelp]

thx :)

[shy kelp]

which resolution did you set that?

[past cradle]

So the thing I was trying to do was take a picture of some stitches that someone had on their forehead and I added to it a little bit until it occurred to me it looked like a nine inch nails cover. And I wanted to lay the graphics over and everything and I know that sounds super simple but, I couldn't figure out how to get the rights to overlay tools trademark over the top and then just create an album cover

[shy kelp]

the base model training data set requires you apply those mentioned resolution then upscale later

[coarse kelp]

1080x1920 and plugged into something that sets it to 1mp

or is it just the resolution combination that matters and not that the final resolution ends up being close to 1mp?

[shy kelp]

yes the resolution combination matters on rendering image by pixels

you can't apply random dimension

image would look distorted

[coarse kelp]

i thought it would be fine as long as it ends up being 1mp

[past cradle]

Hey look guys I'm a little bit new to this place. I mean I got on here years ago but I just haven't been on a lot so help me out I'm not even sure who I'm talking to at this fucking point

[coarse kelp]

that clears it up 😄

thanks

[past cradle]

To be honest I'm not even sure if anybody is actually speaking to me

[coarse kelp]

i'm so confused

[past cradle]

You tech wizards are way above my pay grade but what I was doing was I took an image and I tuned it up a little bit and just for a gag I was trying to turn it into a fake 9 inch nails cover and when I was trying to find a way to overlay good vector graphics of nine inch nails graphics over that image it just wasn't working out for me because they're protected or maybe something like that or I don't know maybe you guys are better at that than this that's why I came here Jesus fucking Christ

[coarse kelp]

it feels like you're being very hostile and i haven't even done anything

i would gladly try to help but at this point it feels more like being held at gunpoint lol

[past cradle]

Not at all my friend you know in fact if I could get any help at all or if you could even lead me in the right direction I would appreciate your help I'm sorry if I came off at Host as hostile I didn't mean to come off that way

[fallow pawn]

light show inspired by Song Dynasty aesthetics, soft blue and cyan hues creating a tranquil and elegant atmosphere, gentle blue waves and reflections integrated into light projections, delicate flowing particle lines, showcasing modernity intertwined with elegance, minimalist modern design, intricately expressing Song Dynasty beauty traits, unique visual effects that resonate deep emotional connections with the audience, serene ambiance, immersive experience, 4k --version 6 --ar 16:9 --chaos 0 --quality 1

[coarse kelp]

i can't do it directly for you because i gotta go to work soon but if you could download comfyui (which is a UI that makes it easier to generate images locally on your own pc without dumb copyright stuff restricting everything) then you could make whatever you want 😄

[past cradle]

Well anyway I could show you the image that I had. It just sort of like when I saw it it reminded me of one of the earlier cover images and it just made me think how can I repackage it into a another you know faux label. Not to try to fool anyone but just to you know share with my friends who also love tool as much as I do

[coarse kelp]

sure ofc

[past cradle]

I mean I don't have any problems sharing the image with you because you could probably do a way better job of it so here you go. That is is there a way for me to just upload this right now I'll try it

[coarse kelp]

if you click this button you get a bunch of options and then there should be one that says upload file or something

this one

but i can't make anything right now because i gotta go really soon. wish you best of luck with it all though and feel free to ask if there's anything

[past cradle]

As I have said I'm I'm pretty new to this format because I haven't been back here in a long time it's probably been four or five years since I've been on here and I'm sure a lot of things have changed since then and I was already an idiot even back then

[coarse kelp]

generally the way to go about it would be to do this:

1. download comfyui
2. go to civitai, find a checkpoint model (.safetensors or similar file types :))
3. in the default workflow inside comfyui you can change the checkpoint model to another one if you download one

[past cradle]

Yay I did it

[coarse kelp]

amazing :)

[past cradle]

So do you see why I thought that would be a great background for a tool cover?

In case you're concerned, that's not pornographic. It's just a picture of my forehead this very weekend

[shy kelp]

it does look like metal yeah

[past cradle]

They blocked my last message because I said they stapled it up with metal instead of some weak floss, some weak ass loss

They didn't like the word puvvs axx floss

Anyway my stupid idea was just to take that picture that I took of my stapled up forehead and turn it into the tool album cover that never existed and not use it for any kind of ulterior motives but just to just you know have a laugh my friends and go you know hey here's an album they never released it's called the weekend when I broke my fucking head open and had to go to the hospital etc I thought it would be a great joke

[hasty smelt]

Guys, in which folder in my ComfyUI should I add the file "pytorch_model-00001-of-00002.bin" for Ultrapixel? I'm downloading this manually, since I always get an error message when running Ultrapixel for the first time, according to Chat GPT this is the file I'm not able to download.

[past cradle]

Are you telling me that this is the first time you can't find an uploaded image and I'm the one that's provided you the image that you've never been able to find before or am I miss hearing you

[shy kelp]

he's talking about something different

Ultrapixel is an AI model

[hasty smelt]

this error

[shy kelp]

I am not sure of the answer because its Cascade architecture

if you wait until the Cascade guys come online they will know

[past cradle]

Okay I'm so out of the loop here I'm sorry everybody's laughing at me Jesus Christ fuck me okay well but I am also the only asshole that split his own head open on the table and went to jail for splitting his own head open fuck me

[shy kelp]

I skipped Cascade completely and went from SDXL straight to SD3

[coarse kelp]

nah man nobody's laughing

[shy kelp]

ye we were just dealing with wub's issue

he can't get his model to work

[coarse kelp]

we're all in this mess of constantly having to learn stuff we thought we knew

[past cradle]

Hey vipraz I appreciate that message you know I love that attitude

[coarse kelp]

of course man

[past cradle]

No you don't know you don't know that I love that attitude but I'm telling you right now I certainly do love that attitude

[hasty smelt]

There's a guy here in the group, he's posting the most incredible images I've ever seen, he helped me, he recommended using Cascade within ComfyUI + Ultrapixel.

[past cradle]

Hey guys can I ask a really uncomfortable question here what's the age range can everybody know that I'm going to ask that that's I'm pretty much already know it's early twenties to late 30s maybe if that I'm not that but what I was trying to do I'm sorry I stumbled into this room to just try to figure out how to do something simple but like I'm a little lost do you mind that you guys kind of point me in certain directions maybe??

[coarse kelp]

i'm not sure how old everyone else is but i'm 21

[shy kelp]

yeah sure what are you trying to do

[hasty smelt]

I have 36, old man

[past cradle]

Yeah that's about what I expected so I'm 44

[meager canopy]

All of your Ultrapixel models should go into {your ComfyUI}/models/ultrapixel, but that pytorch model isn't one of them. The controlnet directory has a canny.safetensor file (for Cascade).

[hasty smelt]

This is not a dating group, so your age doesn't matter.

[hasty smelt]

Thanks Galaxy

[past cradle]

Okay I appreciate that I know it's not a dating group come on guys!!

But for some of us blokes we feel like you know I mean come on now I'm 44 I'm kind of not where it's happening anymore so I understand that if I'm outside of the spectrum of everything that's cool now I start feeling like you books dads and I'm I probably could have wanted to be one of you guys dads Jesus Christ that fucking oh my God I should just leap into the river right now

I don't mean it I'm not going to leave in any kind of river but you know what I mean maybe

[coarse kelp]

before you do, download comfyui or automatic1111 and see what it can do 😆

[hasty smelt]

I get this error when trying to run Ultrapixel for the first time here within ComfyUI, I already installed Cascade today, and everything is working fine, do you know how to solve this?

[past cradle]

Okay well what are those things designed to do please tell

I mean only in a sketch you don't have to say it all

[coarse kelp]

those are both UI's (user interfaces) so they have a bunch of buttons you can click and they will do most of everything you want

[past cradle]

At least give me enough so as to not throw a hitch over the timber and turn over a stool

[meager canopy]

Not really. Do you have a Huggingface account? Sign in, delete the ultrapixel model and custom nodes directories, reinstall the custom nodes and try again to let it download the files.

[past cradle]

I know what the fuck a UI is, as well as a GUI what do you take me for. And that'll be the last thing that I wrote my note after I have credited you Galaxy time machine

Hi now I don't have a hugging face

[coarse kelp]

i don't know what you do and don't know so i'm starting at the start

[past cradle]

I'm not sure how this interface works are we in private message right now when I respond to you right now?

[coarse kelp]

no we're in https://discord.com/channels/1002292111942635562/1089974139927920741

inside of the stable diffusion server

[past cradle]

I have so much to learn

[meager canopy]

The message isn't addressed you. Were you like this before the bump on the head? Perhaps simmer down a bit.

[past cradle]

And now who am I speaking to GTM?

Okay guys to everybody here's what I don't understand I think I'm I'm okay here's what I've here's what I think is happening right now. I think everybody is having a laugh at me because I have no idea about the format of this stuff and you guys are absolutely right I have no idea

But there was no there was no cohesiveness or cohesion towards any of this or between any of him and my siblings and so it all fell apart and I continue to hear about this platform through all the podcast that I would listen to and where they would try to like really get this thing off the ground and it seemed to me like a sort of fell apart in large part

But I I feel like it still held together and kind of an underground way and in a way that I would like to have been a part of but I just never really got into it and found a way to you don't like I didn't ever really find a way into it so I never did

[meager canopy]

https://tenor.com/view/thor-all-i-heard-blah-blah-blah-gif-22363686

[coarse kelp]

i'm not sure i understand

[past cradle]

Yeah FairPoint

Okay fine my point was who the fuck is even on here

[meager canopy]

Why does it matter to you? This is how "social platforms" work.

[past cradle]

Okay I'll just answer you from my point of view I don't know what we're talking about necessarily when we talk about the social network but I mean I got on the socials when they first popped off and then I got burned off of them I got tired of them I got off of them at some point because they are not serving me.

[shy kelp]

oh quite a lot of people quit social networks by now yeah

they are not as big as they used to be

[past cradle]

I hear stories about back when people were on VBS stuff and people who were early adopters and that sounds like what I would have really rather been on and I I was not that early of an adopter but I was on earlier than even that because I was on the phone tag stuff do you guys even know about that?

[shy kelp]

not sure what phone tag is

[past cradle]

Yo it was wild it was like, you would pick up your phone and you would call a certain number and then you would leave a voicemail. And then other people would do the same thing and then after a while you could call into that voice bank and you could press certain numbers and they could press certain numbers to be like oh let me hear your response to the thing that I said and so forth and back and forth and so people would just get into these riots in local neighborhoods within their local areas and just fight back and forth and that was before the internet it was even popping off although the internet did exist back then just nobody knew about it because it hadn't because the Internet hadn't popped off yet

[meager canopy]

Sounds like an LLM that's gone off the wire

https://tenor.com/view/crazy-alert-crazy-alert-alerta-alerta-loca-gif-166249322369313028

[past cradle]

Yeah that was the crazy alert back then because that's when it started it's like when everybody was started but that was like the earliest social media there ever was

[coarse kelp]

almost like an argument except you could leave it until you had a solid enough thing to say lol

[past cradle]

That was in the early 80s when that started happening and if anybody could have had the foresight to you know see the future but you know who could have predicted it

[coarse kelp]

i wonder what would have happened if modern social media as we know it today wasn't introduced

[meager canopy]

Were you even old enough to use a phone then? 🤔

[coarse kelp]

me? no. but if i stopped wondering about things that existed before i did, it would be a boring place

[past cradle]

Hey Galaxy I guess you're asking this Perez because I'll tell you right now I was born in 1979 but, I think you both are much younger

[meager canopy]

I'm 58

[past cradle]

Oh wow I'm surprised to hear that Galaxy cool I guess that makes me not the oldest person in the room

[coarse kelp]

i'm still the same age as before, guess there's not much i can do about that 😂

[past cradle]

The thing that makes me most upset in the world is that I'm not going to live to see what comes, I want to see all the things that comes and the fact that I cannot see those things it's the most bummer thing that I can possibly imagine

[meager canopy]

This is a channel about SDXL, can we get it back on track? 🤷🏻‍♂️

[past cradle]

Well tell me GTM what is sdxl?

Because I just wandered into this room looking for something that was apparently completely unrelated

If anybody's interested I just had a hell of a day. I'm just happy not to be in jail

But please tell me what is sdxl

[meager canopy]

Software that lets you describe an image and it then creates it.

Google can tell you more

[shy kelp]

these were testing vectorscope node

and that one was FreeU node but with the settings pushed the wrong way

(lower B and raise S)

[meager canopy]

Mine are all Cascade/Ultrapixel

[past cradle]

@coarse kelp hey I guess we got disconnected because I said some things that got banned I didn't know they were so sensitive over here my bad

Hey but I was really enjoying some of the chit chat we had going on earlier

I hope I didn't buy me out with anything I said I didn't mean to be offensive honestly I was just really seeking out answers but you know it was you and that one other guy really seem to give me some feedback and I really appreciated that

Your work looks really great for whatever that's worth I'm sure you get that a lot but it does.

You know I'm really not in the space but I really enjoy the conversation so if you'd ever like to get on a meet up for coffee

Strictly speaking in a techno space I just mean I really enjoyed kind of back and forth but I'm kind of a goofball myself and I don't also anyway I don't get out a lot it was nice to have some conversation I'll shut up now

I speak better in real life this was just voice to text so please forgive

[shy kelp]

not really getting good results with xxmix9realistic ..

[deft goblet]

Don't forget to use trigger words in positive prompt

[shy kelp]

hmm i totally overlooked that

i'll give it a try again

[deft goblet]

[shy kelp]

but then again, a checkpoint shouldn't require a trigger word

i mean its not a lora

but thanks for pointing that out, im copying the model again from my backup drive

[deft goblet]

Yes, but the owner of the model said that there are trigger words like lora, so you have to use them, haha.

ps. If you use it with Hires fix, I recommend adjusting the denoising strength = 0.3 and then using Upscaler (Latent) + 4x-UltraSharp (Pixel). If the denoising strength is 0.4 or higher, maybe the anatomy is wrong sometime. If using only denoising strength = 0.3 + Upscaler (Latent), the picture will be blurry as well.

[shy kelp]

ok ...i will make note of that

atm, im just trying a render w/o upscaling at 896x1152 res

preview looks better with xxmixgirl

what a big difference, thanks for the vital mention on trigger

[deft goblet]

https://tenor.com/view/keanu-reeves-you're-welcome-your-welcome-welcome-gif-14530451123339000551

[deft goblet]

for best result using with ADetailer (person_yolov8n-seg, face_yolov8n, mediapipe_face_mesh_eyes_only) as I wrote in the previous post.

for demonstration only (not cherry-picked )

[shy kelp]

i haven't been using all those detailer feature, been using it for only face, curious to try those 3 features together now

[deft goblet]

Maybe 2 features are enough (face_yolov8n + mediapipe_face_mesh_eyes_only)

[shy kelp]

im gonna play with it to see the difference, any thought on using adetailer for hand ? i haven't much luck with hands

[deft goblet]

As for the hand, if it doesn't come out well, using Adetailer is useless. You may have to send it to img2img >> inpaint and then mask the hand and use it with Controlnet openpose full.

[shy kelp]

there is a tool that can help hands

[deft goblet]

more process = more gen time haha.

[shy kelp]

i get fairly good results with pony models when it comes to hand but sdxl seems to fall short

https://old.reddit.com/r/comfyui/comments/19dlbp2/hands_fix_meshgraphormer_impactpack/

meshgrapho

meshgraphormer

[deft goblet]

I haven't used it yet, I'll give it a try.

[shy kelp]

nice.. ty for that, gonna try it out

[shy kelp]

using adetailer for face and eyes seems to do a pretty good job ...

[shy kelp]

its a long video but he shows that we now also have separated detailers for eyes, nose and mouth as well

https://www.youtube.com/watch?v=UTmwyxHQ7pM

they're pretrained detectors so they don't need any fine tuning

[hasty smelt]

I was going to sleep, thanks for posting this, when I wake up I want to study this.

[spiral lintel]

Is it better to use hires fix, or upscale a piece after it's already made?

[spiral bridge]

Ty so much u did it ill 100% try it today and let u know how it is

[vast galleon]

Looking forward to seeing some great images from it. 🙂

[spiral bridge]

I'll try my best depends how creative my brain works for prompts when I try it

Depends on time of day I'll try my best to not do it so late otherwise my brain dies

[vast galleon]

I know what you mean lol

[spiral bridge]

Later it gets my brain dies and I can't think of anything creative to do so prompting gonna be a challenge for me basically

So I'll try ur model at 4 pm so that's when I'm most creative

[vast galleon]

Sweet! I should probably do the same and stop releasing my models so late at night. Lol

[spiral bridge]

😆

[hasty smelt]

Hi guys, I tried SD3, but honestly, my impression is that SDXL is a few years ahead. Is that right?

[tawdry current]

Sd3 is a base model, not a fine-tune. Most of the sdxl models you've used have been very fine tuned for certain common concepts like portraits of pretty women. In doing so, those models tend to lose a lot of potential variety in things. Base models are meant to be a jack of all trades and sd3 is very versatile for almost anything you throw at it. Compared to sdxl base and 1.5 base, it's light-years ahead

[humble cape]

Hello what do you think about this image ?

[shy kelp]

maybe try an img-to-img pass with a lot of denoise

[shy kelp]

you might get something out of it

[glass forge]

Yeah faces come out good with adetailer but help me please with the hands...

hands are a hit and miss at best

[glass forge]

i like it its different, not typical ai vibes

[humble cape]

Good idea

[humble cape]

I typed uncanny valley

[shy kelp]

meshgraformer can help

*meshgraphormer

[worthy orbit]

Nice ones

[vast galleon]

Thank you!

[rotund surge]

Cool but strange, reminds me of gmod with that weird jpeg mods that follow you 🤣

[pulsar bloom]

/prompt a woman

[copper kraken]

Here is the image you requested.

[plain stone]

(Buzz Lightyear from Toy Story, design schematic). (detailed line drawing, black and white outlines, colored 3D main body, annotated labels, dimensions). (DSLR camera). (macro lens). (midday). (technical blueprint style). (Kodak T-Max 100 film) --ar 3:4 -- v 6

[humble cape]

Next bots 😂

[fierce heron]

anyone knows how to make an inpaint model in comfyui?

[shy kelp]

powerpaint v2 and brushnet

[silk smelt]

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

[stiff dawn]

/prompt a man

[uncut gull]

New model from the original makers of stable diffusion, Black Forest Flux
https://huggingface.co/black-forest-labs/FLUX.1-dev

(I added the boxes.) First model to understand shadows.

Transparency and reflections.

[uncut gull]

Very solid, not perfect. close up of the G chord being played on a guitar

Oh! So close. These are not cherry picked though. 1 generation each prompt is all I'm doing.

a painting by John William Waterhouse absolutely flawless. 😮

Perfect text of course.

More challenging text and it doesn't even blink.

"A word fitly spoken is like apples of gold in pictures of silver." At first you think it messed up and left out the word apples. And then you realize it put actual apples there. Is it really closing in on AGI??? 😐

[silk smelt]

Can I use FLUX with comfyui with 12GB vram?

[rotund surge]

You can but you will ran out of vram and ofload it into ram (which will be slow)

[rotund surge]

Im hungry now 😁 nice images

[tawdry current]

yeah, i'm running it on a 2080 with 8gb vram and it's slow, but works. 15 seconds per iteration lol... not really worth it, but it's an option

[hollow kraken]

and im here with my 20sec/it with just loras

[bright ice]

I'm getting good at styles.

[limber kindle]

Ooh this looks really good

[worthy orbit]

More like a Asus2

[hoary saddle]

Damn, it is good with adherence

[glass forge]

flux is like ideogram at home

[balmy dust]

https://tenor.com/bWcZg.gif

[tepid geyser]

Simply stunning

[glass forge]

hands of flux are so awesome after all the crap we had to put up with so far XD

[peak dove]

After ClownSharkBW

[pure rapids]

hello

[meager canopy]

...is it me you're looking for?

[uncut gull]

Flux on 4090: 22 seconds per image. It's a base model that does hands and eyes and keeps geometry pretty stable. I can knit-pick one missing wire on the leftmost wing strut and the right wheel's stalk joining the chassis too far back, but otherwise this is legitimately a well-constructed biplane. The model understands solid things and puts them together like block pieces. It feels so surreal.
(And this is the quantized fp8 version. The full version is probably better, but it's not even like even the 5090 is gonna help. This is as good as it gets consumer-side probably. 😐 )

three dice on a table, the dice have all rolled sixes Zero number-bleed! It didn't really get the concept of dice rolling sixes, but it understands that I want three dice, and it understands that I want three number-sixes, and it put them all on a table. This is just a case of it not having seen enough properly-labeled dice in its dataset, but the architecture is 100% capable of separating number-quantities from number-labels. 😊

Oh that is sooo good! 😱 No question it understood the concept of a reflection, not just two people standing next to each other. An anime girl leaning against a mirror. A girl beside her reflection.

(Reflections and counting are emergent intelligence. They're not behaviors you can get just by adding more data and labels. They only emerge if the architecture is intelligent enough to build generalized models of physics.)

Gah so close! No! 😭 This was really just so beautifully close. It gets the concept, it just doesn't want to write text backwards. I feel your struggle, little model. Gotta pass those text benchmarks. A sign that says "MIRROR" sitting on a bathroom counter beside a mirror. A sign and its reflection. The sign is reflected in the mirror, with mirrored text.

(Also bear in mind I'm one-shotting these, no cherry picking. You can always get better performance out of a model by running multiple generations and picking the best output.)

I can't get it to generate trading cards. 😦 They're not in the dataset, or they weren't labeled. That could easily be fine-tuned though.
But I'm still amazed by its sense of symmetry, the solidness of these costume parts, the fingers and the eyes and the overal body proportions despite using a non-standard aspect ratio, the structure of the sword, the correct folds and drapes in the fabric. This model just understands physical objects so well.
a trading card game trading card, gameplay element, card design, tcg card

512x512, 20 steps, 19 seconds

512x512, 10 steps, 11 seconds

512x512, 4 steps, 14 seconds? Why slower? 🤔 Also not usable quality, but it does seem a lot stronger than SDXL was before lightning.

Sampler / scheduler make a huge difference, no time to test all those now.

Bottleneck on my 4090 is swapping the VAE out of VRAM, so you don't get any meaningful speedup by lowering the stepcount at 1024x1024. 😦 It's always going to be 20-40 seconds.

It understands underwater hair and lighting so well.
swimming underwater, long hair

That's some flawless riding anatomy. It's just so perfect everywhere. The constructed lines, perspective, shadows, anatomy, lighting. Didn't give me motion blur though? Hmm. Well, no more tests until tonight after work.
derby, horseback riding, anime girl, motion blur

Had to try one more. pixel art anime girl fox
But styles can always be added with fine-tuning and loras, so they're really not important. But yeah, that's some absolutely flawless-looking pixel art right there. Doesn't look 100% accurate grid-wise. Not sure how well it would hold up if you actually scaled it down. Gave her two tales, but the style is good.
Gotta come up with some more physics tests since it's aced my normal ones. See what I can think up before tonight.

[loud spoke]

dream

[tawdry current]

With a massive 12b model, I'd hope so. That's a whole ton more room to store information, assuming they use it efficiently and correctly

[hoary saddle]

I haven't updated it in a year, but try my old workflow

https://comfyworkflows.com/workflows/5b30ac83-de7c-4d79-87d0-77436332e8f9

Eventually flux will have a cnet and loras and life will be good again 😂

[tepid geyser]

OOOOOOOOOHHHHH

[peak dove]

Are you having a smorgasbord at all?! 😄

[tepid geyser]

Yes

[peak dove]

🥳

[ancient terrace]

Ryan McCree

[edgy sparrow]

@copper kraken I'm interested in learning more and using Ultrapixel. Can I dm you for details, custom nodes, etc., etc.,?

[cursive raptor]

Impresionante! Que PC se necesita para poder crear una imagen de este calibre?

[hoary saddle]

cnet or a lucky prompt?

[copper kraken]

no luck

no cnet

[copper kraken]

yeah that's fine, also fine to just ask in general with images since we don't have a stable cascade channel anymore for reasons no one understands

might be other ppl with questions

[edgy sparrow]

Ok

[copper kraken]

feel free to ping anytime

[edgy sparrow]

Thanks!

[hoary saddle]

flux can do it, always get powered cocoa in SD

[glass forge]

so realistic

but i do see those vertical bands

[queen whale]

sexy girl

[shy kelp]

that's a latent lol

[glass forge]

You know what

to this day sd15 is the best for the last moment finetunes

like the last few steps to kind of finish the skin and eyes

😮

[unreal canyon]

Yup looks fantastic on the hair, lighting et all.. except the left leg 🦵😅

[uncut gull]

Yep. Some weird calf distortion.

[glacial star]

One message removed from a suspended account.

[glass forge]

Go away Garfield. It's not Monday!

It's the weekend! Leave evil cat!

Why.

looks at hand the robots just don't get you...

These are not the hands you are looking for. XD

[stray warren]

Wow, amazing!

[shy kelp]

really nice

I love AI sci fi images

[gloomy lark]

cascade

[subtle nova]

hellp

safasd

help

[shy kelp]

Now guess, where she is shopping 😄

[limber kindle]

The second one made me spill my beer

[radiant dome]

which model is this

[gloomy lark]

that one's cascade

[silk smelt]

The story of the three detectives, in a universe similar to ours. Being closer to other universes in the extra-space, more anomalies happen. Commonly golden fishes glide through existence, as if from another place. The bureau of the Blue Rose is investigating this matter, trying to capture one of these ghostly appearances. The blue box was specifically designed for the purpose. Then more gold fish gather, showing up like an easy target. Goldfish slowly took over, grew bigger in size and numbers. An even larger shark suddenly attacked and snatched a goldfish. But instead of being devoured full, the goldfish got into the spirit of the shark. In this new body, goldfish had the power to take over. A massive war between the flying goldfish-shark and mankind resulted, which continued for many years into the future. TBC

[glass forge]

This better be an AI generated promt or someone is doing a lot of weed.

[silk smelt]

puts the lighter down

[glass forge]

😆

[coarse kelp]

you can only really see the seam if you look for it, but i'm still gonna try to fix it

[glass forge]

Anomalies at first glance.

[coarse kelp]

what about this one? 😄

this is 2x upscaled

[glass forge]

ok much better

[coarse kelp]

i'm gonna try 4x in a while but it's gonna take a moment

or maybe just 3x

[glass forge]

upscalers amirite

[coarse kelp]

we got a freaking scooter now

[glass forge]

they take up like 80% of generation time

[coarse kelp]

true 😂

[glass forge]

Yes, a scooter with freaking flowers! XD

[coarse kelp]

from this far away nobody can tell xDD

sigh

this is gonna take a while

lol

[glass forge]

[coarse kelp]

see you in 10 minutes

[glass forge]

Whats the denoising?

lower denoising makes it faster i find

[coarse kelp]

0.43

0.5 was too high and 0.4 is too low because some of the details don't get fixed imo

[glass forge]

i stick to around 0.2 because i try to maintain original details

theres no magic formula

[coarse kelp]

true, and the most confusing part is that each image needs a different value lol

[glass forge]

i know!

no magic bullet fits all

[coarse kelp]

if you have lots of patterns you can get away with a higher value

a plain color with a slight variance in color will look different to everyone

but a square of slightly different leaves or bricks won't stand out as much

[frosty rapids]

Favorite SDXL model for anime?

[shy kelp]

A1111 + SDXL

[shy kelp]

SDXL explaining the big bang 😛

[mental minnow]

#1010934719455707218

[shy kelp]

A1111 + SDXL

[midnight panther]

Which model?

[midnight panther]

And this.

[shy kelp]

These are from a personal mix i made. It is cleaned and i also removed the noise and finetuned it with 15k images.

[vast galleon]

This one is from a new version of Mangled Merge XL that I'm working on. But if you're looking for pixel art, the current version of Mangled Merge works great.

[shy kelp]

Still need some work, but the details are good. Not many loose ends or washed out details.

[vast fulcrum]

Fine dust particles from charcoal barbecues, as recorded with an electron microscope, dense

[shy kelp]

Here is the image you requested

[vast galleon]

Trippy

[shy kelp]

3D printer in A1111 + SDXL

[quaint nexus]

draw a dog

[tawdry current]

https://cdn.craftymorning.com/wp-content/uploads/2022/10/easy-dog-drawing-for-kids.jpg

[quaint nexus]

The world is so beautiful when one looks at it with the innocent and hopeless eyes of a child: the moon rings and stars in the night sky are beautiful, the streams, the beaches, the forests and the rocks, the goats and beetles, the flowers and butterflies are beautiful.

[subtle bolt]

two beauty

[shy kelp]

anyone got this canny working with flux on comfyui? https://huggingface.co/XLabs-AI/flux-controlnet-canny

[tepid geyser]

Nice!

[dusky cairn]

artem skibidi toilet

[vast galleon]

I think it's finally got "holding things" down.

almost lol

the platforms are an issue though ...

[raven elk]

Hi! this image is Upscaled with Ultimate SD Upscale, generated with Flux.. I like the upscale exept for the color bleeding ang visible seams? any tips on upscaling images?

[midnight panther]

How to create images like realistic and high quality like that? When I want to increase quality, it moves away from reality.

Like this. You can see it in its iegs.

And wings.

[midnight panther]

How can you do that? This quality.

[lusty wolf]

Flux. ComfyUI

[midnight panther]

What is flux?

[lusty wolf]

https://civitai.com/models/618692/flux

[midnight panther]

Can I use that in A1111?

With xl models?

[lusty wolf]

Its is a model, like SD3. Yes, I think there was an update for that. Search your custom nodes.

[vast galleon]

You can also try a lower CFG. A lot of sdxl model tend to show more realism at a lower scale.

[shy kelp]

It is all about the model you use. As written above, this is a custom made one.

[thorny glacier]

Try the tile_colorfix preprocessor to avoid color bleeding. What I also like to do is to have two ControlNETs on—not just ControlNET Tiles, but also with IP Adapter, as IP Adapter dramatically reduces the hallucinations and allows for much higher denoising strengths to be used, allowing for much higher realism. Here is my example:

Here is the initial image.

This is the initial upscale. I used IP Adapter at full strength and let it stop at 0.5 so that the details would look right, using the DPM++ 2M Karras sampler. I had ControlNET Tiles at full strength (1) here and full time (1), and used a Denoising Strength of 0.75:

After that, I manually fixed some errors via drawing and digital tools, then inpainted them to make it more natural:

After that, I upscaled with the same settings I said above, while also using the same initial DAT upscaler in Ultimate SD Upscale (4xFFHQDAT). I didn’t mention this for the previous one, but for all of them I have the mask padding and the mask blur turned all the way up to reduce seams:

After that, I did some more inpainting and corrections, blurring any seams in unfocused areas, and then it came to the final result:

@raven elk You can also do this using SDXL upscaling too (I used a 1.5 model for the previous upscale), like in the examples below:

[glass forge]

Flux is SD3 if it was good.

[placid spade]

Like this. You can see it in its iegs.

[vestal island]

Like this. You can see it in its iegs.

[hoary saddle]

Like this. You can see it in its iegs.

[thorny glacier]

Like this. You can see it in its iegs.

[vast galleon]

Fighting an unarmed man. LOL

[glass forge]

I bet she knows kung fu.

[vast galleon]

And she's not afraid to use it! 😂

[vast galleon]

I gotta say Dare/TIES merging really is something. I'm throwing tons of photorealistic lora into specific style blocks expecting it to bleed into the 2D outputs, but it actually seems to be making them better.

Welp ...