✨｜sdxl

[heady vale]

no idea. I also havent used A1111 for a while either

[dim steeple]

all good thanks for the help

[heady vale]

turn off highres fix and just do 1024

compare then

[dim steeple]

i thought it is off

[heady vale]

and no controlnet. theres no way its the same speed as dynavision at 20 steps

[dim steeple]

i thought when it turbo released it was only 512x512?

[heady vale]

yeah turbovision was merged so you get nearly the same speed but can use normal resolutions like other XL models

[dim steeple]

oh i see

restarted auto, its faster

[peak dove]

3060 ti and TensorRT can do same in 7 seconds

[lapis gale]

how long for a standard SDXL image? comparing apples with apples and that 😉

[carmine gale]

is this still the most appropriate AR to go by:
1024 x 1024
1152 x 896
896 x 1152
1216 x 832
832 x 1216
1344 x 768
768 x 1344
1536 x 640
640 x 1536
?

[steep sentinel]

a chinese woman,30 year old woman,(35mm style:1.1),analog style,masterpiece,raw,cinematic film still from 1990s horror movie,film grain,night,dark dramatic,film grain,Exaggeration,Dark Aesthetic,Nostalgiacore,Backlight,light rays magical lighting beams,dust in beam,moody lighting,dark color,Exaggeration,Dark Aesthetic,'dark,ominous,haunting,foreboding,gloomy,stormy,sinister,shadowy,harrowing',storm,lightning,Realism,VFX,FXAA,TXAA,RTX,SSAO,Shaders,Denoise,De-Noise,Ambient Occlusion,Tone Mapping,Moonlight,super resolution,ultra hd,megapixel,(8k resolution:1.10),8k,8kres,8k res,high details,detailed and intricate,intricate details,high intricate details,absurd amount of details,absurdness,high resolution,(high quality:1.2,masterpiece:1.2,:1.21),(mature woman:1.21)

[native knot]

#1072220168534642768

[lapis gale]

211 tokens overkill, lots of repetitions - touch of voodoo? 🙂

for comparison, this is simply
"chinese woman 30 years old concerned face thunderstorm dramatic lighting at night, dark mood"
(no finetuning or cherrypicking at all)

[dreamy plover]

A hoodie inspired by space suit

[dim steeple]

I am suprise there is not a custom modeel trained yet for a libary of celebrities?

Is this because of copyright?

[pure crystal]

yep

try more steps and a further fine tuned SDXL model

[lapis gale]

everyone popular is in the training set anyway, this is "the whole internet" as a baseline - or what do you mean with celebs?

[lapis gale]

"perfect strange 3d objects shattered bliss timelapse" 🙂

low poly version

[cyan crown]

https://civitai.com/models/233667

[lapis gale]

"renaissance style charcoal sketch of a moon rocket, beige canvas" no lora at all 🙂

and another well known guy

tesseract

[sharp vale]

sdxl base to sd1.5 upscale

hm more like rescale

[cyan crown]

good, but it''s too "clean"

[lapis gale]

adding "rough hatching hatchures lines"

"renaissance style charcoal sketch of a man with his head in his hands, beige canvas, rough hatching hatchures lines"

"renaissance style charcoal sketch of a man with his head in his hands, beige canvas, rough hatching hatchures lines shading"

for fun and profit

[lapis gale]

cute

[cyan crown]

before doing this Lora i used similar prompt. But again in the drawing part it''s too clean. Very good but it doe not gives the sense of drawing that I was looking for

[shy kelp]

Anyone have any tips for making lighting / shadows consistent between two faces in an image? Want to preserve the facial features as close as possible and just alter the lighting.

[lilac raven]

What things should I look for when releasing a character LoRA? I think it works (as long you use 0.75)

[vale eagle]

https://civitai.com/models/235649 Released my human preference lora alpha version

[lapis gale]

just slightly deranged

[lapis gale]

zany

[lilac raven]

What openpose works with SDXL? I tried a few and they seemed to not work well to actually keep the pose

[uncut gull]

16K ship diagram. 🙂 Exact dimensions 16384*16384

[alpine scaffold]

Hi, first contribution

[glacial axle]

Beware of the Vikings !!

[steep pecan]

哈哈

[sullen laurel]

Mecha musume

[carmine gale]

is there a hughe loss in quality if my training images are 1024x1280 vs the official sdxl 896x 1152??

[cyan crown]

with A1111 1.7 and TensorRT I get 70 steps (with refiner) in 14 sec with 4070 12GB instead of 48

[cyan crown]

the problem is that I can''t compile Loras

[cyan crown]

well with branches it seems to work

[steady grove]

tensorRT is such a gigantic technical hurdle for

[cyan crown]

it's a GB-eater

[cedar bobcat]

I don’t know, but if you’re concerned you could always downsize the images, or upscale then downsize (might preserve quality slightly better).

[carmine gale]

1024x1280 gives me good results. It’s seems to be in the list for official sizes, but not the core sizes. Just not sure. About to format my entire dataset library and want to make sure I’m not doing something I will regret later in training

[lapis gale]

i thought all training was at 1024x1024?

[steady grove]

could be that the high resolution pixel level deformities are more prominent. scaling it down sharpens the irregularities out.

I always train sdxl at 896, 896 with buckets and repeats to help even out batch lengths. i've tried training with one image size versus buckets and i've never noticed much difference in quality.

The biggest differences to quality is the source image sets and their pixel resolution quality. Often we only look at images at smaller than native sizes. Training is working on the pixels though.

[carmine gale]

ok i’ve focused on high quality curating of images. removing text and logos on image, topaz upscale one by one, to get very decent quality faces. I throw out anything that was shot with a bad camera or which has too much post processing and is simply below a high standard. I am getting already a good result with 1024x1280, but one cant be sure, there might be some flexibility issues or what now that one would never know about without proper comparisons

just cant be sure

[peak dove]

TensorRT will not work in SDXL on 8Gb VRAM, as TRT uses only xformers. 8Gb must use --medvram or --lowvram

[carmine gale]

still not sure whether to proceed on these resolutions or the more official one 896x1152.. ive been hearing people say just run with yours

[lapis gale]

starcraft style - no zergs were harmed in the making of this image 🙂

alternative version

[steady grove]

i wouldn't use topaz on training material. It's okay if one or two of your source images are lower resolution / quality, but you should not need to upscale a training set.

topaz lleaves really obvious pixel level artifacts and works less good on already very small iimages.

you'd be better off using sd15 with cnet tile to upscale a low quality image to native sdxl resolutions

[carmine gale]

That’s a very good point. I totally agree about low quality. Topaz is not adding detail, and will more than likely add artifacts like hard lines. But for medium quality images where there is something it is pretty good

Yes I will definitely add the comfyui upscale of interesting images that don’t have the quality

Thanks for the great comment

[steady grove]

but also, with a good mix of high / low, i got wacky style results. used a lot of really old photos from the 90s to make a lora that turned out really well

and high res shots of newer versions

https://civitai.com/models/207437/madball-xl

no lie, these were in the training data straight up. let kohya upscale them.

[soft bone]

client i trained

[slender coral]

Anyone have any good reccomends for training stacks? I've only used auto 11

[carmine gale]

This is more like the kind of stuff I do. But I take your recommendation home with me. Maybe it’s that, ‘use lofi images to give the model more flexibility’

[steady grove]

for faces i def wouldn't use 300x pixel images lol. the lofi approach works better for concepts and styles i think

[steady grove]

was thinking of making a hollowknight model, but juggernaut clearly knows

[sudden lichen]

Prompt：

(1girl), long hair, perfect body, windy, blue sky background, full shot,natural light,(masterpiece,high quality),Highly detailed,

Negative Prompt：

nsfw,logo,text,badhandv4,EasyNegative,ng_deepnegative_v1_75t,rev2-badprompt,verybadimagenegative_v1.3,negative_hand-neg,mutated hands and fingers,poorly drawn face,extra limb,missing limb,disconnected limbs,malformed hands,ugly,

[rustic garnet]

it's fine. SDXL can deal with many kind of resolutions without any problems

[misty steeple]

a 3d rendering of a city with a ferris wheel, in the style of cyan and black, national geographic photo, anglocore, precision of line, light black and indigo

[uncut steeple]

#1072220168534642768 @sudden lichen

[crisp timber]

New lora for #1186697881390956645

[ruby karma]

can someone tell me how to achieve a look like this with a logo but the trees naturally forming the logo instead of it being so clean and white?

i did this with juggernaut xl and controlnet in automatic11111 but cant get it to look natural

[vagrant valve]

I would play around with the strength... maybe the prompt a little too.

Here are 3 examples, the first with the highest strength and last with the lowest strength.

[ruby karma]

could you tell me what prompt you used for these? @vagrant valve

[vagrant valve]

I used QRCODEMONSTER with SD 1.5 for these a while ago. The prompt was ( Deep Space, the pillars of creation, nebulas, powerful colors )

[ruby karma]

I think part of the problem is that controlnet sees the logo as thin lines only?

QRCODEMONSTER is a controlnet model?

[vagrant valve]

Yes! The most popular one is the SD1.5 model by monster-labs on hugging face but it looks like they have a SDXL version too. Linked : https://huggingface.co/monster-labs/control_v1p_sdxl_qrcode_monster

[ruby karma]

how do I select the model here correctly?

[vagrant valve]

I have limited knowledge of Automatic1111 and do not know. I use comfyUI. Here is a video I found that might be a help https://www.youtube.com/watch?v=CmIxW_VOo8U

[ruby karma]

what is this $50 stuff

[vagrant valve]

Most likely a scam... I would not click on that link

@smoky patrol Sorry to tag but I believe these are scam bots.

[ruby karma]

@vagrant valve ur a champ

thank you

[vagrant valve]

Hell to the yeah

[vagrant valve]

H e r r i

[ruby karma]

woah

how do you get such a good negative effect?

[vagrant valve]

I'm just playing around with it. I do not know

[vagrant valve]

What was your prompt for this?

[ruby karma]

any idea how to word this view

looking up at looking up at snowy tree crowns isnt producing anything

[vagrant valve]

try something like " worms eye view of a deeply wooded forest " with a strength of .70 - .90

[ruby karma]

the strength is the cfg scale?

[vagrant valve]

no. I keep my cfg around 6 - 8. On comfyui i am able to control the strength of the controlnet model... not sure how it translates to automatic

[ruby karma]

ohh yeah

mine is 0,75

[vagrant valve]

fantastic

This is kittens with a strength of 1.4 lol

[ruby karma]

[ruby karma]

https://huggingface.co/monster-labs/control_v1p_sdxl_qrcode_monster

@vagrant valve I think this isn't released yet?

this is not the controlnet model right

[vagrant valve]

that is it

[ruby karma]

oh

[vagrant valve]

or at least thats the one i downloaded and am using with juggernaut XL v7

[ruby karma]

I tried to do pomegranate seeds

getting super weird results with the xl model

weird

[hoary saddle]

is there a hidden "ugly xmas sweater" lora anywhere yet? only 1 on Civi and it's not ugly

[cyan crown]

don't use refiner with qrcodemonster

[peak dove]

I have used cd to get to here - B:\ComfyUI\custom_nodes\ComfyUI-Manager>

How do I go back one level to B:\ComfyUI\custom_nodes?

[vagrant valve]

just do $ cd and it will reset your current directory

sorry exclude the $

[peak dove]

Thankyou

[crisp owl]

if you want to go back a single directory, cd ..

[native knot]

or if you want to stay in the same directory, cd . 😛

[peak dove]

Or if you want a dvd and not a cd ... ? 😄

[crisp owl]

You could always create an alias for cd so if you type dvd instead it'll do the same thing 😉

[peak dove]

Using Plush-for-ComfyUI OpenAI Node

Prompt = Surrealist Art of frail children morphing into dust particles, an American Gothic house melting into a dystopian landscape, Dustbowl Okies with faces of anguish transposed onto weathered barn doors, homeless figures merging with the shadows of despair, a family dinner table set in a dysfunctional swirl of storm clouds, a shriek transforming into a flock of disturbed crows, a scream materializing as a cascade of shattered glass, anger personified as a fiery sunset over a barren field, style of Beksinski, Kowch, and Rockwell. Model = CounterfeitXL

[crisp owl]

While these are nice images, this is the SDXL channel, not for 1.5 based models.
#🏞｜general-with-images

[shy kelp]

I wasn't aware, thank you. Will merge them over there!

[bright valley]

You can also generate text with my model, no control nets or anything required

[crisp owl]

it's true, I occasionally will use your lora with even a bit of a lower weight to help integrate text better into some of my images.

[bright valley]

whoops an extra I

[bright valley]

awesome!

You should check out HxSVD if you haven't yet, it's pretty darn cool

[crisp owl]

huh?

[bright valley]

It's called the HxSVD Workflow, it's for comfy, and it makes these ^

[crisp owl]

Ah nice!
I'll bookmark it and check into more later. I'm behind on some of my stuff I'm working on and trying to catch up (never ending game it feels like lol)

[bright valley]

oh yeah you dont gotta tell me twice lmao

[twilit nexus]

what are these new sdxls turbo zero etc?

also what is really good for making stylized models like leage of legends or prince of persia or things like that

[twilit nexus]

anyone using this model?
https://civitai.com/models/84040?modelVersionId=247214
recommend or no?
currently using juggernaught

[ruby karma]

how?

[bright valley]

Hah, I literally just told you my guy, using my model Harrlogos: https://civitai.com/models/176555/harrlogos-xl-finally-custom-text-generation-in-sd?modelVersionId=214296

[peak dove]

Wow! This Plush node for ComfyUI sure does produce some powerful imagery!!!

[peak dove]

These next ones are by Distillery@Discord (FollowFox)

[ruby karma]

I didn't see 🙂 will check it out after work! Thank you

[rustic garnet]

what's your upscaling workflow for SDXL? Do you use SDUltimate Upscale + an 1.5 SD Model for it, or do you use an SDXL model? Do you use 1.5 Tiled Controlnet?

[peak dove]

Topaz GigaPixel

[ionic dragon]

4x-ultrasharp

[rustic garnet]

upscaler oO?
They never looked good for me

[ionic dragon]

they worked fine for me, made the details sharp

[celest heron]

anyone else having issues with clip encode? I keep erroring out on comfyui whenever I change the text of the prompt, no other changes to the workflow, just the text. Restart Comfyui and it works again until I change the text

"Error occurred when executing CLIPTextEncode:

Cannot copy out of meta tensor; no data!" is the error message

[unreal sandal]

Hello, I'm getting a "model not found" error with the animation SDK using the SDXL- 1024 model and was wondering if anyone has info about it. Haven't used the SDK in a while and got this error out of nowhere when running my exact same configs on the same project as before.
Only happens after 1 frame has rendered. Log: """
8%|█▏ | 1/12 [00:07<01:22, 7.50s/it]Received RpcError: <_MultiThreadedRendezvous of RPC that terminated with:
status = StatusCode.NOT_FOUND
details = "The specified engine (ID stable-diffusion-xl-1024-v1-0-comfy) was not found."
"""

[old parcel]

https://tenor.com/view/shocked-black-guy-shocked-gif-27437284

[wet nacelle]

@smoky patrol GET THEM!!!! GET THEM!!!!

[uncut steeple]

For tech related issues check out #🤝｜tech-support

[sharp mason]

Anyone here been able to finagle sdxl to work on 5xxx series Radeon cards (like a 5700xt)?

[peak dove]

This makes excellent prompts! Plush for Comfyui ... https://github.com/glibsonoran/Plush-for-ComfyUI

[carmine gale]

Need to Upscale tight crops for face training. Want to use Comfyui to generate details. Would it better to work from pixelated small (around 300x360px) or an upscaled Topaz upscale (1024x1280) to remove pixelation to then denoise as a sharp image?
Note, the right image is Topaz'ed

[crisp timber]

https://civitai.com/models/226324/rankinbass-studios-lora

A great Christmas Lora for SDXL

[keen nimbus]

https://tenor.com/view/rick-rude-gif-7954254

[bitter juniper]

baddass jaffa in construction...

[jolly brook]

jaffa, kree!

is that a LoRA you trained?

cause I def need it

[wet nacelle]

@jolly brookhuh row

[jolly brook]

hi

[dense skiff]

hey guys im having a hard time getting SDXL to run with Vladmatic anyone willing to lend a hand

[dim steeple]

There was a really helpful link I saw on the sd reddit regarding all the available SDXL controlnets, but I can't find it

any help?

[peak dove]

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

[bitter juniper]

yep a lora. I try to adjust parameters to get flexibility and details. Then I will start a lora for all the stargate universe.

[peak dove]

Hao

[cobalt canopy]

你好

中原

[ionic dragon]

when's the next july 18?

@high skiff any insider info

[peak dove]

Pixart-alpha

[eternal fog]

Been playing around with the new Insightface IPAdapter and it works pretty well.

Input image:

Outputs:

There can be some weirdness with the eyes, but it's a lot more accurate than the regular IPAdapter models

[high skiff]

I do have a little bit, but I am bound under multiple NDA's from my job, so there will be no sharing this time lmao

[crisp owl]

What's your process, or video you followed (or website) describing how to use it/apply it?

and in Comfy? or elsewhere?

[eternal fog]

Comfyui and I followed the sample workflow on the IPAdapter comfyui addin GitHub page.

I've modified it a little as it's a sd1.5 model.

I make the image with SDXL first, then Img2img that image with sd1.5 and that IPAdapter.

[crisp owl]

aaah, ok. I was looking on the github for the sdxl info and didn't see it, that makes sense now

[eternal fog]

Hopefully they'll make an SDXL one next

[bitter juniper]

And I thought the mandalorian was cool...

[nocturne dove]

Member Since Dec 21. Will not make it to Dec 23.

[vale eagle]

the third version of my human preference lora

[bright valley]

Somebody: harrlogos is too inconsistent!
Harrlogos:

[ionic dragon]

You joined SAI?

[peak dove]

What's the name of the latest iteration of IPA? My ComfyUI Update has broken my Pixart-alpha (amongst other things!)

Speaking of ControlNet - where is the correct folder for this model at all?

ComfyUI cannot "see" this, which is under Models/Controlnet

(I often ask in tech-help, but get no answers)

[glad grove]

prob in lora folder since its a lora

why you using the control lora tho that thing is outdated

[peak dove]

OK, I will try and copy it over to there ... I am using Portrait-Master-v2.2

So I need that Control-LoRA

[high skiff]

Oh no, I don't work for them. That ship is long sailed

[peak dove]

I see the problem - when I d/loaded the Control LoRAs package - everything BUT the LoRAs d/loaded!!!

The d/loaded LoRA folders are empty!!! 😄

OK, seems to work now 🙂

Some examples from Portrait-Master-v2.2

[high skiff]

Is it supposed to look so cooked and messy?

Or am I missing what's going on here?

[vale eagle]

which one is better? ABC

[rustic garnet]

interesting. I had only bad experiences with IPAdapter. Like it is nice to play around and mix images, but it is totally sensitive to styles and less to content. It never adapts to "faces" , but just to style and composition. So it seems that this insightface is working better ins this regard?

[pliant mesa]

/yep a lora. I try to adjust parameters to get flexibility and details. Then I will start a lora for all the stargate universe.

[peak dove]

I used the wrong Checkpoint - now using NightvisionXL and it is much better...

They need denoising and sharpening (I shall use Topaz Sharpen and Denoise)

[peak dove]

Where to d/load? 🙂

[cyan crown]

I think it''s only for 1.5 and not sdxl

[peak dove]

I'm using SDXL, seems OK - I will try 1.5

Anybody got a json of the InsightFace/IPA w/flow at all?

[thorny glacier]

Cute bees:

[stone fossil]

https://civitai.com/models/241477

MoDernart - Newskool

Introducing MoDernart, an innovative LoRa (Text to Image) model that goes beyond conventional art boundaries, bringing a fresh perspective to creative expressions beyond the -abstract.
By simply invoking the trigger word "MoDernart, you'll unlock a transformative experience, elevating your images into the contemporary realm of modern art and more.
MoDernart has been meticulously trained on a detailed dataset, featuring 6000 steps of hand-captioned content also containing the official artwork name and the artist.

All example images are made with Paradox 2.0 as base model.

Feel free to load the model onto your preferred Stable Diffusion application, give it a spin and let us know what you think.

Use trigger word: MoDernart

[cyan crown]

And with A1111?

[peak dove]

I rarely use A1111 - but when I do I will check it out - or are you saying I NEED to use A1111? 🙂

[peak dove]

D/loading ...

Is Mo Dern related to Laura (or is that LoRA) Dern at all? 😄

[ionic dragon]

are they working on sdxl 2?

[stone fossil]

Nope. 🙂

[glad grove]

yes it will release along with controlnet 2 part 3 season finale,but only the super expert higher up ceo know this information since its very secret and crucial for the company

[ionic dragon]

@uncut steeple @smoky patrol

[uncut steeple]

[peak dove]

I'm using MoDernart LoRA with Nightvision Portrait XL - Prompt = MoDernart peruvian arpillera style of victo ngai henri rousseau vladimir kush (the last two pictures are without the LoRA)

Turning the CFG up to 16 (from 8)

[high skiff]

If I just told you I am under multiple NDA's that could not only end my career but also put me in legal trouble, what makes you think I would possibly be able to answer that question

[peak dove]

If they swabbed me for NDA, could they find out the truth?! 😄

8^(|) d'oh!!!

[eternal fog]

https://github.com/cubiq/ComfyUI_IPAdapter_plus for the extension and documentation on how to use in Comfy

https://huggingface.co/h94/IP-Adapter-FaceID/tree/main for model files.

[peak dove]

Thankyou

[eternal fog]

The SDXL Face Plus model can get pretty close to faces in my experience. The insight face one gets closer, but I've noticed it does a few strange things, mostly with the eyes. I think that's something to do with how insight face captures the faces. It tends to change eye colours and also makes the eyes look directly into the "camera" which can look a bit weird. It gets the likeness better though.

[peak dove]

I'm finding the same about Portrait Master - the eyes are always straight on - so it seems like a crime-scene police line-up photo everytime!!! LOL

[eternal fog]

It will even move where eyes are looking if you do a light img2img

They've only just released it though, so maybe some more work required on it. It does generally get facial features pretty well though. Main thing I've found with the input images is it needs to have clean consistent lighting.

If any part of the face is in shade or has coloured lighting on it, it tends to screw the whole thing up.

its resonably good at face swaps with img2img, looks better than reactor in my opinion

At the moment I've got it setup so it does an image with SDXL and the IPAdapter Plus Face to get the general face shape and idea. Then a 2nd pass over that image with SD1.5 and the FaceID IPAdapter.

[peak dove]

Good setup

[eternal fog]

You can do it without the Plus face on SDXL, but sometimes it then doesn't match the face shape very well.

[peak dove]

Send me a workflow PNG/json? 🙂

[eternal fog]

But there's about 6 settings for the IPAdapters and LoRA that you can change to tweak it.

[eternal fog]

I'd suggest using the workflow in the examples on the github repo. I just have 1 workflow that I have a bunch of stuff in and it's a complete mess with nodes you have to bypass or mute to have stuff work. It wouldn't be usable without explaining the whole thing.

My workflow also still has the problem that upscaling anything makes everyone look like they are made out of plastic and I can't seem to fix it.

[peak dove]

I always Upscale via Topaz GigaPixel - much much faster! 🙂

Some more MoDernart LoRA

[vale eagle]

SDXL base with my human preference lora.

[ocean ermine]

Yeah, that's the main problem with upscaling--it loses the definition that was so clear in the base level gen. I recommend decreasing denoise, using an upscaler model that helps keep details, perhaps fatal_anime (good for everything, not anime alone) or Remacri. Also, more steps in the hires pass--try 15-25 depending on the sampler you're using. The sampler itself also matters. If you're not already using 2M Karras, I can recommend that or DPM 2M/3M SDE Exponential (minimum 15 steps for the upscale, as exponential requires slightly more than karras).

I usually use a three step upscaling process, with a 0.4-0.5 denoise on the first, and then a 0.25-0.32 denoise in the second. The second is ultimate SD upscale, so it goes tile by tile and helps keep details sharp. You wouldn't want a high denoise on a tiling generation anyway. I'd love to use multidiffusion/tiled VAE as it's a far superior tiled upscaler, but that's not available in ComfyUI, and generally doesn't work well with SDXL anyway.

[eternal fog]

Yeah I've tried stuff like this and still looks plasticy to me unfortunately

[ocean ermine]

There's a fair amount of suggestions there. You've tried all of them?

[eternal fog]

At some point or another yes

[ocean ermine]

I see.

[ocean ermine]

Does the base level gen look plasticy?

[eternal fog]

Sometimes they do, but even when they don't the upscaled result still looks crap. Always has with SDXL for me.

I tried so many things that eventually I got bored and just stopped trying lol

[ocean ermine]

Well, if you'd like to give me an example image with a workflow, I can take a look. It could also be on the model side; some models really don't do well with upscaling, and it's only in the last month that I've seen cohesion improve well enough to bring up the denoise on an upscale to 0.5.

[eternal fog]

I don't have anything at the moment, I might have a look again later.

[peak dove]

This Upsaling method claims to have more detail after Upscale https://github.com/ttulttul/ComfyUI-Iterative-Mixer/tree/main

[elfin flare]

Aftrer trying SDXL turbo for a bit

yeah nah I think for now until it gets better overall quality I'll stick with SDXL

Also, I just like 1024x1024 more

[eternal fog]

Yeah I looked at that a little earlier, but it doesn't look like it's a tiled upscaler, so it will take ages or i'll run out of VRAM

[ocean ermine]

This looks very promising, at least for the initial upscale.

[ocean ermine]

If it works properly, then break it down into a couple upscales, max 1.5x each. The first can use the Iterative Mixing and it keeps most of the detail with the higher denoise. The second you can have a much lower denoise, and from there it's upscaling and re-genning just enough to get rid of the upscale model artifacts.

[eternal fog]

Have you tried it out. I just tested the example workflows and it doesn't work properly. It gives awful noise on everything.

Oh lol it's a "known issue" that it leaves noise.

So it's kinda crap then

Yeah this is awful

[bright valley]

I also really like gigapixel!

[bright valley]

Wow this pixel art looks great!!! What model is this?

[elfin flare]

This is OpenDALLE 1.1 + A combination pixel art and voxel art lora at .50 strength for both

[bright valley]

Pixel art style is one of my favorite things I do with my model

[elfin flare]

I wanted to mess with stablezero123 with this stuff to see if I could attempt a 3D Model

But yeah no I got nowhere

Makes me very sad

[bright valley]

Yeah it's not true pixel art/voxels/whatever

It's just pretty

[elfin flare]

No it's not even just that

I just can't get consistent output outta that model

🤷‍♂️

[bright valley]

Oh, I haven't tried, I'm a 2d guy

Although I make 3d looking things

[elfin flare]

I saw this.

Yeah.

The bobbing is weird, but ye.

[bright valley]

That's from the HxSVD workflow, super fun

[elfin flare]

Lol

How much VRAM did you need to process that?

[bright valley]

Have you tried svd? You can't really control the motion

At least 8GB

[elfin flare]

Well I have 16

But since I'm on Intel Arc

big think

Where is this SVD workflow? I'll try it

[bright valley]

https://civitai.com/articles/3355/hxsvd-harrlogosxsvd-txt2video-comfyui-workflow-generate-and-animate-text-with-svd-v2-out-now

Here if you wanna check it out

[bright valley]

Yeah there's 2 versions for dl

The second version has 2 svd passes and RIFE interpolation

But I haven't updated the templates that I provide to v2 yet

[ocean ermine]

I haven't tried it yet. Wow that looks terrible. Is there no way to tell it to stop after a certain point?

[eternal fog]

Seems like it's potentially broken, as there's a lot of people on the reddit post also saying that it does this and they can't replicate anything he's done in his examples.

[elfin flare]

I added a turbo lora to this, so now it generates stuff like this in 12 steps.

Before, I needed to do 70.

[thorny glacier]

@eternal fog @ocean ermine You absolutely need to use ControlNET Tiles when Upscaling to get a natural upscale, especially with Ultimate SD Upscale, though it can be helpful with Tiled Diffusion/VAE. You can even go to denoise as high as 0.55 when using ControlNET Tiles, but mostly likely you will need to upscale the image/generation with a 1.5 model. Here are some examples of some of my upscales, along with the initial image at a lower resolution:

[ocean ermine]

This is, of course, the ideal, but Controlnet Tiling isn't available for SDXL. Unless that has changed since I looked last?

Ah, didn't read your "but mostly likely you will need to upscale the image/generation with a 1.5 model"

We'll just have to wait until X-Adapter comes out.

[thorny glacier]

Unfortunately, it hasn’t changed. BUT, something that can help (at least for initial generations) is Kohya’s Deep Shrink, as it helps the composition to stay coherent at higher resolutions

[ocean ermine]

I've never been able to get that working right. Even at 1.5x scaling, it's far too unstable and incohesive to use properly at high resolutions.

[eternal fog]

I've been using SDXL I'd forgotten about the tiles ControlNet, I'll give that a go

[thorny glacier]

You could try inserting an initial SDXL generation/image inpainting edit and then upscale with a 1.5 model; it always works very well for me

[ocean ermine]

Tell me, what does your end goal resolution for most of your gens look like?

[thorny glacier]

There are certain settings you have to do that can be a little complex to get it working properly, as the extension doesn’t tell you what the “base” resolution is—or, your software isn’t fully updated

[elfin flare]

This guy's got a lot to do today.

[thorny glacier]

Almost always 4K (4096~4160px on the long side), as shown in the images I posted

[elfin flare]

I greet you, my good sirs.

[eternal fog]

Yeah that's what I've been doing but I forgot about the tile controlnet

[ocean ermine]

And your denoise? What's that usually set to for tiling?

[shy kelp]

XL is still the most up to date/best model right? any word of a newer better model for 2024 🤔

[elfin flare]

I've been using OpenDALLE 1.1 so far

Which is just a specialized SDXL model

[thorny glacier]

Since I usually upscale with 1.5 based models, I can use ControlNET Tiles; therefore, I usually denoise as high as 0.45~0.55, but I usually do 0.52—obviously without ControlNET Tiles, there is no way one could go that high normally. Of course, I always use DPM++ 2M Karras, a deterministic sampler, for Upscaling as I don’t want to introduce hallucinations and weird stuff that ancestral/SDE samplers do in Upscaling (perfectly fine for base gens, though)

[cyan crown]

I use Tensor RT. 79 steps are faster than SDXLTurbo 12 Step

[thorny glacier]

@ocean ermine Also, I do such a high denoise because it allows for the most natural details; at lower denoise, while it is easier to keep the composition, it tends to not look as natural as at 0.45~0.55

[ocean ermine]

And you somehow don't have issues with characters or unintended things showing up with that high of a denoise? Are you removing everything from the prompt at that point?

[elfin flare]

I cannot use TensorRT.

As I'm not a Nvidia user

[ocean ermine]

Surely if a tile is over a patch of nothing, things will find their way into the tile.

[thorny glacier]

No, actually, I can keep stuff in the prompt because of ControlNET Tiles

[thorny glacier]

Do you ever use non-deterministic samplers (SDE samplers, any sampler with “A” behind it) when Upscaling? That can cause problems. Also, ControlNET Tiles is essential to not get such hallucinations at such a high denoise, as even deterministic samplers like DPM++ 2M Karras will create such hallucinations without it

[ocean ermine]

Generally I use 2M Karras or 2M/3M SDE Exponential. But I've only attempted using Controlnet Tiling once before months ago.

[elfin flare]

I've actually never used Controlnet tiling

[ocean ermine]

Have you heard of the X-Adapter though? And the coming ability to use 1.x plugins with SDXL?

https://showlab.github.io/X-Adapter/

Controlnet tiling is specifically listed in the testing they've done thus far. Very interesting stuff.

[thorny glacier]

Another thing to mention is that I don’t do latent upscales; I usually use UltraMix Restore as it keeps texture generally and the composition

[thorny glacier]

Yes, I think when/if it comes out it will be amazing!

[ocean ermine]

Yeah, latent upscaling is a lot more trouble than it's worth. After doing a fair amount of back and forth testing months ago between upscaling models and latent, I found that it's much better to just use a model.

[thorny glacier]

Me too honestly, latent Upscaling introduces too many random artifacts, even with ControlNET Tiles

[ocean ermine]

Eh, you should probably zap that first image. It's a bit too nsfw for the server.

[eternal fog]

lmao, there is a slight bit of shaft on show there

[thorny glacier]

Okay, reposting (It is really amazing, try it! The coherency you can get is stunning. Here’s another example):

[eternal fog]

How do you apply the controlnet tile in comfyui again. Do you generate an image and then pass that image through the controlnet?

[ocean ermine]

But anyway, the main issue here is that you still can't use SDXL models for the actual generating. So with a high denoise, you'll be removing all of what makes the base level SDXL gen great. Better overall cohesion, a greater awareness of ideas, the ability to natural-language prompt without issues, not to mention the styles.

[thorny glacier]

This is a good workflow:

[thorny glacier]

You can still use the SDXL generation as a base generation, as I did for most of these. And ControlNET Tiles does a really good job of keeping the integrity of the image while adding appropriate details

[ocean ermine]

That likely won't work so well for things like illustrations or anime artworks. There's just too much of a stylistic change between them. Especially if it's of something like an up close face, where you'd want to see individual brush strokes or the like.

Well, all that being said, I haven't used controlnet tiling in a while so you could be right.

[thorny glacier]

You can also use multiple ControlNETs in case it keeps messing up something. For this one for example, it kept messing up the wheel so I also used the LineArt ControlNET to keep that coherent along with Tiles:

[ocean ermine]

I'll have to look at it and do some testing to see how it works for me.

[thorny glacier]

It’s true that it does struggle a bit more on illustrative artworks—at least when they’re DIGITAL style illustrations. For analog (oil painting, drawing with graphite, watercolor, etc.) it actually upscales pretty well:

[thorny glacier]

Of course. Hope it works for you as it did me! And don’t be afraid to reach out for advice

[ocean ermine]

Yeah, I deal primarily in illustrations and artworks, heh.

You can get pretty far without controlnet though. It's just a matter of learning what works best with the tools you have.

[thorny glacier]

Makes sense! I primarily like to upscale realism/3D render style, so that’s what my workflow is optimized for. I find that Tiled Diffusion generally does a better job than Ultimate SD Upscale when it comes to artworks. This is an example here:

[thorny glacier]

These are really nice!

[eternal fog]

wow the Ultimate SD Upscale takes so long compared to the tiled ksampler I have at the moment

More than twice as long

[ocean ermine]

Nice. Love the painted texture

[thorny glacier]

It definitely does, but when you use it with ControlNET Tiles, the quality is just amazing

[thorny glacier]

Thanks!

[tiny summit]

GiveRep <User:User> [Num:Whole number]

Invalid arguments provided: Not enough arguments passed

[eternal fog]

I'm not finding much improvement with tile compared to what I'm already doing and if I put the denoise up as high as 0.5 like you suggested it starts making noses and eyes in random places. It's not as bad as without Tile, but it still does it.

[thorny glacier]

What was the image?

[eternal fog]

Weird mark on the head, noses and lips appearing in the creases in the clothes.

compared to what I'm currently doing

[thorny glacier]

Can I see the initial generation?

[eternal fog]

As in before upscale?

[thorny glacier]

Yes

[thorny glacier]

I’d actually like to try it myself to see if there’s maybe something else different that I’m doing; I also want to try Upscaling it

[eternal fog]

And Ultimate upscale takes 130 seconds compared to 55 seconds with what I was doing before.

[thorny glacier]

Did you use a 1.5 model to take advantage of Tiles?

[eternal fog]

Yes it wouldn't run otherwise

[thorny glacier]

Makes sense. Anyway, I’ll try it and get back to you

[eternal fog]

I'll run this one again without tile just to make sure it is actually using it properly.

it looks like it is because I see even more eyes appearing on the preview without it

[thorny glacier]

Oof

[eternal fog]

Yeah tile is definately working, just seemingly not well enough

The main thing that bothers me is the horrible plastic skin

SDXL is really bad for it

[cyan crown]

no plastic face

[ocean ermine]

Just try some anti-prompting.

[ocean ermine]

And use a model that's known for making realistic skin. Not fashion models, heh.

[eternal fog]

What model, I've got Juggernaut, nightvision, realvis. All seem to do the same thing for me.

[cyan crown]

last 3 are from standard SDXL

[ocean ermine]

Better than words v3.

[thorny glacier]

Another thing I should mention is that I also have a massive buffer zone (high mask blur and padding) so that it has a lot of context to go off of with no visible seams @eternal fog

[eternal fog]

I tried the settings in the screenshot you posted

I'll try another image

[thorny glacier]

I am Upscaling to 4K (2.9 times) so it’ll be about 13 minutes until I get the image 😅 I’ll show you my results

[ocean ermine]

It takes you 13 minutes to get to 4k?? What's your gpu?

[thorny glacier]

RTX 2070

[ocean ermine]

Ah.

[thorny glacier]

What’s yours? 🤣

[ocean ermine]

3080

[thorny glacier]

I see l for me 💀

[ocean ermine]

Still nowhere near enough to do most decent LLMs and to run SDXL on Auto1111 properly. (but there's other issues with running SDXL on auto anyway.)

[cyan crown]

i'm with auto without problem

[ocean ermine]

Yeah, so everyone keeps telling me, lol.

[thorny glacier]

For real though optimization is just not good there for SDXL… on my machine I can only run 4-bit 7 billion parameter LLM models

[ocean ermine]

Try running multidiffusion with an upscale model (hires fix) on an SDXL model and get back to me

[cyan crown]

I use img2img for upscale

[thorny glacier]

Wait, how long does it take you? I wouldn’t have guessed there was that much of an improvement

[ocean ermine]

And multidiffusion?

[ocean ermine]

Eh, lemme see if I can find that info

[cyan crown]

no

[ocean ermine]

That's the difference.

[cyan crown]

but 70 steps with SDXL in 14 sec on a 4070

[wet nacelle]

4070 with what Vram?

[cyan crown]

12gb

[ocean ermine]

Aight, I'm going to regen this and see how long it takes. #✨｜sdxl message

[cyan crown]

100%|██████████████████████████████████████████████████████████████████████████████████| 70/70 [00:13<00:00, 5.14it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 70/70 [00:14<00:00, 4.79it/s]
100%|██████████████████████████████████████████████████████████████████████████████████| 70/70 [00:16<00:00, 4.19it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 70/70 [00:17<00:00, 4.06it/s]

[thorny glacier]

Well, at least my computer’s saving grace is that I at least have relatively high RAM (32 GB)

[ocean ermine]

Yeah. That's a big boon to LLMs especially. I have 64 myself and I've been able to at least RUN a 70b model. You just have to be super patient with 0.15-0.25 token speeds, lol.

[cyan crown]

64 VRAM?

[thorny glacier]

I am hoping to upgrade my machine eventually to get 16 GB of VRAM and perhaps even more RAM, but for now I have to deal with what I have since getting that can be expensive

[ocean ermine]

Ram.

[ocean ermine]

I'll be honest, I wouldn't recommend anything under 24gb of vram if you're gonna use it for AI and can swing the extra price.

But it's definitely a big difference in price there, heh.

[cyan crown]

ok me too

[thorny glacier]

I mean, you’re right; the more VRAM the better. Of course I’d want to get 24 GB of VRAM, but budget is important too

[ocean ermine]

Just thinking of the minimum requirements of models out there right now. Heck, even the new text to 3d that stability ai released is min 24gb.

[thorny glacier]

It’s really absurd, even if understandable, and I’m hoping they optimize it so us “little” guys can get some too lol 💀

[ocean ermine]

Oops. Left some stuff on for the gen that I forgot to turn off, it'll be a moment, heh.

[ocean ermine]

Left detailers on for the body and face, and they're trying to gen the entire 2696x3936 at once lol

[thorny glacier]

Okay!

[thorny glacier]

OOOF oh god

[thorny glacier]

Okay, I just tried it and it didn’t do any of that. There might be another issue that you have, it’s weird that it’s hallucinating noses there because ControlNET Tiles usually identifies what something is properly…

Denoising strength 0.52

[ocean ermine]

Ok.
Prompt executed in 175.45 seconds

Now granted, that's for 2696x3936

[thorny glacier]

Wow, that’s pretty quick! It looks like they made lots of improvements in the 3000 generation!

True, mine are usually 4K

[ocean ermine]

Let me do 4096x4096 and see.

Though, that's quite a bit more than 4k, heh.

I'll also optimize by changing the upscaler from hugepaint to fatal

[eternal fog]

No idea, I just fed that image in and used the settings in the screenshot you posted

[thorny glacier]

I mean, I usually do 4096~4160 as “4K,” so that’s that. Even though “3840x2xxx” also changes it

[ocean ermine]

Ah, gotcha

[thorny glacier]

Maybe you could try reinstalling your ControlNET Tiles file, or updating. I’m not sure why you’re getting those issues

[ocean ermine]

Oh boy. I can already tell I'm going to like this one.

[thorny glacier]

I like the stylization! Which model are you using?

[eternal fog]

Using the same file as you. Are you on comfyui or Auto, it's possible it's a comfy thing.

[eternal fog]

Could you send that with the workflow embeded, so I can check if I've got some weird setting?

[ocean ermine]

I'm glad you asked! It's Eface's model, BPN (BrightProtoNuke 1.3) https://civitai.com/models/148871?modelVersionId=248861
It's the most cohesive, most prompt following general purpose model that I've used for SDXL thus far. It's amazing.

[elfin flare]

I'm just waiting for X-adapter

so that I can use controlnet tile with SDXL

[ocean ermine]

ʏᴏᴜ ᴀɴᴅ ᴍᴇ ʙᴏᴛʜ

[lapis gale]

wot? 🙂

[ocean ermine]

https://showlab.github.io/X-Adapter/

[ocean ermine]

There we are. 4096x4096.
Prompt executed in 205.92 seconds
About 3 and a half minutes. Not too bad.

[ocean ermine]

The D E T A I L though

[eternal fog]

@thorny glacier Just checked my file and the checksum is the exact same as the one on Huggingface, so no idea why it's not working I've tried again on a few different images and they are all adding too much extra stuff and making it look awful. Nothing like the one you've done.

Hmm, this one is better, but there's still a random face up in the top left

[tough trellis]

Any advice on how to display those motion/speed lines anime-like with SDXL?

Should I look for specific models/checkpoint? Or is it more like a prompt-tuning topic?

[elfin flare]

Negative: Disfigured, Mutated, Ugly, Out of Focus, Blurry, Text, Watermark

Parameters: Steps: 12| Seed: 2351330683| Sampler: Euler a| CFG scale: 2| Size-1: 1080| Size-2: 1080| Parser: Full parser| Model: OpenDalleV1.1| Model hash: 673887ace8| Backend: Diffusers| App: SD.Next| Version: 00281e1| Operations: txt2img| Lora hashes: LCMTurboMix_Euler_A_fix: 813ea5fb1c67, VoxelXL_v1: 796529775279, pixel-art-xl-v1.1: dbe5cdf5f8f4| Pipeline: StableDiffusionXLPipeline| Full quality: True

Time: 19.16s | GPU active 12516 MB reserved 12516 | used 12516 MB free 3740 MB total 16256 MB | retries 0 oom 0```

It took a bit of time to figure out the cfg

since the main model needs a cfg of 8

but the lcm lora I use decreases the cfg requirement way down to 2

Perhaps this wizard will guide my way

with his torch.

[elfin flare]

Maybe I want something a bit more detailed.

But still pixelly/voxelated

[thorny glacier]

Hello! I’m so sorry this took so long, I had some work come up! Here’s the proper workflow and the final image:

The one that I uploaded was something similar I found online rather than tailored to this

[thorny glacier]

@ocean ermine In case you were interested in my Upscaling workflow

[ocean ermine]

Hmm. Why chess as the mode_type instead of linear? What's the difference there?

[thorny glacier]

“Chess” makes the Ultimate SD Upscale do every tile like a checkerboard (not left-to-right, top-to-bottom as normal). Why I do this is because the less linear order makes seams a lot less obvious, or even non-existent

[ocean ermine]

Ah. That's a good idea.

[thorny glacier]

And… this one took me 820.79 seconds (to get to about 4K) 💀 continue

[ocean ermine]

Multidiffusion has always been my jam and I've never liked Ultimate Upscale so I guess I've never looked too far into its specific settings. If you're running that with a 1.x model, isn't the tile size too large at 1024x1024?

[thorny glacier]

It actually works much better than doing it at 512x512 because it dramatically lessens the effects of seams. Also, the model I’m using was trained with 768px as the largest side, so it still remains coherent while reducing overall time and tiles. What’s more, ControlNET Tiles enables the model to maintain consistency at that resolution, much like Kohya’s Deep Shrink does

[modest swan]

Not really sure what channel to ask this question in but here goes... can anybody help me figure out how to upscale without losing detail, particularly in the skin for realistic images? I've tried a dozen different methods in A1111, Comfy, and just tried Fooocus with its own upscaler that has the exact same issue. It will add detail in certain places, like the eyes and teeth and hair may look better, but it flattens out the skin, removes blemishes and marks like freckles, and just generally makes the image look like it was doctored/edited.

I just want to be able to scale up an image while retaining the skin quality of the original.

[ocean ermine]

I also notice that you're running it at a relatively quite high amount of steps for a 0.52 denoise.

Usually I'm only at 12-15 steps for an upres.

[thorny glacier]

26 steps? I guess so.

[thorny glacier]

Oh, I see, I usually do the steps that I do for details

[ocean ermine]

Considering the denoise is only changing the image by half, you would theoretically only need at least half the amount of steps as usual.

But... I've found in the past that very high step counts on a hires upscale can help give extra detail.

So I can't recommend changing it from what you have, lol.

[thorny glacier]

Makes sense lol 💀

[ocean ermine]

I just don't usually do that these days because it can take so long.

No wonder it takes you 13 minutes, lol.

[thorny glacier]

Maybe you might want to reference the workflow I posted above

[ocean ermine]

If only you knew how we've been talking about that exact subject on and off for the last few hours, lol.

[modest swan]

Guess I came in at the right time then

I'll pop back a bit and read the chat, thanks

[thorny glacier]

True… but I guess when it comes to certain specific images, I don’t mind the tradeoff—for many years before AI I was a “traditional” artist, so this is—as you can guess—is a lot faster lol

[ocean ermine]

I get that. As I've told others who've tried to espouse the wonders of turbo and LCM, I'd rather it take 10x longer and have a fantastic generation rather than a quick great one.

[thorny glacier]

Exactly! Thats my philosophy. And I think no image expresses that better than this one, for me. For years, I dreamed about being able to get my characters in a hyperrealistic 3D render style. The best I can do manually is a painterly realistic style. But with AI, inpainting, and more, I can finally achieve that. It’s such an amazing feeling!

[ocean ermine]

Nice! Great stuff

[thorny glacier]

Thank you!

It seems like you enjoy more horror/surreal images?

[ocean ermine]

Absolutely not, hahaha. Just used my previous gens for reference. I generate all types, but they usually fall into anime or non-anime. I actually hate horror movies/games/etc. so I find your statement quite humorous, though I can see how you got that.

[thorny glacier]

Then again, you have an anime profile picture and a Japanese-y name, so I should’ve known lolll 💀😭

[ocean ermine]

You can find a showcase of my best works here: https://civitai.com/user/Yumeshiro

Though I also have a local "hall of fame" folder that I store iconic or specifically good gens of mine and from others to keep on hand.

I'm a big fan of ultrawide, that's for sure.

[thorny glacier]

@ocean ermine This is great! I’ll follow you on Civit. If you see someone named KTTB follow you on Civit, then that was me. I don’t really post much there but I really should

[ocean ermine]

Oh nice. Thanks

Yeah it's great when you can avoid the nsfw stuff.

But things keep finding their way through the filters.

[thorny glacier]

Unfortunately… 😅

[modest swan]

Ok... I'm seeing some good initial results from an SD1.5 upscale with Ultimate Upscaler at around .25/.3 and controlnet tiling

Much better than what I was getting

[thorny glacier]

I’m glad our discussion could help!

[modest swan]

Thank you 🙂

[ocean ermine]

You ever work with dynamic prompting and wildcards? They can really help with coming up with testing subjects/backgrounds around a style or theme.

[thorny glacier]

Oh my god YESSSS here’s some I got recently with a somewhat complex prompt:

[ocean ermine]

Oh wow! I love the look these have

It's like straight out of a bing background or something lol

I like the orb, heheh

[thorny glacier]

The prompt was: photo of old {Japanese|Atlantis, flooded|European| |Shirakawa-go|Middle Eastern|mystical fantasy village full of treehouses|jungle|Inuit|Alaskan|creepy|Kidcore, pastel colors, curved whimsical architecture|apocalyptic|Russian|European castle|Thai temple|massive colorful fantasy trees massive curved church|African tribal|African|futuristic tiki|rocky tropical Hawaiian volcano|ancient Mayan mountain|ancient Aztec canyon|serene Zen Thai waterfall cyan buildings mountain|full of bubbles spherical buildings|colorful curved exotic tropical fantasy village with terraced rice paddies|pastel colors exotic angular triangular fantasy village with massive mushrooms|whimsical curved Art Nouveau architecture buildings|intricate steampunk} {village|town|city|slum}, {evening|noon}, clouds

Negative: (Semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime) text, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly

[thorny glacier]

Thank you! Yeah, I just fired off whatever I could think of for the wildcards

[ocean ermine]

I've put together a list of backgrounds and D&D classes, armors, etc wildcards, over the past year. My original intention was to make a D&D portrait creator, and I think I achieved that to a certain extent at the time, but I need to get back into it with SDXL being so much better at recognizing everything.

Here's a few background wildcard examples with a dark style I was testing

[thorny glacier]

SDXL is really a lot better at getting characters right; like, if you want two subjects, 1.5 almost always merges or duplicates just one of the subjects, but SDXL actually respects it. It’s also better at generating random characters like Doraemon or SpongeBob, even if not perfect

That’s beautiful

[ocean ermine]

Yeah, they're super fun.

And as you can see, all illustrations, heh. When I try the controlnet tiling I really hope it's able to keep that. I'll probably have to use Dreamshaper, SDVN3RealArt, or something similar to retain the art style.

[thorny glacier]

These were some I enjoyed too:

[ocean ermine]

Haha love the jeep through hobbiton

[thorny glacier]

For illustrations, I actually tend to use Deliberate v2 since it’s a very generalist model, but it depends on what you want; it’s best to experiment

[ocean ermine]

Wow, that's an old one.

[thorny glacier]

It’s definitely quite old, but it still works well

[thorny glacier]

Yeah, that one was legit hilarious!

[ocean ermine]

The best for illustrations that I've found with 1.x was probably Dreamshaper v7. That was before it got super photorealistic

[thorny glacier]

I should probably try that version of DreamShaper out then; as it’s one of the most popular models of course I have it, just not that version…

[ocean ermine]

This was originally with Dreamshaper 7. Slightly photoshopped in one tiny part.

[thorny glacier]

Okay, I’ll definitely have to download it now, that is pretty good; I’m always on the lookout for models that can do paintings better

[ocean ermine]

I believe this is also dreamshaper if I'm remembering correctly.

[thorny glacier]

And then of course, don’t forget about mixing models, either (like generating with one model then switching to another to exploit both). That enables for even more powerful results

[ocean ermine]

Yep. That's how I made this. Started with Dreamshaper or something similar and then switched to an anime model.

Looks like it's DreamshaperV8 to BreakdomainAnime

[thorny glacier]

That one looks so much like a human-made artwork, it’s actually crazy. AI isn’t usually great at low-frequency (lots of flat, iconic shapes that often have to be just right) styles. The texture is on-point

[ocean ermine]

Such things are often more easily achievable by breaking the model through composable diffusion. Essentially telling it to make two things at once. This allows for very interesting results.

[thorny glacier]

Do you have any workflow? I’m intrigued

[ocean ermine]

Much of this is wildly different through the months, heh. The ones above are when I used 1.x with Auto1111. Were you looking for something in particular about a workflow?

[thorny glacier]

If you had a more specific one; if it’s just with Auto1111 I could figure it out easily

This was a unique one that I upscaled using Deliberate; again, it’s still a pretty powerful model:

[ocean ermine]

Oh neat!

[thorny glacier]

Some work came up so I’ll have to leave for a little bit. Happy generating!

[ocean ermine]

Thanks!

[bitter juniper]

polishing the art style...

[ocean ermine]

Was getting a lot of hallucinations at first, but seems to be better at the moment for some reason. The single upscale with controlnet really does help keep the style of the original gen. Thanks for the upscale tips!

[spice heath]

Are there any tricks to getting text output with things like HarrLogos? I'm having a <1% success rate getting the text I want to show up

[sharp mason]

Anyone here been able to finagle sdxl models to work on 5xxx series Radeon cards (like a 5700xt)?

[peak dove]

I was seriously contemplating a 4090 system; but it seems to have a "melting wires problem!" Then again, its price'll drop, as the new 5090 is coming. Mebbe I'll invest in a 4090 system in the Summer?

[tough bough]

The 4090 issues were a storm in a tea cup. They are fine so long as you plug the power cable in properly. Gamers Nexus and others did a good breakdown for this. They are still bonkers expensive though and likely to stay so as nvidia is using all their production capacity on AI server chips.

[lapis gale]

fully agree, and IIRC they now have better connectors as well (modified to avoid mistakes). important to have enough space in the case and no pressure/pull on the cables to be on the safe side. i measured the 4090 during the burnin test with an IR camera and all was fine

[steady grove]

yeah. The vast majority of 4090s exist without issue. Youtubers signal boost anything that get clicks and views.

Seemingly reputable outfits like gamer nexus doing a lot of quality reputable journalism, are actually largely a yellow journalism outfit. Most of their content is overblown and they're closer to british tabloids and tucker carlson style reporting than they are any kind of real industry whistleblowers. They might even have the same PR outfit as drama alert / keemstar. i wouldn't be surprised

ltt have been doing clickbait marketing for years too and only fluff up their reputability as much as they can, while they operate first and foremost as a yellow journalism outfit

same thing is happening with starfield lately. Drama got slow for a while, so they started manufacturing this whole situation that starfield is a bad game. Meanwhile, the vast majority of players enjoy it fine.

[peak dove]

Portrait-Master v2.3 #ComfyUI #SDXL

[slender coral]

Anyone have a comphy node suggestion to take a new line from a text file for each prompt?

[shy kelp]

very creative

[lapis gale]

santa, fighting stance

[ionic dragon]

https://fxtwitter.com/EMostaque/status/1738613529169563897

When is our new Xmas present dropping

[snow wasp]

Trying to figure out if I can stream and generate shit at the same time

https://www.twitch.tv/bigmarklarper

[shy kelp]

they should. i feel stable need some hype

[bitter juniper]

Bow before my Jaffa...

[jolly brook]

https://tenor.com/view/apophis-stargate-helmet-armour-gif-18440901

need serpent guard lora

[bitter juniper]

https://tenor.com/view/stretchy-pants-nacho-libre-gif-1496345603287220583

[native knot]

Sithmas Tree:

[crisp owl]

@cyan crown By assistance of your renaissance Lora, plus some other magic

#✨｜sdxl message

[cyan crown]

very fine!

now I'm working on 3 new loras

[crisp owl]

Nice, I really like how Lora's play with SDXL compared to 1.5 models. They fit so much better

I never used em in 1.5 because they were wonky often

[cyan crown]

yes in some way Lora with sdxl makes some magic to the base model

this is a result I'd like to achieve with next one

[rich crescent]

what is the best refiner?

[rustic garnet]

no one

most models do great without refiner

[ocean ermine]

I'm surprised how well some Midjourney prompts I've found online are working with SDXL. The gap is a lot smaller than most think.

[west breach]

https://civitai.com/models/141592/pixelwave

New version of PixelWave out now!

[cyan crown]

https://civitai.com/models/243571

this is the first of 3 loras I'm working on

[cyan crown]

I've also made the 2nd Lora, but it's not perfect so at the moment it's unpublished

These are some results

[lapis gale]

merry termination

[cyan crown]

it's not so bad but I must use with 0.6 Strenght

[lapis gale]

i also admire the SDXL anime capabilities, good enough for me, haven't touched a "real" anime model for ages

[stray wave]

What is the best model like Rev Animated?

[lapis gale]

what is "best" for you? style? anatomy? lighting? composition?
i was a big fan of revanimated back in the days, but IMHO it's overtrained, all the women look the same

only SDXL for me now

[swift wind]

Why are my images so bad and also in bad quality

[lapis gale]

just kidding

[lapis gale]

that's more a question for #🤝｜tech-support than this channel... what are you using, what are you trying to generate, what is the result instead, what do you consider "bad", etc etc

[cyan crown]

So I published a revised version

https://civitai.com/models/244376

[cyan crown]

ok stop

[slender coral]

Anyone know if it's possible to use alpha in inital image prompt (Without removing backgrounds)? The black is the alpha channel in my inital image, exported as 0,0,0 for an output.

[left dock]

https://civitai.com/models/43977?modelVersionId=262973

[dim steeple]

is this contorlnet for sdxl? https://huggingface.co/webui/ControlNet-modules-safetensors

or is it 1.5

[cyan crown]

https://civitai.com/models/244499

I've also made another Lora, but I don't think I'll publish it because I don't know if people can like this style. I post some images.. What do you think ?

[bitter juniper]

I love it but I wonder if it is not possible to get the same thing with only prompt

[lapis gale]

looks quite cool... i'm trying to approximate it with "high contrast charcoal drawing, white and red and black color scheme, a rose, grunge red background" but it's not 100% the same

[bitter juniper]

probably need to add something like red monochrome and ink color

[lapis gale]

"high contrast minimalist charcoal drawing, white and red and black color scheme, a rose, grunge red background , smears, crimson"

[bitter juniper]

or ink drips

[lapis gale]

"red background, oversaturated high contrast minimalist abstract schematic rough charcoal drawing, red and black color scheme, black cat in a poppy field, grunge red background , smears, crimson"

knowing some art terms helps 🙂

"red background, oversaturated high contrast minimalist abstract schematic rough charcoal drawing, red and black color scheme, samurai warrior moon, grunge red background , smears, crimson"

all with sdxl (realisticfreedom), no loras/embeddings

[bitter juniper]

not red enough

almost

[lapis gale]

"red background, oversaturated high contrast minimalist abstract schematic rough charcoal drawing, red and black color scheme, ((horus)) egypt falcon head god full body, grunge red background , smears, crimson"

this is fun 🙂

yeah, and her 🙂

[bitter juniper]

Oh yeah....

I got it

love this art style...

oh yeah...

[cyan crown]

again result are similar but different, because subject are too defined with only prompts

[lapis gale]

rock on

[bitter juniper]

I got an idea....

[lapis gale]

needs more metal 🙂

cinematic style instead

psychedelic

cubist, and done for now

[fierce hollow]

Has anybody managed to train a lora for segmind's vega/ssd-1b or turbo? kohya doesn't work with these apparently but also getting some cryptic errors with peft, not sure what sort of tinkering this requires

[elfin flare]

He is ready.

[bitter juniper]

I cannot wait to master Animatediff and SVD.... I need to create video with these loras....

[bright valley]

Have you tried the HxSVD Workflow?

[bitter juniper]

I will try it soon. Thank you

[bright valley]

Absolutely! v2 makes animations like this, 2 SVD passes, and RIFE interpolation, ending at 16sec of animation, but it "pingpongs" by default
(Goes to the end, and reverses back to the start)

[bitter juniper]

I have to make it work...

[meager canopy]

https://tenor.com/view/walk-like-an-egyptian-egyptian-the-bangles-bangles-bangles-egyptian-gif-16182356

[lapis gale]

good prompt idea 🙂

retro!

[vale eagle]

https://civitai.com/models/160738

[cyan crown]

new Lora is on the way

[bitter juniper]

from this

to this

or this

[lapis gale]

fight 🙂

[lapis gale]

ha... "a cat themed darth vader chibi", what do you say now 🙂

[left dock]

Guys, a question, what should you use for checkpoints for hyperrealism?

[cyan crown]

Elsa from Frozen

[cyan crown]

I use SDXL Base plus loras

[left dock]

Thank You!

[native knot]

I had a bunch of these already....lol

[cyan crown]

Lora is baked. It's too late I'll publish tomorrow. It's the strangest Lora I've done till now

well in the end I published.......

https://civitai.com/models/246474

[bright valley]

Are these with harrlogos?

[native knot]

No.

[high swift]

Hi

[ivory solstice]

What works better as SDXL upscale: "Tile resample" XL controlnet, or Kohya highres fix with img2img?

Want your opinion guys

For me, the latter is better as it doesn't add depth of field

[wise helm]

#1134502659647348756 a cup of tea on a table with toxic vapor coming out of it, ultra realistic, canon camera, realism

[peak dove]

I adore a new lora 🙂

[peak dove]

Using Future Cubism LoRA from @cyan crown 1st picture 50%, then 65%, 75%, 85%, 100%

Without Future Cubism LoRA

No LoRA

[peak dove]

This is cool! Way good

[peak dove]

A mixture of LoRAs - Shaun (the Sheep); Melting Candles, and Futuristic Cubism, all in one workflow.

[molten gull]

...you nearly got it, ai 🙂

[cyan crown]

https://civitai.com/models/247039

This style is a bit 'disturbing

[cyan crown]

yes that's easy.. Try to match Joker o Pika i.e. without IPADAPTER

[vale eagle]

just try some idea on my new fine tune.

[cyan crown]

it's good but it's quite different

[left dock]

can someone suggest a good prompt for realism?

[vale eagle]

use camera and year for realism

[bitter juniper]

I rush my apophis lora... It doesn't work as I want...

https://tenor.com/view/failure-feel-like-a-failure-im-a-failure-failed-disappointment-gif-15601002

so average...

[bright valley]

Lol, okay. What did you use to make them? They're pretty cool

[native knot]

They were just prompt work and an SDXL model...probably JuggernautXL, via ComfyUI. But it's been a bit, so I don't recall.

[bright valley]

Ah interesting, thanks

[bright valley]

Reminds me of this epic post in my models gallery, that's where I got the image I posted from
https://civitai.com/posts/830731

[bitter juniper]

Jaffa !

[upbeat goblet]

basically "evil goddess of primordial chaos" etc, mostly img2img

[peak dove]

Using Mixlab Nodes - this is Node 5 - GPT

[peak dove]

GPT phrased prompt Mixlab Node 5

[lapis gale]

smile!

[dim steeple]

Is rendering at higher square res than 1024x1024 going to ruin my image?

using img2img and controlnet

[molten gull]

happy eiffeltower year ! 🙂

i wonder how the happy new year 2025 pictures will look like 🙂

[cyan crown]

Testing 2 new loras

[cyan crown]

https://civitai.com/models/248269

[peak dove]

D/loaded

[cyan crown]

hope u like

the other Lora is a 50/50. Good but not great. It should emulate a toy camera. I don't think I'll publish

what do u think? it worth to be published?

[cyan crown]

well playing with it it seems better than I thought

https://civitai.com/models/248225

[lapis gale]

i think the analog SDXL style gives a nice retro look with better quality (ok the toy camera is supposed to be... a toy, i get it)... not sure

but we'll see how the downloads evolve 🙂

and i do admire your creativity

[cyan crown]

thanks. Yes this doesn't want to give an analog look. I wanted a toy camera like a lomo.