✨｜sdxl

[copper kraken]

It can get in the ballpark but it struggles with things like word count targets

Ahh gotcha

[gloomy lark]

yeah exactly. gpt would do it, but it would put stuff in random boxes. tail over here, body on the other side of the yard type of thing.

even with regional prompting, I don't think I can match this sd3 shot

[copper kraken]

Dark arts kicks ass at the original clownshark prompt

[copper kraken]

Miiiiigtt be possible maybe with comfy nodes for that

Def won't commit to saying I could do it though lol

"A high quality cinematic photograph of a Pickle clown shark tickle in a messy apartment, garbage clown freak shrooms, gun tree, missile launch snake, drug shrimp"

Certain combos of words just drive sdxl bananas

[gloomy lark]

ok I'll try that one with claude

[copper kraken]

Claude rewritten prompt? Or does it have an img gen too

[gloomy lark]

it's making the prompt

I think the censors are kicking in for the missle launch and drug shrimp. so what it's giving is seriously twisted.

[copper kraken]

Wow that prompt blows sdxls mind so badly it got carosello to generate photos... Nothing I've tossed at it just did that

Those are great

[gloomy lark]

A pickle clown shark tickle, in the style of photorealism ADDCOL mushroom clown freak, in the style of surrealism ADDCOL gnarled gun tree, in the style of dark fantasy ADDROW missile launch snake, in the style of sci-fi ADDCOL drug shrimp chef, in the style of satirical realism ADDCOL messy apartment interior, in the style of gritty realism

whoa. better one coming up

this is right up your alley for random stuff smashed together.

A whimsical pickle head sporting a jester's hat, in the style of Salvador Dali ADDCOL
A shark fin protruding from a tangle of human limbs, in the style of Hieronymus Bosch ADDCOL
A bouquet of vibrantly colored fungi, in the style of Ernst Haeckel ADDROW
A pistol-gripped branch laden with bullets, in the style of Giuseppe Arcimboldo ADDCOL
A serpentine rocket exhaling smoke rings, in the style of M.C. Escher ADDCOL
A shrimp adorned with pill capsules and syringes, in the style of Yayoi Kusama

[copper kraken]

Dark arts goes haywire with this one "A high quality cinematic photograph of a Pickle clown shark tickle cannibal in a messy apartment, garbage clown freak shrooms, gun tree, missile launch snake, drug shrimp, song bong gong gang bang, warp slug, beep freak"

Is that Claude?

I love whatever that hat head is

With the eyes staggered in a ring lol

[gloomy lark]

i'm impressed that it can do photos as well as illustrations

so the prompt was generated by claude, using SDXL regional prompter with dark arts model

[copper kraken]

Dark arts getting its brain fried ^

Carosello can handle it

[gloomy lark]

ideogram doesn't care about violence, but it rejected this one. 🙂

probably because of the shrooms

[gloomy lark]

it doesn't always work with dark arts, but the cute 3d render lora sometimes adds a depth to it.

[copper kraken]

Oh that's fantastic

Forgot to try that

I want to make the cutest menace

Q

[gloomy lark]

claude 3 rocks. This was from highly detailed panther with jewelry in blue lighting with large paws and sharp claws is gripping into a tree

[copper kraken]

Wow

Gonna have to check this out tomorrow

[gloomy lark]

good stuff

[gloomy lark]

yeah that's really good. the weapons and pointy faces are neat

[copper kraken]

Check out those with pnginfo

That was gpt 4 output

Dumped a big chunk in there for safe keeping lol

AI really is very good at prompting ai... Lol sounds like the opening of a Terminator sequel

[gloomy lark]

here i am thinking that new 512 token limit for SD3 is awesome and you throw 732 tokens at it

whoa. that might be the best moby dick novel in the prompt box to image quality I've seen yet

I assumed that was regional prompter on your part

[copper kraken]

Hahaha

Nah just dumped it in

I think forge might be splitting the prompt in blocks and concatting them

I've noticed it always shows tokens used in increments of 75 and everything in that prompt affected the output

[gloomy lark]

right

i'm assuming it's forgetting a lot of it, but it's also keeping a lot

1 of something is hard, 10 of something is easy

[copper kraken]

Those numbers are why I'm wondering about that

Once you hit 76 its 76/150, then 151/225 etc

Like it's splitting it into blocks of 75 and concatting all together

I need to do some systematic tests on this

[rustic garnet]

do they concatenate them? I thought they blend/average them, but I'm not sure.
In principal SD could take any number of tokens, but it was trained with a fixed length of 76 tokens. Similarly, CLIP was trained with a maximum of 76 tokens, so it doesn't work well if you add more to it

[stray warren]

Would you mind sharing your instructions to Claude?
Not to Claude, I mean sharing the instructions you use for Claude

[mellow tendon]

How do you get Claude to generate images? or did it just give you a better prompt for it?

[shut lantern]

Thank you, let me try i t

[gloomy lark]

I want you to create a terse text to image prompt. For that image, I want you to split the image up into 6 pieces; 2 horizontal rows of top and bottom, and 3 columns per row, left and middle and right. I want you to describe what is in each piece, starting from top left, then top middle, then top right. Then for the bottom row, bottom left, bottom middle, and bottom right of the image. Use the word ADDCOL to delimit columns and ADDROW to delimit rows. Don't add words denoting which image piece it is. Don't mention more than one subject per image piece. Put each image piece prompt on its own line. A subject is allowed to span vertical image pieces. Determine an appropriate artistic style for the overall image and mention ", in the style of " with that style at the end of each prompt line, but before the row or column delimiter. Please make a text to image prompt for:

[gloomy lark]

I want you to create a terse text to image prompt. For that image, I want you to split the image up into 6 pieces; 2 horizontal rows of top and bottom, and 3 columns per row, left and middle and right. I want you to describe what is in each piece, starting from top left, then top middle, then top right. Then for the bottom row, bottom left, bottom middle, and bottom right of the image. Use the word ADDCOL to delimit columns and ADDROW to delimit rows. Don't add words denoting which image piece it is. Don't mention more than one subject per image piece. Put each image piece prompt on its own line. A subject is allowed to span vertical image pieces. Determine an appropriate artistic style for the overall image and mention ", in the style of " with that style at the end of each prompt line, but before the row or column delimiter. Please make a text to image prompt for:

[copper kraken]

The way you describe it makes more sense intuitively, I'm referring to the name of the comfyui node that behaves pretty similarly to BREAK in a1111/forge

I just did a test

checkpoint: darkArtsImages_v10Abyss, steps: 20, sampler: dpmpp 2m, scheduler: karras, seed: 2711398504

clown clown clown clown clown clown clown clown clown clown
shark shark shark shark shark shark shark shark shark shark

clown clown clown clown clown clown clown clown clown clown BREAK
shark shark shark shark shark shark shark shark shark shark

clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown
shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark

clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown clown BREAK
shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark shark

so if you have two chunks that are each 75 tokens, separating them with "BREAK" makes no difference

so i think it must be using that when you sail over the limit

[gloomy lark]

(clownshark:87.3)

[copper kraken]

shark
clown
Volcano
Skyscraper
UFO

shark BREAK
clown BREAK
Volcano BREAK
Skyscraper BREAK
UFO

[meager canopy]

WTF?! @copper kraken

[gloomy lark]

interesting. merge vs. also

[copper kraken]

Typing past standard 75 tokens that Stable Diffusion usually accepts increases prompt size limit from 75 to 150. Typing past that increases prompt size further. This is done by breaking the prompt into chunks of 75 tokens, processing each independently using CLIP's Transformers neural network, and then concatenating the result before feeding into the next component of stable diffusion, the Unet.

[copper kraken]

I don't think it does? Could be wrong

[gloomy lark]

Ok I tried your prompt and it just said "RuntimeError: Sizes of tensors must match except in dimension 2. Expected size 7 but got size 8 for tensor number 1 in the list.

(claude:3)

[copper kraken]

shark
BREAK clown
BREAK Volcano
BREAK Skyscraper
BREAK UFO

shark
BREAK
clown
BREAK
Volcano
BREAK
Skyscraper
BREAK
UFO

[meager canopy]

Just gives me this, with or without the BREAK

[copper kraken]

shark BREAK clown BREAK Volcano BREAK Skyscraper BREAK UFO

[copper kraken]

Same checkpoint same seed same params?

[gloomy lark]

RuntimeError: Sizes of shark queue must match except in dimension 2. Expected size 5 sharks but got size 87 for shark number 1 in the list.

[stone fossil]

You need more breaks in your break so you can break while you break.

[meager canopy]

Yes. It seems to discard all the "clown" tokens, or it thinks there is a real "clown shark"

[copper kraken]

Weird

[meager canopy]

Ah...Cascade

[copper kraken]

Oh that keyword isn't in comfy

The comfy equiv is the conditioning concat node

Take the parts on each side of BREAK in separate prompt boxes and "merge" them using the conditioning concat node

[meager canopy]

I don't think I ever go over the 77 tokens anyway 😄

[copper kraken]

This was 732 tokens lol

[copper kraken]

I found last night that having gpt 4 generate a very detailed description of god knows what will almost always result in a very interesting image, even with it being 400+ tokens

Yeah, there's obv tons of prompt neglect but every chunk does indeed have an effect

[gloomy lark]

I have the regular image generation prompt if you're looking to stop the prompt vomit from gpt

Act as a creative agent who generates a terse but highly creative image prompt derived from the prompt I send you. Include descriptive visual elements of the subject, lighting and surroundings. Specify an artistic style or camera settings at the beginning of the sentence, using descriptive elements that pertain to this artistic style. Include no more than 10 elements presented as discrete descriptors in one long sentence without story. Put the most important descriptive elements at the beginning of the sentence. Here's the prompt:

You can make it bigger or smaller by adjusting that 10 elements number

I bring that way down for cascade since it does best with prompts way below 77 as many have pointed out.

[copper kraken]

Thanks haha I will def give that a shot. Just been finding it interesting to see what happens when you use plain English descriptions and concat all the chunks together vs writing something that describes something similar that's within the 75 token limit

If you're trying to get something specific... Not a great approach

[mellow tendon]

But this only works if you are using an extension like Regional Prompter right? The ADDCOL stuff seems to have no impact in Autmatic1111 with SDXL without an extension like https://github.com/hako-mikan/sd-webui-regional-prompter

[gloomy lark]

Correct. You need that.

[gloomy lark]

Set it for attention instead of latent, and turn on the don't convert and and break at the bottom. Do not set a resolution and column number etc below, those tags in the prompt itself will create the columns and rows.

[copper kraken]

With your prompt festivalman

[gloomy lark]

You're welcome. lol

Even without Claude, I'll just take what it creates for me and start separating things out in the prompt box to have more influence on where things go

I like the little house on his back with the big ham radio antenna

Broadcasting shark waves

I also look forward to a better vae in sd3 so everything futuristic isn't "neon lit"

[meager canopy]

https://tenor.com/view/rainbow-rainbow-coke-neon-neon-rainbow-coke-neon-rainbow-cokegif-gif-12086451

[wet nacelle]

@meager canopyBurger

[meager canopy]

It looks quite edible 🤤

[wet nacelle]

https://tenor.com/view/jerma-eat-burger-whopper-gif-19831425

[meager canopy]

That was mine! 😦

Stop...I need food now

[stray warren]

Thank you very much for sharing. I'm afraid this ADDCOL and ADDROW won't work with Draw Things, but still very helpful

[gloomy lark]

Sometimes the local mixtral model can churn out good stuff as well.

Anthro horse in a bar setting, drinking beer, in the style of Impressionism.
ADDCOL
Anthro horse with a mustache and monocle, enjoying a pint of beer, in the style of Pop Art.
ADDCOL
Anthro horse wearing a Hawaiian shirt, holding a coconut filled with beer, in the style of Tropical Art.
ADDROW
Anthro horse wearing a beret, sketching another anthro horse, in the style of Cubism.
ADDCOL
Anthro horse sitting at a bar, sipping a martini, in the style of Film Noir.
ADDCOL
Anthro horse dressed as a bartender, pouring beer for a patron, in the style of Art Deco.

[pallid path]

oh this is great

[willow dagger]

Uhm... just had time to check your workflow out... and that has nothing to do with using Differential Diffusion without putting the whole image through VAE encoding first.
Your source image gets turned into a latent right away - and there is no step that combines the original pixel image with the difdif'd latent/image.

I think we were not talking about the same thing / process.

[copper kraken]

You can do the same thing with image masked composite

[willow dagger]

Alright. I'll try that.

[copper kraken]

Yeah just add that at the end

You'll have to composite all the masks you use in total of course

But if you're only doing one in painting step then you don't need to worry about that

[willow dagger]

I got it. Now that of course also leaves the usual slight differences because the inpainted area got Re-VAEd of course. But at least now I can always choose what is more important to me.
Thanks.

[stone fossil]

xD

[sharp scarab]

https://www.youtube.com/watch?v=xfF69OTaJi4

[orchid cobalt]

how to change folder to which comfyUI saves images?

[gloomy lark]

in the save image thing, just put foldername/filenameprefix

no leading slash

[orchid cobalt]

I am confused... Is it not possible to declare a full path? if it is, where would the subfolder end up?

[gloomy lark]

it starts in the output folder, so if you put sdxl/comfy, it'll put it in comfyui\output\sdxl\comfy934893.jpg

haven't tried a full path though.

[orchid cobalt]

alright thanks

[wet nacelle]

ah shweet

[willow dagger]

Did you test the Puppeteer prompt in XL? I'm actually surprised that he posted that one when it didn't include the puppeteer at all.

[gloomy lark]

the one I just posted above is SDXL with regional prompter

an evil pupeteer looking down ADDROW
an evil pupeteer looking down ADDROW
evil pupeteer hands ADDROW
a cute rusty robot head made of garbage ADDROW
a robot torso made of garbage with a sign with "I M SD3-BASED" written on it ADDROW
rust robot legs

but yeah, he posts an SD3 picture and it doesn't actualy prompt adhere. that said, I think his prompt is kinda messed up, so I don't blame SD3 yet.

[willow dagger]

That's actually great. You've been playing with that extension all day, eh?

[gloomy lark]

When I get moments. 🙂

[gloomy lark]

Some from today.

[pallid path]

Holy crap the one with Donald Trump and the scientist

I love these

Sorry if you were asked 300 times before but is this regional prompting in Comfy or in Auto1111/Forge

[gloomy lark]

automatic. the functionality for it in comfy is extremely complicated, just ask clownshark batwing. he's spent the time to get it working in comfy but it's like a swiss watch in there.

automatic is just throw simple column and row markers and you're done.

[pallid path]

I have seen some examples with noisy latent compositing and area compositing

but those require literal pixel values

[gloomy lark]

that said, specific models with specific settings work best. others just don't do much of anything

[pallid path]

Yeah.. Thanks for informing

I'll see what I can do in Comfy

[gloomy lark]

these settings and model work really well

[pallid path]

Thanks

[pallid path]

You should.....

Treat yourself.. NOW!

[elder plume]

https://huggingface.co/lllyasviel/sd_control_collection/tree/main?clone=true
https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/tree/main?clone=true
https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0/tree/main
is this a representative sample of the relevant controlnets for sd1.5 and sdxl?

am i missing anything significant?

[elder plume]

what interface is this from?

i have the ip adapters in a separate list

[gloomy lark]

automatic 1111, 1.5 and sdxl, for ControlNet v1.1.440

[elder plume]

is there even an instruct pix2pix for sdxl?

it doesn't look like htere is

[gloomy lark]

it's in the picture.

[elder plume]

i'm saying that as far as i know, "InstructP2P" has only been created for SD1.5

alright thank you

[copper kraken]

there's some other biggies... there's inpainting and tile ones for sdxl

but also illumination ones for sd15

and the qr code monster ones, an opticalillusion one

mediapipeface

[zinc cargo]

is there a SD3 channel already?

[copper kraken]

i'll get you a list of some later for sure

i've collected basolutely everything i could find

[upbeat summit]

I guess for people that have been invited to the preview

[south jasper]

Can any one say free ways to generate best ai images

[sturdy kindle]

nihaoya

[copper kraken]

Here is the image you requested.

[meager canopy]

It is possible.

[hardy pike]

red dog

[slim wren]

🔴 🐕

[nova temple]

been trying to train lora on sdxl but the speed, it feels unreasonable slow at around 15-30x slower than sd1.5 lora training, any1 got settings that work well with 2080ti? i use https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

[livid basin]

Looks like more SD3 pictures showing up on twitter

[tawdry current]

It's going slower because you likely don't have enough vram. If it uses any shared memory at all, it will be exponentially slower. (Shared memory is in your regular system ram). SDXL base resolution is 1024 vs 1.5 being 512.

There's a huge latency overhead when having to shuffle calculations back and forth between ram<>vram if your vram is maxed out.

[quasi elk]

Does anyone have a working SDXL controlnet adapter for lineart?

[sharp scarab]

https://www.youtube.com/watch?v=xfF69OTaJi4

[glass forge]

This page is like a bad acid trip.

[tepid acorn]

Architectural clausura. From a modern city to a city of the future: man, environment, technology

Subject:
The territory of the center of the residential area

Goal:
To design these spaces as a kind of transition to the city of the future, corresponding to historical and cultural traditions and images of identity, socio-demographic, natural, aesthetic, environmental and technological trends of modernity.

Tasks:
In the project, it is necessary to ensure a balance of three groups of criteria: environmental, aesthetic and functional.
Landscaping elements should use the latest achievements (existing or projected) in the construction and finishing materials industry, eco-technologies, “smart technologies” that contribute to the creation of an environmentally friendly living environment.
To offer a graphical interpretation of the phased implementation of the project from the present to the future.
To express the spirit of the place of design (linking to a specific territory you have chosen).

/prompt home blue

[copper kraken]

Here is the image you requested.

[copper kraken]

Sadly no not that I know of

[quasi elk]

that's very unfortunate since my entire workflow depends on me first drawing the lineart before letting ai color it in, so SDXL is essentially useless to me... unless there are some other ways?

[south flume]

Hi.Can anyone explain the diference between RealVisXL 4 and the Lighting variant? Plus, its not explained in detail, is one supposed to use it with SDXL refiner? https://civitai.com/models/139562?modelVersionId=344487

[heady vale]

same model. the lightning one generates faster in less steps. like turbo model variants

[south flume]

Tks, so I guess lower RES as well, BTW: do you happen to know if its expected to use SDXL refinder in the normal variant?

[heady vale]

you will get almost the same detail and res. No refiner needed

[south flume]

ok,tks

[rustic garnet]

Lightning is like Turbo but works on normal SDXL resolutions

(however, most Turbo merges also work on normal SDXL resolutions. So it's subjective what you like more)

I never use Refiner and I think most people don't use it. Custom models are often much better than refiner anyways

[south flume]

I guess Turbo has its uses, but I just really want higher RES, so I stick with SDXL , Trying variants , but not sure if those need refiner or not

For lanscapes I ussualy use base+refiner at 50/50, and get what I consider good results, but maybe I'm wrong

[rustic garnet]

as said: the Turbo Merges support native SDXL resolution

[south flume]

With basic SDXL, that is , just trying variants now, I'm mostly into real landscape scenes

[rustic garnet]

Turbo -> 1 step generation, but only 512y512

[south flume]

Er..isnt Turbo supossed to work on 700 something pixels?

[rustic garnet]

Turbo merge, LCM, Lightning -> 6 step generation, but 1024x1024 (and other) resolutions

yes, but turbo merges are turbo models merged into sdxl models

thus, they behave somewhat in between

[south flume]

Haaa, so Turbo merge is yet another variant?

[rustic garnet]

yes

[south flume]

Ok, got it now

tks for explaining , so something like that REasl Vis Lightning should get comparable results but with 6 steps or so?

[rustic garnet]

yes

I only use classic models with more steps if I need negative prompts or have very complicated prompts

cause these low step models need very low cfg

otherwise, I get MUCH better results with, say, DreamshaperTurbo than with SDXL base+refiner

[heady vale]

if you want larger res gens you can use the Kohya deep shrink node in comfyui. Ive done many 3840 x 2160 native gens with turbo models

[south flume]

Ive gone really huge before, using just regular SDXL, its tricky but doable , Im no expert but got decent results with Controlnet Canny and somewhat higher than regular denoise for upscale (in A1111) , I+m not really into speed, just want better (realistic) results

I mostly use img2img

From this rough client provided as reference:

[pallid path]

highresfix before and after

[upbeat summit]

model: Aetherverse Lightning by @icy brook

[zinc cargo]

with a lora?

[gloomy lark]

No Lora, it's just the dark arts images model from civitai

[twilit trellis]

imagine prompt:A close-up image of a vibrant, colorful Jungle Juice Popper bottle surrounded by dense tropical foliage and exotic fruits, under a mysterious, moonlit jungle canopy. --v 6 --ar 16:9

[shy kelp]

😭 my damn lora doesnt want to learn

😔

is the LR too high?

[south flume]

Ok, got it now, a person gets out of touch for a month or so and everything is changed, so Lightning is actually a speedier Turbo version, so what models have base resolution of 1024? All Turbo merged,Turbo and Lightning or just some of those?

[stone fossil]

Myself I would not recommend to use any fast / turbo models as they degrade the output to much vs clean XL (finetunes), the only one method using mentioning is using DeepCache but that also degrades.

Unless you need it do something video like.

[south flume]

No idea what DeepCache is, you have any link I can read about it?

No, I'm using mostly to enhance rough images composed on PS

Mainly landscape backgrounds or props/extras to include in my real work

[stone fossil]

You can download a image of my Rembrandt model page: https://civitai.com/models/282340/rembrandt-infinity-sd-xl-10
I believe all workflows where done with it.

But once again only if you want to degrade your image go turbo / fast / deepcache.

[south flume]

Civitai webpage isn't very mobile friendly is it?

[native knot]

Works fine on mobile for me.

[south flume]

What browser? Firefox doenst pan

[native knot]

Firefox is what I'm using and I have no issues.

[south flume]

Chrome is ok,tks,

[native knot]

Skill issue. 😛

[south flume]

So, what is exactly DeepCache? Is that some kind of Lora?

[native knot]

https://github.com/horseee/DeepCache

[south flume]

Tks

[south flume]

I know things aren't always straightforward but assuming there's a (presumably small) quality degradation on Turbo and Lightning models is it a usual workflow to try stuff with those and then send to img2img to "enhance" with base+refiner SDXL?

And BTW: are custom (regular) SDXL based models ment to be used without refiner?

It's almost a full time job trying to keep up with the latest

I've read that on DeepCache but not certain it would work for me (I'm on Macintosh)

Speed is good , but ultimately I want the best quality possible and most of the time I need stuff to be at least 4x4k pixels

[wet nacelle]

https://tenor.com/view/life-is-pain-life-is-pain-i-hate-jerma-jerma985-aesthetic-gif-24010152

[native knot]

I don't know what to think about this one...if that was an actual glass sphere, odds are that the refraction would flip the internal image.

[upbeat summit]

I agree, it's certainly not physically correct heh. maybe it's a portal

[native knot]

Yeah...it's mind-bending. It actually took a moment to register with me.

[upbeat summit]

trying out some things 😉

[native knot]

heh...wild

[native knot]

I've got something interesting, but it's Cascade-related. I'll mention it here conversationally, though.

Essentially, I've found a word combination that's got an overwhelming strength to it. The example is that if I use the word "bumblebee", I can easily tilt cascade one way or the other by adding "transformer"/"car" or "insect". Pretty standard up to this point.

However, add the word "flower" and now you are locked into the insect and you pretty much can't escape that unless you re-add "transformer" or "car" at which point both a car and flowers will appear. But good luck with modifying the bee itself with anything other than generally strong tokens like batman, beans, or the usual list. (Obviously, there's more that will be overpowering sometimes as wel've seen in all models.)

What's more though is that it's so strong in Cascade that this positive:
(police:1.1) [bumblebee:1] on a flower
with this negative:
(bumblebee:2)
still results in an image of a plain bumblebee on a flower. By all other logic, the bumblebee should pretty much be eradicated from the image.

(...and yes, I've also tried (bumblebee:-1) just in case something with brackets was broken, but even then, the negative should've had some impact.)

In essence, the combination of bumblebee and flower is a combination in Cascade that is rather overwhelming and I found that interesting. Typically the tokens that we find are overpowering are single words or they end up being names or styles.

[upbeat summit]

That's a very interesting find! Thanks for sharing it. There are a couple of token combinations in each model that really fight for dominance.

I've done some research with Cascade when it was published, but there are so many new SDXL models I'm testing right now with amazing fidelity and SD3 arriving that I haven't spend much time with it the last few days.

But what I've noticed it, as many others have, that prompt tokens behave quite different with Cascade. Longer prompts can overwhelm it and you get images like you have used crazy high CFG settings. Also distilled models like Lightning based but especially Turbo behave more restricted in versatility and sometimes it's hard to break out of concepts.

I'm noticing this a lot right now since I'm mostly testing Lightning based models and I'm dissecting prompts quite more often to find a token that doesn't let me move in the latent space

there's certainly a technical reason for it (distilling etc), but this is how behaves for me on the user side

[native knot]

Yeah, the long prompt issues were immediately apparent to me within moments of toying with it. It's great with most short prompts, though.

[native knot]

You've probably seen enough of my images and discussions to know how interested I am in TXT2IMG versus some of the other stuff. (Not that I don't use them sometimes, but prompt engineering is so fascinating!)

[upbeat summit]

yeah - absolutely the same for me 🙂

[peak crag]

draw arch count

[cyan crown]

https://civitai.com/models/350433

[pallid path]

rgthree image comparer is so good

it helps me appreciate highres fix even more

[native knot]

Yeah, I like that compare tool as well.

[pallid path]

epic

[gloomy lark]

I see your batman, and raise you 1000 more

[surreal depot]

tell me guys how many gb you have on your video card and how long you are generating SDXL

[gloomy lark]

24giblets

[surreal depot]

and how long you've been generating?

[gloomy lark]

shrug. a year?

sdxl wasn't released all that long ago

[surreal depot]

I have 4gb and it takes 30 minutes to generate

[copper kraken]

Ouch

[glad grove]

yea u need at least 8gb otherwise it moves the model to ram and takes forever

[surreal depot]

Don't mind my English, I'm just learning

[gloomy lark]

this took about 60 seconds on a 4090 with that new lightning lora. it's pretty great.

[shy kelp]

the legend of zelda

[surreal depot]

you're either very rich or you've already sold your kidneys😂

[gloomy lark]

it runs really fast on a 3080 also.

the lightning part makes it all crazy fast.

[glad grove]

20 secs for a 1024x1024 img on forge with 12gb vram on a 3060 and that card is not rly expensive

[surreal depot]

I can't install the 3060 on my laptop

(

[glad grove]

yea laptops suck unless u buy one with a 4090 but those ones are rly expensive

[gloomy lark]

I don't want to know what a 4090 on a laptop would be like for heat and fan noise

[glad grove]

its the downgraded 16gb version

[native knot]

Plus, a lot of the laptop GPUs are not really the same as their desktop counterparts.

[surreal depot]

No the laptops are top notch just bad that they are cut down but I can play all the games on ultra

The biggest disadvantage is the cut-down graphics card(

[native knot]

Play City Skylines 2 on ultra. 😉

[surreal depot]

I generate a regular SD in 60 seconds, but SDXL takes 30 minutes

[glad grove]

because its a 4gb card

sdxl model weights 6gb so not enough room

[surreal depot]

well I'm not interested in this game but I see what you're getting at)

[copper kraken]

Here is the image you requested

[shy kelp]

What about genshin impact

[copper kraken]

Here is the image you requested

[gloomy lark]

ok that's funny. is that pandora?

you need blue men riding those things

and the many faces of his dog.

This next one is hilarious

[gloomy lark]

hah i love the eyebrows

[gloomy lark]

it's the one in the cabinet that gets you.

[dusty tundra]

Working on a project where creatures are morphing into objects.

[manic vector]

anyone have trouble inpainting on sdxl?

im trying to inpaint to fix eyes but it just doesnt work out right

[chilly vigil]

I have just started using this one it might be over kill https://youtu.be/6OO37ZjjbSs?si=raxW3xnZZOMwzjFW its a bit weird because it uses ESAM to segment what you tell it instead of drawing the mask yourself.

[manic vector]

thanks

[tawdry current]

If you're doing inpainting, check out "differential diffusion" with comfyui. It's amazing and works with any model. The more steps you use, the better it blends

[tawdry current]

Here's an example of inpainting using the new differential diffusion method. Make sure you blur your masks (old inpainting was a binary black or white thing, this can handle gradients and you'll see that while it's diffusing)

And like I said, it works with any SDXL model you throw at it. This is just using the base 1.0 model. I wouldn't recommend it with turbo models because you really need more steps to get clean transitions

[hasty orbit]

Here is the image you requested

Here's an example of inpainting using the new differential diffusion method. Make sure you blur your masks (old inpainting was a binary black or white thing, this can handle gradients and you'll see that while it's diffusing)

[torn heath]

/a girl

[neon jolt]

woman of stunning beauty, very exuberant, with long hair with a lot of phosphorescent white shine, big hair like Rapunzel, flowing, FULL BODY FLOATING in a dark CEMETERY full of mist dressed as if a cloud came out of it, white in color, ultra realistic, ultra quality, ultra detailed, "-c", in the background a large moon and white fireflies Pale makeup, illuminated skin, SHINY EYES, HDR, 8k, glitter particles and crystals throughout the image

[neon jolt]

Is it a boot like dream? How do I enter?

[native knot]

You don't. You read:
#1047610792226340935

[final notch]

do not enter the boot

[glass forge]

sdxl has mature dbut still color bleeds bad. if you request any thing in a color (clothing eyes etc) then ull likely end up with a lot of thing in that color.... sigh

[azure oxide]

You learn not to specify colors in prompts and save that for inpainting. It's probably best to not expect a masterpiece artwork on the first generation since AI isn't at that level yet. Even artists today first draw a rough draft first then detail later.

[copper kraken]

diff diff creations

[pallid path]

damn

[hallow summit]

experimenting with workflows and settings for SDXL-Lightning (Dreamshaper XL) + TCD (Trajectory Consistency Distillation) LoRA:

[gloomy lark]

what does tcd do?

[pallid path]

SIGMA

I'll look that node up

https://github.com/dfl/comfyui-tcd-scheduler

[hallow summit]

https://mhh0318.github.io/tcd/

[gloomy lark]

interesting, thanks

[hallow summit]

TCD scheduler node from that repo works, sampler not so much - but samplers compatible with SDXL-Lighgtning (like Euler) seem to be working with TCD LoRA. though for pure SDXL-Lightning it's usually with eta 0, and for TCD you need bigger value

[pallid path]

imma try that out cause I'm using lightning anyway

[gloomy lark]

Anyone do outpainting in comfyui? I'm trying to turn a square image into a 16:9 one. It's doing a good job, but it's got these seam lines as you can see in the second image. Ideas on what to adjust? I've tried changing the feathering number but that didn't get rid of the seam lines.

[heady vale]

remove them in photoshop and do an img2img pass to clean it all up

[gloomy lark]

well, I'm trying to have it be part of my automation.

where it renders at 1024x1024 and then widescreens it all via json and thrown to the comfy api

automatic1111 does this without giving seam lines.

so i know it's possible...

[heady vale]

Ive not had perfect outpainting using any software, so automating it all will be tough

[gloomy lark]

example of a1111 outpainting. no seam lines.

[heady vale]

yeah im not sure what causes the seams. it does suck

openoutpaint for a1111 was really nice

[gloomy lark]

i think it's not denoising the whole image, whereas a1111 is.

ok, that was actually the fix. denoising the whole image afterwards as another step

from comfy

[heady vale]

nice. surprising it didnt create something from the left over noise

[gloomy lark]

I'm just going insane with left/right expansion. 🙂 it's absolutely just making stuff up, but it's still inline with the prompt since it's pulling from those.

[heady vale]

a little repetitious though

almost like it mirrors the sides

[gloomy lark]

yeah, I'll try my lorax gang members on the subway one.

hah, finally an image that fits my superwide monitor

that last step really fixes up any glitches, like hands holding swords not lining up.

[stone fossil]

Myrlin Clownroe.

Kissie kissie. 😉

The round table of faceless people.

[latent citrus]

The phrase coming soon with summer colors and a colorful golden rose in background

[latent citrus]

" coming soon" with summer colors and a colorful golden rose in background

[latent citrus]

The phrase coming soon with summer colors and a colorful golden rose in background

[latent citrus]

https://tenor.com/bWn1z.gif

[native knot]

#1047610792226340935

[rich olive]

Who is this?

[minor pecan]

Dranald Aubrey Trumprake Graham

[finite heath]

You can imagine a quirky, cute and sweet girl with long hair and braids

[glad grove]

i would rather not

[copper kraken]

Here is the image you requested.

[copper kraken]

ready for school kids?

[cobalt musk]

fill the shadow of the image with coffee beans

[gloomy lark]

why u make me do dis?

[copper kraken]

Here is the image you requested.

[cobalt musk]

No. thanks

[copper kraken]

Here is the image you requested.

[copper kraken]

aww cute

bugs are cute

espec spiders and cannibal variants

[gloomy lark]

recent revelation: adorably plump - makes everything awesome

[gloomy lark]

i love dark arts, but playground 2.5 upscaled is pretty crazy

click open in browser

an anthropomorphic centipede-pope hybrid in a business suit giving a high five to a trump as a little boy who is smoking

[copper kraken]

wow that's awesome

[lyric cove]

what's the most photorealistic base model for portraits?

[gloomy lark]

probably juggernaut

on civitai.

[copper kraken]

Check out imperfectportrait too

[gloomy lark]

I just zoomed up that turtle. that's really good

[rich olive]

Thanks im experimenting with the AND prompt syntax, things get pretty weird

[gloomy lark]

with an extension or just in general?

[rich olive]

Just general

[gloomy lark]

So what's an example prompt look like where AND is helping in some way? vs. just batman and putin.

[rich olive]

Something like photo of batman :1 AND photo of Putin :1

[gloomy lark]

what's interesting is that it's just putin under the cowl. it's his lips

[rich olive]

You can get something like this

[rich olive]

Yeah is basically just combining the 2 prompts

[copper kraken]

yeah and it's diff from concatting two embeddings with BREAK

and it's dif than |

[rich olive]

Vladman

[gloomy lark]

prompt: photo of batman-putin-centipede-bus hybrid - it gave me lots of wheels like centipede legs, batman's cowl horns/ears, and it put it in Russia. 🙂

[rich olive]

Yeah i don't like using Break or |

[copper kraken]

where's the obligatory mushroom cloud given the setting? 😛

[gloomy lark]

TRIGGERED

[native knot]

I made a thing that like one or two people would care about. If you're one of those couple of people, thanks for checking it out:
https://docs.google.com/spreadsheets/d/1es_BjyKJDg6LB-2sM0HpF1fLLoV7OYEDCPKWBJmh6nM/edit?usp=sharing

[gloomy lark]

putin batman goes hard.

[gloomy lark]

it's so beautiful. 🙂 I've actually spent way too much time trying to figure that stuff out. thanks.

[native knot]

You're one of the 2 people, then! ❤️

[copper kraken]

putin goes hard

[native knot]

He doesn't look hard in that image.

[gloomy lark]

What does a sharp minded ballerina wear?

[native knot]

😛

[copper kraken]

how could you tell with the ballerina skirt in the way

[gloomy lark]

an astutu.

(cursed astute-tutu hybrid:1.5)

[native knot]

I mean, if he's to be believed, you'd definitely know, even with that on.

[copper kraken]

lol

[gloomy lark]

putin in a ballerina outfit, being lifted by batman in the nutcracker suite.

[gloomy lark]

if you look closely.....

[native knot]

Forbidden burger.

[gloomy lark]

ok now i want dumplings.

[copper kraken]

time to feast

[native knot]

mmmm....rock loaf

[rich olive]

magma dumpling?

[wet nacelle]

@heady vale@gloomy lark

[heady vale]

whats a highlight?

[rich olive]

Ideogram 1.0

[wet nacelle]

Don't really know actually. Found it funny though.

[rich olive]

Ideogram 1.0

[gloomy lark]

yeah I figured out the outpainting. it actually exposed me to the truth about latent space upscaling. no more latent to image to latent to image, losing quality at every step. so much higher quality images now, and far faster.

[copper kraken]

What are you referring to

[rich olive]

Ideogram 1.0

[copper kraken]

That's a static image up above (the highlight) not. A link

[gloomy lark]

this is my new workflow. I'm sure it's old news to people such as yourself, but I was using ultimate sd upscaler for everything. for resolutions ~3k and below with the 24 gigs of vram I have, I can do much faster upscales, maintaining much higher quality by skipping all that.

and I can use ultra fast dpmpp 2m at 20 steps sampling and still get crazy high quality

[copper kraken]

Oh letwe get you my workflow

[gloomy lark]

the denoising takes care of all the stuff I used to spend tons of time with dpmpp_sde on, just so the fingers and hands would be right. this solves all that.

I can literally watch it fix hands and arms in realtim

with the fastest cheapest sampler there is.

[copper kraken]

There's a few really good ones I can get you in a min here

[gloomy lark]

and it does a 1920x1080 res upscaled image in 7 seconds.

[copper kraken]

That are actually simple

Not your usual clownshark monstrosity

Ultimate kinda sucks

[rich olive]

Dalle-3

[rich olive]

Dalle-3

[copper kraken]

there's a few good methods

regarding the straight up latent upscale you showed, try getting NNLatentUpscale and swapping in that node, i've found it's usually better than the base one

[gloomy lark]

yeah I tried the methods they have there, the "area" one is majorly better than the others.

[rich olive]

Dalle-3

[copper kraken]

here's the simplest one that i liked the most when giving it just a bit of freedom to tweak or add details

if you're using regional prompting, you can't use that node, but you CAN use the "tile diffusion" node which patches the model

[rich olive]

Ideogram 1.0

[copper kraken]

then feed that into regular ksampler etc with your regular regional prompts but upscale em first

[gloomy lark]

so this is going form latent to image, back to latent, then back to image, similar to the old method I was doing. When I looked closely, I was losing quality of the image each time I did that.

[copper kraken]

switch the tiledksampler scheduler to exponential to get less change

this is better than latent upscale

you lose more detail from the latent upscale than from the conversions

but this was the best combo i could find including with the upscale model

the main advantage is you don't need as high of denoise levels to get satisfactory detail

tiledksampler takes a long time but i'm the happiest with those results, at least with the weird shit i tend to make

[gloomy lark]

well, the high denoise I found was to fix the lack of correct limb placement by the quick simple dpmpp 2m sampler which is prone to that stuff.

[copper kraken]

the other reason i like tiledksampler is cuz you're sticking to the standard resolutions

you're not inducing mutations, you're usually fixing them if anything, unlike a straight up latent upscale

this is with NNLatentUpscale instead of the latent2img->upscale-img2latent

it might be better for your purposes

i tend to kinda like the weird details that sometimes pop up with more aggressive settings unless they flat out change the fundamental composition

[gloomy lark]

I keep going back and forth between speed and highest quality every nook and cranny looking amazing. I ran that one, which took 75 seconds on mine, vs. 7 seconds. 🙂 I find I can riff faster and respond to cool ideas faster when it's 7 vs. minutes that I'm used to waiting.

[copper kraken]

here's another method, iterative upscale

yeah tiled will always be slower

[gloomy lark]

this is pretty good for 7 seconds.

I played around a lot with turbo and lightning models in the last day, I thought the lightning stuff was amazing, until I started comparing it head to head against non-lightning. It's pretty much trash. amazing loss of detail and prompt adherence. lorax gang members with tatoos, bandanas, switchblades, etc etc. 75% of that was gone on the lightning version compared to regular.

[copper kraken]

yup

i have no interest in any of those models really

[gloomy lark]

So I bit the bullet and ordered more 4090 hardware instead. 🙂

[copper kraken]

if there's a loss in quality, unless that loss can act like a preview and you can just swap the settings back to get the same thing with details restored (never happens really...)... then it kinda sux

ha did you order a second one?

[gloomy lark]

why stop at 2?

[copper kraken]

here's a faster method

hahahah do you have three??

this is straight up unsample resample

not faster than the latent upscale

but does give diff results

[gloomy lark]

yeah I figure with SD3 coming shortly, and with all my scrips and stability swarm being able to use all of them at once, I'll just have the 3 4090's and the 3080 running in parallel for maximum silly pictures.

[copper kraken]

and there's opportunity to do some really interesting stuff by changing the prompts on the unsampler alone

love it haha you're insane

[rich olive]

I found it!

[gloomy lark]

it's a set of 6 upscaled high quality images against 2 models every 40-50 seconds now with a 4090 and 3080. so adding 2 more 4090's will speed that up by a lot. so I can throw prompts at it and get stuff back quickly to have fun just throwing ideas a it.

I meant to ask, were you around here when the stability ai bots were working here? how fast were they on requests?

[copper kraken]

nope, i wasn't

i first hopped on here shortly after cascade launched because i got annoyed at how crap the info on reddit was

[gloomy lark]

With the new SD3 bots coming along, I'm wondering if there's benefit to extending what I have going to other people or is the bot stuff here going to cover any need.

[copper kraken]

2/18 that was when i joined this discord

[gloomy lark]

yeah I was shortly after

[copper kraken]

i'd imagine it would cover any need of a certain type, but not the need the Sharkb0t is currently covering

[gloomy lark]

hah what's the sharkbot/

?

[copper kraken]

me lol

spewing fucked up images

[gloomy lark]

ah ok

well, when I get the new hardware and can figure out proper queueing of multiple rapid fire requests, where the next request doesn't skip the line of the first one's images, I'll let you know.

probably a couple weeks.

[copper kraken]

cool

i finally got around to setting up tailscale so now i can use comfy wherever i want in town

[gloomy lark]

neat. I'm finding that the discord bot method works best when I'm out. the ability to submit a prompt and then close the phone and check it later is great.

obviously full comfy is best if you're trying to adjust stuff.

but for just having an idea and want to see what that would look like, discord bot is great.

[copper kraken]

for sure

[gloomy lark]

saw this on reddit the other day, so i made a couple

reminds me of the cruise ship in fifth elemtn

[copper kraken]

big badda boom

[gloomy lark]

exactly

[copper kraken]

^_^

another thing you can try too...

the second is with noise injection, this is a straight up latent upscale with NNlatentupscale

it brings a bit of those details back that you lose from latent upscales

the thing that sucks with them is a lot of more complex textures are lost

assets disappear or are simplified and textures get smoother

for comparison, this is the same NNlatentupscale except with tiledksampler

check out the grass

[gloomy lark]

hah the person in the background is normal vs. dead body in the first 2

[copper kraken]

lol

yep, the downside of tiledksampler... dead bodies come back to life

[gloomy lark]

leeloo dallas multi-pass, falling past futuristic flying cars, multi-level, floating city, futuristic fashionable people.

[copper kraken]

and here's one last one for when it's imperative to preserve the source composition

[gloomy lark]

yeah, like that high denoise I'm doing. I accept that it's going to give some weird things happening every now and then, but it's the tradeoff of returning lots of images quickly.

hah yeah exactly. torso crotch

but sometimes that's awesome

[copper kraken]

yep

[heady vale]

and fish lamps 😆

[copper kraken]

bmut yeah absolute best i've found for preserving the original image has been tiledksampler or tiled diffusion for regional prompting needs

if you want high levels of detail, dpmpp_ancestral with 30+ steps

source composition is generally fine at 0.5 denoise, 0.55-0.6 if you want shit added, beyond that it'll go bananas

[gloomy lark]

2s ancestral at 35 or 70 steps is my favorite. best looking out of all the samplers. second only to dpmpp sde's accuracy.

[copper kraken]

yeah for real

i love that sampler

res momentized is getting a real soft spot in my heart too

that is a very very good sampler

[gloomy lark]

yeah i was doing 0.7 etc for big side outpainting. 1:1 to 32:9

[copper kraken]

what are you outpainting with

[gloomy lark]

I was doing the latest padding for outpainting. and then denoising after. that's actually how i stumbled upon the whole denoising of an image afterwards, and that led to upscaling the latent and denoising that.

then i also did dark art images render, with latent upscale but with juggernaut9rundiffusion.

that led to this one.

[copper kraken]

got a workflow? curious what you're doing exactly there

[copper kraken]

i used that some for a while, then ditched it cuz i got tired of patching subtle seams

[gloomy lark]

so dark art, did a good one of this, but doing a juggernaut denoise massively increased the realism of his face and skin

the denoise afterwards gets rid of all seams

[copper kraken]

it worked really really well in sd15 but i've had issues with the inpainting models in sdxl

[gloomy lark]

that was my complaint.

[copper kraken]

ahh so you're denoising the entire image afterwards

[gloomy lark]

i said to the guy, a1111 has no seams! and then I thought, it's because they're denoising the whole image, not just the masked parts on the sides. voila! denoise the whole thing, no seams. looks great.

[copper kraken]

ha, yeah

that's another place where tiledksampler fn rules

diff diff could probably solve that issue too

what i've noticed too though is you still end up with issues with "soft" seams

where changes to the structure of the image have been introduced

sigh, so many things in sdxl are just so much crappier than sd15 and inpainting models are def one of them

[gloomy lark]

ok here it is

makes weird compositions, but with realistic textures.

those weird dark arts compositions, that juggernaut would never do natively

[copper kraken]

yeah for sure

i need to get a workflow set up to make carosello img2img realistic outputs

that model is fn incredible for making the weirdest of the weird

and yeah regarding the upscale def give NNLatentUpscale a shot, maybe it'll suck in your case but i've gotten better results

[gloomy lark]

so this made the dumplings look more photo style

it's just an interesting option.

I ended up just using this method without the separate checkpoint for the 7 second upscaling.

[copper kraken]

what i'd love to do down the road is develop a workflow that automates outpainting one chunk at a time in a ring around an image until it's doubled in size

then downscale it 50% and iterate a few times

then upscale and tiled resample for a while

[gloomy lark]

it's interesting that you bring that up.

[copper kraken]

could do some really wild looking shit

[gloomy lark]

So when we render at 1024x576 for example, what's in the image will be significantly different than if we do at 1536xwhatever. it's not just more texture details, but what makes up the image, the complexity of how many subjects etc is very different. I'm noticing with services like ideogram, they're getting very high complexity of subjects and interactions, all at very low res with their default 1280x720. Same with midjourney. very complex details, even when it's just 1024x1024. stable diffusion XL, doesn't do that. resolution directly dictates what's going on in the image.

I'm really curious to see how SD3 handles this. if 1024x1024 will be more complex now.

[copper kraken]

yeah i'm wondering if they're hiding shit

what i'd be tempted to do is upscale, get the extra details, then downscale to hide the artifacts

iterative upscale is great for adding

[gloomy lark]

exactly. i've spent countless hours on the iterative upscale.

just to get that kind of complex world building.

[copper kraken]

ahh cool you've used that one then

when you see my crazy monster workflows, tbh they're not as crazy as they really look

usually i'm doing shit like reproducing an iterative node, or something conceptually similar that doesn't exist yet

so that i can actually play with all the parameters internally

[gloomy lark]

so I just plopped that nnlatentupscale in there.

is it supposed to make things better?

or do i need to use that with the tiled ksampler? just with that one node swapped in, the image is a lot software now.

softer

[copper kraken]

Huh interesting

Maybe the sdxl version isn't as good now that I think about it

[gloomy lark]

switched back to the original upscale latent by and it's much sharper again

[copper kraken]

I just realized Ive been doing the ldsr plus workflow since I started using SDXL full time

So maybe it was only better with sd15

[gloomy lark]

that's the partner of dpmpp_sde. slowest. upscaler. ever.

[copper kraken]

Err not that uhhh

Whatever it was I had embedded in the Json

[gloomy lark]

ok, the output of tiled ksampler, is different than just ksampler advanced with denoise that i was using

[tawdry current]

That scheduler does two passes per step, that's why

[gloomy lark]

this one is the tiled ksampler

this is the original ksampler with advanced denoise I was using.

As you can see, on the crucifix on the second one, is there's little jesus trump with trump hair. the clear winner. 🙂

the details are very close, but the trump on the cross, and also if you look at her hair bun, it's a little better on the non-tiled ksampler.

but most important, ksampler with denoise takes 7 seconds, compared to 20 for tiled.

[copper kraken]

Yeah I'm sure there's some cases where you might get something more preferred that way

[gloomy lark]

all that said, the "area" method is really the star of the show. all the others are blurry like that nnlatent upscale node.

[rich olive]

Sapphire dumpling?

[copper kraken]

The stuff I've made tiled with that upscale model (whatever it's called, away from comfy now) has been the best pretty consistently

Nature stuff in particular

[gloomy lark]

ok it does NOT want to give me sapphire pork dumplings.

[copper kraken]

Ha

[gloomy lark]

all that said, tiled is how to get real big. this latent upscale thing only works up to around the 2k-2.5k resolution, then it starts blowing up the 4090.

but for 1k res upscale to 1.7 or 1920 res? it's perfect.

[copper kraken]

Yeah

What I've found is the farther you get from native resolutions the worst latent upscale gets

That's obv going to depend a lot on the checkpoint too

[gloomy lark]

definitely. If I set this to 2x resolution it starts putting pixelated squares everywhere. 1.5x seems to be the sweet spot

[copper kraken]

Yeah

2x the big thing I see is mutations

[trim orbit]

turned on fp8 yet?

[copper kraken]

Freak limbs all that and a loss of assets

[trim orbit]

i breach 2k with 16gb 4080

[copper kraken]

A beach with pebbles turns into just sand, etc

Tiled won't do that

[trim orbit]

kohya hires fix is my favorite tool for raw gens at high res

[rich olive]

Lava dumpling?

[gloomy lark]

love me some lava dumplings.

[trim orbit]

When Po becomes a fire bender? or lava bending woudl be earth bending wouldn't it?

dumplings are for sure po he invented them

[gloomy lark]

times like these when it puts out stuff like this, i think i need to go back to the censored model

[solemn wigeon]

/me beautiful girl

times like these when it puts out stuff like this, i think i need to go back to the censored model

[gloomy lark]

hah good stuff. can't wait until these things can hold swords correctly

[copper kraken]

Here is the image you requested.

[solemn wigeon]

/me What a sweet weird cute elf girl

[copper kraken]

Here is the image you requested.

[rich olive]

one of my SD1.5 finetunes

[gloomy lark]

these anime models are hilarious. it doesn't matter if you say they're wearing power armor, they're no more clothed than without.

[gloomy lark]

why are these models even trained on this stuff

[rich olive]

SDXL 0.9

[copper kraken]

Who said the part umad responded to

[gloomy lark]

i don't see that reply in his twitter posts or replies, so i feel like it can't be recent

[copper kraken]

Gotcha

[gloomy lark]

yeah, it's gonna be pipelines i guess. something that takes "man running" and looks through it's library of controlnet openpose files for that, then uses that on what the model IS trained on to give you a man running.

they say that's what dall-e does, but who knows. it's all black box

I doubt ideogram does that.

[rich olive]

Have you heard of ELLA, basically it's equipping any SD based diffusion model with LLMs for enhanced sematic alignment, there's a whole paper on it, i think Ideogram might be using something similar, maybe SD3 too

[gloomy lark]

yeah they said they'd release it in a week, it's been a week. 🙂 maybe this coming week.

having that kind of prompt adherence with the maturity of a lot of the existing models would be amazing.

with sd3 we're starting over again, i'm hopeful, but it's gonna be a while. kind of like with cascade, I personally feel like I've hit my limit with it until finetunes come out, which'll probably be never.

[rich olive]

Yeah cascade isn't as big as i taught it would be, and the fine-tunes are worst than SDXL for some reason

[chilly vigil]

This actually makes it look really good. Can you please share that workflow?

[tawdry current]

Just look at the second image, it's like two extra nodes from a standard workflow. The only node you'd need would be a mask blur node

Comfyui essentials has the mask preview and blur node. The rest are built in

The big key with differential diffusion is that you HAVE to have gradients in your mask, otherwise, it will just be standard inpainting. So the blur will make those gradients for you or you can do masks in photoshop

[burnt shadow]

what? why? 😄

[tropic wyvern]

#Aishwarya Rai Bachchan transformed into a male character, wearing modern male dress, standing in a futuristic cityscape with neon lights and flying cars in the background, under a purple-hued evening sky, exuding a sense of confidence and mystery.

[rich olive]

I made that for you, "modern male dress", not sure what you mean by that one

[rich olive]

maybe this one

[pallid path]

messed around with the level tool in gimp to make the sky overexposed on purpose

[pallid path]

Tried out SuperPrompt-v1
very OP even with SDXL it seems?

WTF IT ACCIDENTALLY COOKED SO HARD

[rich olive]

Lol

[pallid path]

okay I'm actually going to use clip L and Clip G separately

to see if Clip G can extract that natural language intelligence

I also read somewhere that CLIP G is better at interpreting natural language, while CLIP L is better at interpreting tokenized prompts (our usual comma-separated prompts). Maybe that's the reason?

okay??? nice??

im plugging in the superprompt prompt into Clip G since its more like natural language

and all this with SDXL Lightning (DreamShaperXL)

man it would be great to have that 512 context length, I keep cutting from the prompt to fit in that measly 77 token limit

[pallid path]

what could they be looking at?

[tawdry current]

I'm using llava 1.6 Mistral to describe images in large detail and then feeding that into a Hermes2 model that's good at following instructions (The llava model sucks at following strict directions for some reason). It takes the wall of descriptive text and consolidates it down into really decent prompts. I wrote a few nodes for comfy to interface with ollama server to do it. The hard part is writing the right system prompt for the second stage.

[tawdry current]

Here's an example of my autocaptioner. I just point it to a folder of images, set the seed to 0 and increment, then queue up how many ever images are in the folder. It really does spit out some half decent captions for doing loras and you can put in your own custom tag. On a side note, I hate coding in python...

[pallid path]

dayum

[pallid path]

@graceful nexus Sorry for pinging you, but your SuperPrompt model has been cooking for me in SDXL, I hope to see the comfyui node one day. With SD3 it's going to be even more useful. I also tried plugging the superprompt prompt into clip_g and it gave some nice results as well. Thank you for your work.

[copper kraken]

btw i just checked what that upscale model was

LSDIRplus

about to try the new proteus rundiffusion model

[gloomy lark]

Man, I love proteus. 0.3 especially 0.4 grew on me, but he keeps going down this weird road and now he's full embraced this specialized booru style Pony tagging method. Hard to get on board of proprietary stuff.

[copper kraken]

Oh damn is that what's going on with the new one?

[gloomy lark]

Stability staff have commented on his tweets about the new model, referring to it as a virus, that him and some others are moving us backwards from natural language and back to a sd 1.5 style of comma delimited keywords.

[copper kraken]

Yeah, I saw that

[gloomy lark]

Boy holding a cup is sdxl, boy, cup, holding, doesn't establish the relationship and only works well for single subject portraits.

[copper kraken]

The problem with sd15 style is you're then leaving it almost entirely up to the attention layer to decide what your composition is imo

Exactly

It has to guess

[gloomy lark]

Yep

I've had the LLM give me that style back, and for single subject pics, it actually does better than full language. But if anything more complex, it falls apart.

Proteus is really goood at prompt adherence, it was king before dark arts and deep blue came out. But now he's going backwards.

For the sake of prettier pictures.

[copper kraken]

Yeah

Well that sucks lol

Honestly the models already make pretty enough pictures

IDC about the aesthetic score shit at this point

It's all about getting the right composition now imo

It does seem like the actions issue will remain present like you've said

But cascade has shown us that they should at least have resolved some issues with other image types, like dark scenes in particular

Drove me fn crazy trying to get sdxl to generate an image of a truly dark scene, I don't think I've ever really succeeded, there's always been a significant light source injected or implied somehow

[gloomy lark]

So talk about prompt adherence, I said unearthly beast emerges from portal at end of prompt. Kept getting this.

Only deep blue followed the more complicated prompt

Lemme try the dark thing.

[pallid path]

these are epic

[gloomy lark]

So what couldn't you do with sdxl concerning darkness?

[copper kraken]

Get darker than that

[gloomy lark]

I said - completely dark room. Small anthropomorphic bear shining flashlight at his nervous face. Sense of extreme darkness.

So what would be an example of language that's darker

[copper kraken]

Not the darkest I've gotten with cascade, just what's handy ATM

But I've gotten it pretty easily to generate images with no strong diffuse light source off screen, no candle, almost pitch black

Not sure off the top of my head

But basically what the interior of your house looks like during a power outage at 1am where you can only see after your eyes have adjusted

[gloomy lark]

the last atom of life before heat death of the universe.

Hah still rather bright

[copper kraken]

Lol

[pallid path]

https://media.discordapp.net/attachments/1089974139927920741/1161431788317581312/image.png?ex=66034da4&is=65f0d8a4&hm=88ecad9d2609650015396607759811920619d8f99afc9569ecd89ef086fed150&=&format=webp&quality=lossless&width=1183&height=676

hehe old sdxl image

[gloomy lark]

I feel like plush animal is a wholly unexplored prompt tree for me, soon to be remedied.

a tiny lit candle in the middle of the wooden floor illuminating a legion of horrifying creatures in the room.

That is one horrifying creature.

[copper kraken]

that's what cascade can do

that afaik sdxl absolutely cannot

[copper kraken]

finally got around to writing darkarts a much earned review

"This is a truly incredible checkpoint. I have 240+ on my SSD and this is my #1 go to for almost anything (the other being Carosello). Its prompt understanding is outstanding, and the qualities of the outputs are diverse, with a lot of unique styles (with as many checkpoints as I have, it's obvious how many are inbred). Absolutely fantastic work and I look forward very much to any future versions or checkpoints trained or merged by the author."

just checked, it's the first model released by that author

[gloomy lark]

Yeah I went looking for his realism based on. I have to see if it's better than juggernaut

[copper kraken]

wait there's another somewhere by that author? i must've missed it

[pallid path]

ouch

[gloomy lark]

Yeah there's a whole bunch. Dark arts was a one off to handle this kind of content

[copper kraken]

spoiler that or delete it, that's an epilepsy trigger

[pallid path]

dark arts looks like if Illuminati Diffusion was for SDXL

[copper kraken]

⚠️ @smoky patrol @uncut steeple

[pallid path]

holy shit

this looks clean

[copper kraken]

@slim wren @nova aspen @vestal breach

[pallid path]

22 🤔

[copper kraken]

lol, i like how i got timed out in less than 90 seconds when i posted a pic of a batwinged vampire where the shape of her nipples was visible through her black dress... but there's a racist meme that's been sitting in general with images for nearly 18 hours now, and we've got epilepsy triggers flashing away unspoilered >_>

[stoic turtle]

where do i go to generate images in this discord?