📝｜prompting-help

[atomic wigeon]

And those are defintley chibi

Chibi will force a smush on a scene too. Might be worth trying it.

People think chibi is jsut for characters but I use it all the time in advertsing stuff. Like the Ford Mustang.

That bottom monkey may even be funko, funko pop or funko chibi

If you link it I will try to prompt the style. That's all I can do. You've linked like 15 styles in your few images.

storybook style vector illustration chibi thin blended outline hip hop

Where did you put the subject?

Village?

What was it? Village, mushroom viallge?

Ok lemme try that

That with neg prompt?

[orchid ore]

Any you like?

yes cfg higher is very important

[atomic wigeon]

This is with no neg prompt and 150 steps.

If you're in the ballpark now would be a good time to add an artist like Angela Barrett

She makes this stuff and will blur the lines a bit

Still plucking away at it. No good results so far.

2d cartoon rich color fantastical storybook style vector illustration chibi thin blended hip hop medieval village, Angela Barrett style,

It's ok. Rendering videos on other computers. I dont mind helping. Still trying stuff.

[orchid ore]

If you like mushroom you can try Peyo he made smurfs, so he did mushroom village, just dont know if supported in model

[atomic wigeon]

I seem to have gotten rid of the borders but now it's resembling a watercolor painting.

[atomic wigeon]

Thanks, but he's not after mushrooms. He's after a very particular cartoon style.

[orchid ore]

i like those. Was confused reading what was written before.

[atomic wigeon]

Try this:

illustration, organic color, fantastical, grogu style, funko pop, chibi, 2d,

My results with it at 40 steps looked good. But 150 steps and it's way off. Sorry, I tried. lol Gotta do other stuff now. Good luck.

[orchid ore]

@vast jetty can you post image that is closeths to your idea?

[atomic wigeon]

That prompt in 3d. Kinda fun lol.

That prompt in 3d + macrophotography. LOL ok now I gotta go do stuff.

[severe vale]

what is the correct way to have multiple subjects in a prompt?

[orchid ore]

Are you using A1111? @severe vale
If so best way is regional prompt extension. You can divide image on several quadrants and tell what will be where.
If not it is trial and error, and patience 🙂 Group of smb, pack of smb.
or Bride with flowers. Older groom. Best man.

@vast jetty how about those?

i like the last one

it is from comfUI, i can send you pos prompt

Model of simple dreamy village with mushroom. Color. Stroke outline. Cartoon

negative there is nothing special, only watermarks and so

np! hope it will work. Btw cfg 10

good 🙂
if you installed comfyUI you can get all data out of those images.

oh i forgot.... Hope it will work for you. It is SDXL....

[severe vale]

So there is no special fonting or prompt types that i need to achieve this? i have always had good results just winging it but i want to make sure im learning correctly as well. Im just using the disco bot and a few other platforms like playground ai.

[orchid ore]

If you are satisfied with results, it is o.k.
I for example when i want bride and groom.
Middle-aged bride. Very old groom with watches.
I divided subjects with dots. But all can sometimes fail.
For sure there are experts that can help you probably more.

[placid mica]

Trying to get a view from outside, looking into a small living room... For the life of me I only get a view from inside the living room looking out the window... any ideas?

[orchid ore]

partialy done but they have closed blinds 😄

[placid mica]

it's rediculous how difficult some things are 😄

[orchid ore]

Not best but from outside to inside 🙂

when something is difficult, i add oil painting. But this isnt this case, just photorealistic

[placid mica]

that's exactly what I'm trying to do!!! Why can't I!?!?!? 😄

[orchid ore]

Many tries and some golden sample.
Photorealistic. Looking from street exterior in small window to living-room with opened blinds

[placid mica]

thx, gonna try! what model was that using?

[orchid ore]

sdxl crystal clear

[placid mica]

still just getting interior.... I'll try a bigger batch and see if it's just rng

[orchid ore]

cfg 4 but it shouldnt matter.

[placid mica]

guess my stable dif really don't want to go outside....

[orchid ore]

switched cfg to 8 and from 4 i think 2 are o.k.

[placid mica]

pumped "exterior" up to 1.4 and now I'm starting to get some views from the right side...

no idea what that cat is doing there....

[orchid ore]

it is probably attribute of opened window 🙂

[quiet zodiac]

@vast jetty It's a nice style to be able to leverage. I played around with the prompt suggestions and got close.
old medieval village with rustic mushroom houses nestled in a green natural fantasy woods , cartoon (vector:1.1) illustration style, vivid organic color,chibi, close up

[orchid ore]

Simple dreamy village with mushroom. Color. Stroke outline. Cartoon
Remove word "model", i was lucky not getting proper models. And now nothing but models on desk.
Good night all!!!

[stuck shuttle]

@hollow tapir VAE? Part of me wants to just have someone help to to point where I can create the ref characters I want for stuff and I can stop paying artists on drawing servers. Then reach my brother how to do this since he also wants to learn this. Even if that means paying though I don't really want to I've been lazy about this because its a lot to learn so I'm just sitting down the next 2 days and staying on discord till I've got it done if I can.

[hollow tapir]

VAE is basically the part of the models that transform "latent space" values from/into actual "pixel values". Baby talk, it's what turn math statistical values into picture.

In the real world using a different VAE from the defauly one can alter slightly alter the output picture you get. (it will mainly change contrast/colors/minor details)

VAE quick guide :
1/ What about their weird extensions? if you want you can add .vae. in front of them. It will dictate where you have to put them.
2/ Where do I put my VAE ?

• VAE with .vae.pt / .vae.ckpt / .vae.safetensors extensions go into the models\Stable-diffusion folder
• VAE with .pt / .ckpt / .safetensors go into models\VAE
  3/ How do I use my VAE ? Three possibilities :
• Either you name it similar to another one of your model (eg : Anything-V3.0.safetensors + Anything-V3.0.vae.pt), by doing that it should automatically load the VAE when you load the associated model.
• You manually load your VAE by going to Settings -> Stable-Diffusion -> sd_vae
• You treat yourself and add a VAE dropdown at the top of your page to quickly switch back and forth between all your VAE by adding sd_vae to your Settings -> User Interface -> Quicksettings list

Here's some comparison shot

So that might help with your "lack of colors" problem

Now. What picture are you trying to recreate and what results are you getting ?

@stuck shuttle

[stuck shuttle]

Sure one sec, I'm not signed into discord on my desktop so mind if I just use my camera to take a pic of my screen instead?

I wanted to make this, fairy tail girl pack lora but I got this instead after trying.

[hollow tapir]

Please take proper screenshot if possible. Post the civitai link of the picture you're trying to recreate. And show your whole auto1111 webpage when you're trying to recreate it.

[stuck shuttle]

I'll send myself the images then, I can't sign in on my desktop.

[brazen willow]

read the documentation and sample image prompts on the LoRA because you might need to exhaustively detail all the missing stuff like tattoos and hair bows

the style difference is probably the model you're using

a character LoRA won't duplicate the art style of a show, and that's a good thing

[stuck shuttle]

I'm using the same model and version as they said was used for the other image.

https://civitai.com/images/1106622?modelVersionId=93658&prioritizedUserIds=16870&period=AllTime&sort=Most+Reactions&limit=20

[hollow tapir]

@brazen willow this is a followup from a conversation in #💬｜general-chat message

I've already listed the usual suspects

(thanks tho)

[stuck shuttle]

What exactly do I do to at least get the right art style?

[brazen willow]

the screenshot says it failed to find the LoRA in the lower right hand side

[hollow tapir]

You're not using the same settings:

• you have different initial resolution
• it's using the old AddNet (from "Additional Network" extension) methd for lora
• you probably didn't install the lora correctly because it couldn't find it

the most important part in all of this is the last.

you're not using the lora at all

[stuck shuttle]

I was told download it and then put it in the lora folder. It shows up with the other lora when I hit the red button under generate and click on the lora tab.

Do I not just put the downloaded lora file in the lora folder?

[hollow tapir]

what is added to your prompt when you click on that lora's card ?

[stuck shuttle]

lora:FairyTail_Girlpack_v3:1

[hollow tapir]

then use lora:FairyTail_Girlpack_v3:0.4 instead of lora:FairyTail_Girlpack_v3-00004:0.4

(you'll still be missing VAE, using incorrect res, etc)

[stuck shuttle]

That's what was added when I copied and passed the generation data from the image page.

[hollow tapir]

I know

maybe the name changed since then. or they use filename instead of alias

whatever it is, this is not the correct way to activate this lora for you as demonstrated by simply clicking on it.

[stuck shuttle]

No the res is what I got from the generation data.

[hollow tapir]

the generation data mentions 480x720 and you are using

And it is using AddNet/"Additional Networks"

[stuck shuttle]

I don't know where you got that as the generation data has with and height as 480x720 and that's what I've been using.

That's where the bar sliders are for the image generation.

[hollow tapir]

oh that might be that old pesky bugs where values don't get refreshed correcty sometimes....

I got that from your very own screenshot

right next to "Hires. fix"

Anyways, fix your lora situation, use a proper VAE and it should be much better.

[stuck shuttle]

Now she looks closer to the image I'm supposed to have but its off in just the right ways to make her look creeper.

[stuck shuttle]

Where do I put addnet what?

[stuck shuttle]

How do I fix VAE?

[hollow tapir]

I've told you everything about VAE earlier #📝｜prompting-help message

[hollow tapir]

I have to do some stuff before answering to that. For now ignore that part, the default lora loader should be fine.

[stuck shuttle]

I'm confused 2/ is talking about having a VAE file and I don't nor do I know where I'd get one for this.

[hollow tapir]

How so ?

[stuck shuttle]

Do you need a screenshot or can I just use my camera and not send myself the screenshot this time?

[hollow tapir]

You can get some VAE from huggingface usually.

eg : https://huggingface.co/NoCrypt/blessed_vae

[hollow tapir]

sure ... camera should be fine I guess.

[stuck shuttle]

And what VAE do I use for this image?

[hollow tapir]

whichever you prefer.

cf the comparison shot, google the name of the VAE and voila

[stuck shuttle]

I haven't messed with VAE yet though but...

[hollow tapir]

try increasing the lora's weight from 0.4 to 0.6

[stuck shuttle]

But like idk what's good for matching this image?

[hollow tapir]

pretty much anything but the default one which will give you grey-ish results

[stuck shuttle]

Oh that made it worse, that made it worse.

[hollow tapir]

Also, when in doubt RTFM https://civitai.com/models/20690?modelVersionId=93658

so try with that model, vae and settings (and maybe try installing the additional networks extension (if it even still works... I can't remember right now)

[stuck shuttle]

Idk, instead of learning to match the image exactly right I should spend my time learning how to add the right prompts for the real ref pics I want. Do you think its worth going into editing the fine details to match exactly at this point?

[hollow tapir]

it's your call to make, but you're pretty far from the original image.

[stuck shuttle]

This is so confusing I was told its relatively not hard to get the images I want or to make characters for a ref pic but I'd argued no its very very confusing and hard, I can't even match one image let alone figure out how I'd create a character and edit a bunch of little things on her.

[brazen willow]

it's very hard to get consistent characters

[hollow tapir]

It's very easy to get a picture but it can be very difficult to get the image you want.

stable-diffusion is a very powerful tool but it's not magic and do not read the mind (yet)

[silver valley]

its not that hard to get a specific character if you have the lora for it.

[brazen willow]

they're trying to make a new character from scratch

[silver valley]

ah okay

thats not as easy xD

[brazen willow]

there are some tools like "character turner" that can generate a bunch of self-consistent turnarounds in a single image. maybe you can repeatedly inpaint those to get enough data to train your own LoRA, but i've never actually seen it done

i tried using a single reference image with reference controlnet, and it was a disaster

this is the tool that might work, if we're lucky: https://civitai.com/models/3036

my idea is that you would chop these into separate images to feed into a custom LoRA

[stuck shuttle]

Then why is it so dam hard to match the character of an image I'm not even close to what I'm trying to match. I can maybe get the character I want but how I edit 15+ things like I'd do after the first artist rendition of the character they are drawing I still have zero clue. Plus even for the lora I have been messing with all I can get is a yeah that's her but that's just not right for the cannon character not even the image I'm matching.

[stuck shuttle]

Like yes and no, like I can just create a centaur for a lora and then edit like 6 things and call it a day. As long as I have a base model and can edit like hair and other small things about the character I'm fine, though I get the sense creating them is not as easy as I was lead to believe?

[hollow tapir]

Sorry but I can't help much more for tonight, I broke my sdwebui install trying some new things out ^^"'. I might take a look at recreating your civitai shot tomorrow if I've had time to fix my install.

[stuck shuttle]

Thank you. I go all out with my RPs and sometimes I make new ref characters the reason I got into A.I. was so I didn't have to keep paying people to draw my characters for me.

I'm now wondering if that's the case still?

[brazen willow]

a human can learn from one or two examples, but AI is very dumb and needs a dozen at least. also from what i've seen it has a hard time with imbuing characters with a personality. maybe a compromise is to have an artist draw lots of quick lineart sketches and use controlnet to turn those into finished images

[granite crescent]

how do you describe a prompt of a person's head pointed 45 degrees from the camera. the halfway point between a side profile and looking directly at the camera?

is there a term for this pose?

[brazen willow]

quarter profile?

[granite crescent]

thank you!

[brazen willow]

also called 3/4 view or three quarter

[stuck shuttle]

Idk I was just told this could be used for my ref pics and I'm getting the sense that no way it can at all?

[brazen willow]

sure is easier than drawing by hand

whoops

i accidentally left the model set to rundiffusion-photorealistic, anyway this seems to do exactly what you want about 50% of the time, play with CFG scale to vary the results once you find a good seed: ((solid outline)), solid lines, flat vector cartoon illustration, full frame, cute walking kitten leopard, simple, neutral background - negative prompt: paper cutout, detail, multiple tails, standing upright, shading, shadow, colorful background, watercolor, circle, text, faded, watermark, blurry, complex, photorealistic, rendering, jpeg artifact

also it might help to put the style stuff at the start of your prompt

stuff at the front has more weight

rules? what rules?

i was just trying out prompts and didn't pick a special model, just left it on the default. when i tried various anime models i got much worse results

rundiffusion is a model (and also cloud service) https://civitai.com/models/82972?modelVersionId=88158

training data quality matters i guess, who would have thought

that's a SD 1.5 model. i've never used SDXL. i've been waiting for things to settle down with SDXL while i learn SD 1.5

there has been a ton of work put into SD 1.5 models and getting them to make good output

it will be a while before we get the same level of quality from SDXL

[next flint]

although i do agree many sdxl models arent at the same quality of some of the highest rated 1.5 models i think with good prompts and negatives images from sdxl models such as dreamshaperxl and rundiffusionxl can compete with 1.5 models.

like although at the moment realistic vision 5.1 is probably the most realistic model i think stylistically for photos sdxl can really compete in some aspects

[tired vigil]

I have 1.4 is it like worse than 1.5

Since I installed SD I never noticed which version I was running >.>

[next flint]

like ur webui version is 1.4?

[tired vigil]

It says 1.4 at the bottom

A1111

[next flint]

i mean it adds more features including sdxl support when you upgrade

[brazen willow]

yeah it's hard to find a model that understands both the concept of a mushroom house and a flat vector art style

[brazen willow]

dreamshaper gets it

[fresh phoenix]

I am going to train a dreambooth model for white shirts since I am working on a project for clothes swapping. Any suggestions for instance_prompt and class_prompt for it? (Will instance_prompt: white shirt, class_prompt: shirt be suitable?)

[abstract gust]

I plan on having a nsfw male na'vi lying seductively in a hammock... How do I get that in the prompts?

[brazen willow]

hey i found a tutorial describing the original character creation process https://www.youtube.com/watch?v=iAhqMzgiHVw

[next flint]

Firstly you would have to find or make a lora

[strange meteor]

Exciting news, everyone 🎉!

Monster API SDXL now supports 20 amazing Styles!

Steer your image generation workflow towards a specific style.

Try it out here - https://lnkd.in/gH3j3p9F

[abstract gust]

Have that one available already - just need a hint for the prompt

[waxen wraith]

Does anyone know how to make pictures with the same face like https://www.instagram.com/millasofiafin/

[atomic wigeon]

What does this have to do with prompts?

[glad elm]

I keep getting the things I have on my negative prompts like blurry face and bad hands

how I can help it?

does it have something to do with my settings or sampling steps?

[silver valley]

Hey, yes it also has something to do with the resolution

[glad elm]

What should I change it to?

[silver valley]

The bigger the face the better quality it gets

So for example try a resolution of 512x768
Then add "Portrait of" before the lora

[glad elm]

768 is the height right

like this?

[silver valley]

Yea you can add the word tiefling after the of too

Portrait of tiefling ... Or portrait of name of character

[glad elm]

wont it hurt the image tho since I'm going for a full body?

[silver valley]

If you go for a full body then 512x512 is to small

But then say Full body instead of portrait

512x768 should work better

[glad elm]

Thank you

what is a good amount of sampling steps?

[silver valley]

30 is a good amount

[glad elm]

mine is 60 so overkill

explains why it takes 6 minutes to put out a single picture

[silver valley]

Yea xD

[glad elm]

I dont have space in my laptop so put it on external drive

so everything is super slow

[silver valley]

Make sure you use --medvram --xformers for better performance

[glad elm]

Whats that?

[tidal junco]

Does someone know how to help with these black squares in adetailer? Sometimes it happens when I make some stuff using adetailer

[silver valley]

You need to edit the webui-user.bat and at the COMMANDLINE_ARGS= line you add: --xformers --medvram
Then save and relaunch the webui-user.bat

They will speed up the generation time

And use less vram

[glad elm]

Thank you!

Needed this a lot

Oof got a broken eye again

Is the resolution still too small?

[silver valley]

Yea you would need to use inpaint or upscaling to get better faces

What's your GPU?

[silver valley]

Update the extension or reinstall it. Thats a strange bug. Should work on default settings

[glad elm]

Can you remind me where to check the GPU? its not on system-> about

I keep forgetting it

[orchid ore]

in task manager choose performance, here gpu and should be in top right

[glad elm]

Its geforce gtx 1650 ti

[glad elm]

Also what are inpaint and upscaling exactly and how I should be using them? Sorry if I'm asking too much, quite a newbie here

[silver valley]

Upscaling is enhancing the resolution of an image and getting better quality

[glad elm]

Why would I do that instead of creating it in higher resolution?

[silver valley]

You can do that with the Highres fix option in txt2img or with the ultimate upscale extension in img2img

[silver valley]

Because the 1.5 models are trained on a base resolution of 512x512.
Going over that results in duplicates, multiple legs etc

[glad elm]

What does turning on highres fix do exactly?

I see thank you

I also saw the restore faces option and turned it on

I'm guessing it should help a little with distorted faces?

Also just to make sure, this is how this webui-user.bat should look right

[tired vigil]

yes

[orchid ore]

Just with photorealistic faces, you can do it ex post in extras tab

[tired vigil]

@glad elm try inpainting a face if you can, its much better thn relying on getting it good first try

[orchid ore]

no need git pull, if something wrong happened you update it automatically. It is matter of personal taste.

[glad elm]

What does that mean?

[glad elm]

How do I inpait something and what does it do?

[glad elm]

I would rather keep it automatic cause last time webui didnt even start until I added that line in

[tired vigil]

press that little button under your image

it will send the image to another tab, do some face prompts and draw over her face

it will generate a face with better details and stuff

[glad elm]

Ooh thats perfect! Thank you

[tired vigil]

I can't show you settings right now I can't run SD

[glad elm]

All good, I will try to figure it out, thanks

[tired vigil]

@glad elm example from a video

the original option makes it remember what was behind the mask, and only masked makes you generate only that part again

and denoising strength changes how much it will change of the original image I believe, higher = more changes

[glad elm]

I see, thank you

so if I want minor changes on face like fixing some little blur I do inpaint masked, original, only mask small denosinig str

[tired vigil]

yeah

I haven't messed around too much with it since I can't inpaint with my GPU

[drifting anchor]

League of Legends, Yuumi the magical cat basking playfully on a large open hovering spellbook, gray fur with thin dark purple runes along the cat's back, emerald green eyes, bird's eye perspective
a saw in its right paw

[atomic wigeon]

Tip: Now this video is for PlaygroundAI's Canvas tool. But it applies to inpainting and how to use it in a unique way. In some art circles (Hollywood) this is referred to as AI mapping. It's one solution to clean up faces like you're trying to above. https://www.youtube.com/watch?v=bV2E7CAd2u8

You can zoom in and out on the face, clean it up, and then import it back into Image to image as a stand alone (like in a corner). The AI will fill in the scene based on your prompt.

This process combined with Davinci Resolve is the most powerful use of AI art post production on the planet. Blackmagic and AI art still images are a match made in Heaven. And it's free.

[atomic wigeon]

For the record, I don't work for either PlaygroundAI or Blackmagic. But making tangible and practical use of your AI art only starts with a good prompt. It ends with good post production software. I use Resolve and Stable Diffusion AI art on a daily basis. I only mention it because of the above conversation about faces. Resolve can help you. So can Premier Pro, Photoshop and a million other programs. Hope this helps.

[tired vigil]

how is this pose called

[atomic wigeon]

The name is irrelevant. Use controlnet.

[tired vigil]

okey thanks

i think is working

[atomic wigeon]

It could also be "defensive stance" or "ready to attack". That's what I put in my prompt.

[tired vigil]

thanks

[thorn swan]

How do I upload a photo to SD in Discord?

[twilit arrow]

Hierarchy of prompts never works for me. Maybe I'm doing it wrong. Is there a good guide for the most effective use of nesting with (), BLOCKs, ANDs, anything that helps SD attributes things correctly? I've even seen stuff like Location: a snowy mountain, Lighting: blah, ... but everything I try to add specificity ... (a woman, red hair) and (a man, blue hair) seems to just end up in a blender. Do any nesting promts or grouping techniques work at all?

[orchid ore]

() is to add more attention on things in it. More ((())) you add it will be more accented

[atomic wigeon]

Print Screen on your keyboard will take a screenshot of your entire screen. Control V to paste it in Discord.

Right click on any image and "Save Image As" or "Copy Image". Control V to paste into Discord.

See that + sign here in the chat box? That's how you upload an image from your computer.

[cedar gorge]

what would i need to say if i want to create charakters in this kind of style?

[atomic wigeon]

I wouldn't even know where to start with that one. It's obviously a cartoon illustration. It has a thin outline. Rich colors. Motivated light and dynamic lighting. Red spotlight perhaps. Deep shadows. Medieval style clothing. 2d. Fog. Mist. That's all I got.

[cedar gorge]

how do you know the lighting? also that a big help already

[atomic wigeon]

Experience. It's what I do. Motivated light is lighting conditions that force light on a subject in a dark room. As if the lighting sorce has no origin, it's jsut there. Commonly used in athletics commercials and advertisements. Red glow instead of red spotlight may be the way to go. It could potentially be "layered" as well. Commonly used as a paper effect.

[cedar gorge]

what does layered mean?

and how to do it?

[atomic wigeon]

There is a common art style that uses 3d layered paper, construction paper and cardboard. It was perfected in the Orient, but has found its way into Western art styles. Hold for a prompt. I did some recently.

.
----------- 3D Paper cutout of Abraham Lincoln -----------

papercraft-papercut shadow box of Abraham Lincoln playing a guitar, surrounded by flying peace doves, 3d, 32k resolution, thick brush strokes,

.
----------- 3D Paper Cutout Foodtography Style ------------

CNC cut 3d layered papercut shadow box using a Picasso oil painting, glass gloss reflective surfaces, liquid milk splash art , 32k resolution, foodtography style,

.
------ 3D Paper Cutout Foodtography Style #2 -----

CNC cut 3d layered papercut shadow box using a organicpunk digital painting, high gloss reflective surfaces, liquid milk splash art , 32k resolution, foodtography style, motivated lighting, centered, symmetrical, close up shot,

[cedar gorge]

ok i feel like its not really working 😄

[atomic wigeon]

What's not working? Those are just examples of an art style perfected by Eiko Ojala. They are not really for your project. But you asked.

[cedar gorge]

how do i know what to prioritize? like whats the first thing to say in the promt? that i want high quality or what i want to see on my art

oh i was refering to my project

[atomic wigeon]

You have to build it bud. I can only steer you in the right direction.

Your project style is not paper, but it is layered. Almost like the artist made paper cutouts (or stickers) and merged it all in AI.

[cedar gorge]

ive got this

[atomic wigeon]

That's the spirt

[atomic wigeon]

👍🏻

[cedar gorge]

mushroom warrior that is holding a leaf weapon, cartoon illustration with black deep shadows and dark outlines, rich colors, motivated light and dynamic lighting, red glow, high quality, forest background, darkest dungeon style, 2d, fog, mist

its wonky and not really the style but it kinda looks cool

[atomic wigeon]

It does look cool

Try to add a red spotlight or red glow

[cedar gorge]

i have red glow in the prompt

[atomic wigeon]

I meant red outline

[cedar gorge]

how do i save the top left one?

[atomic wigeon]

Sec.

[cedar gorge]

actually top right is cool as well

nvm im blind

they are all below

[atomic wigeon]

Layered cartoon illustration, 2d, thin black outlines, red outline, motivated light, dynamic lighting, red glow, dark comic style, fog, mist, mushroom warrior holding a leaf weapon

[cedar gorge]

oh you put the warrior at the end?

[atomic wigeon]

Try it. Focus on the scene first. Nail down the art style. Only move subjuct content forward if the AI is being stubborn or you want more focus on it.

[cedar gorge]

it generates right now

7 mins

ill get a drink 😄

[atomic wigeon]

Medieval attire or medieval fashion may be needed. That mask is 100% medieval. From the black plague era.

[cedar gorge]

I mean i mainly wanted the style, but more like the fantasy route with the mushroom warrior

Do you mean the mask in „the style of darkest dungeon“?

this is what i got with your prompt

this is way to bright 😄

[atomic wigeon]

Ya see, it left out the attire. That's important to stay true to your original image. Red spotlight is way too bright ya

[cedar gorge]

wait there is no red spotlight

[atomic wigeon]

Red glow + red outline ya. Too much

[cedar gorge]

i replaced red outline with medieval attire

[atomic wigeon]

I've never tried red outline before. I use red spotlight to force color changes in a scene.

[cedar gorge]

i wish i could run multiple bots at the same time

[hoary galleon]

So I'm having trouble with prompting beacuse it seems that the bot just won't do it lol, so I prompted this: "1968, a woman wearing a space age dress and hairstyle, holding a stereotypical green alien baby with big dark eyes" What I'm trying to achieve is this first picture, with an actual alien baby but retro-ish, the just keeps giving me human babies, how should I re-do the prompt to make myself clear to the bot? Pictures below for comparison of what I want and what the bot is giving me.

[atomic wigeon]

In playgroundai.com you get 1k free images a day. You can run both Board and Canvas at the same time.

[cedar gorge]

what is board and canvas?

sorry for all the noob questions i just installed all this yesterda

[atomic wigeon]

PlaygroundAI is a website that is divided into two tools. Board is one image generator. Open a new tab and Canvas is their AI generator, in painting, edit, kind of all around good project tool. Pretty cool. Free 1k images, $15 bucks a month for 2k. Per day.

[atomic wigeon]

Lemme try something and I'll respond after.

[atomic wigeon]

Ya I got human babies too. Trixy one. I'll keep trying.

[hoary galleon]

Alrigth

[atomic wigeon]

I'm trying "Star Trek style" and a different set on another site.

LOL dang, that's a challenging one.

[next flint]

regional prompting?

[cedar gorge]

oh god that right one is cursed

[atomic wigeon]

I'm using Playground

[next flint]

oh

[hoary galleon]

wow what website did you use? Regardless, I would like to achieve this by just using the bots, it seems easy but its' a hard one I guess.

[atomic wigeon]

www.playgroundai.com gives you 1k images a DAY for free. 2k images a day for $15 bucks a month. Has built in controlnet, filters, edit tools and other toys.

Playground is the second biggest AI generator on Earth.

[next flint]

damn. i hadnt even heard about it before you showed that you were using it

[cedar gorge]

midjourney was the first one right?

[atomic wigeon]

realistic tiny alien baby with blue skin and spots being held by a tall white American woman, 1960's style,

[atomic wigeon]

No. Midjourney came from the StabilityAI split.

[atomic wigeon]

That's SDXL 1.0

[hoary galleon]

did you did it on here?

[atomic wigeon]

No. www.playgroundai.com

[hoary galleon]

This is the closest I have got, it's laughable lmao

[hoary galleon]

Oh that's free rigth?

[atomic wigeon]

1k images a day for free yes. Only need a gmail to sign up.

American company

[hoary galleon]

Did you use negative prompts or anything else?

[atomic wigeon]

No neg prompt on that one

[cedar gorge]

i gave a a few more spins with some edits, but i feel like im not getting closer. actually getting further away. i dont know why there are 2 humans now while i want one mushroom.

Layered cartoon illustration, 2d, thicc black outlines, deep black shadows, Medieval attire, motivated light, dynamic lighting, red glow, dark comic style, fog, mist, mushroom warrior holding a leaf weapon

[atomic wigeon]

I lied, there was a neg prompt. Sorry.

poorly rendered face, poor facial details, poorly drawn hands, low resolution, images cut out, bad composition, mutated body parts, blurry image, disfigured, oversaturated, bad anatomy, deformed body features, out of frame, text, error, cropped, worst quality, low quality, jpeg artifacts, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, used fingers, too many fingers, long neck, watermark, signature, blurry, bad anatomy, extra limbs, poorly drawn face, poorly drawn hands, missing fingers, letters, numbers, fonts, text, words, symbols, autographs, nsfw, pink

[hoary galleon]

okay thanks and also sorry for asking so many questions but, did you use any specific style?

[atomic wigeon]

Thicc and thick and different.

[atomic wigeon]

No. When you use Playground's Canvas it defaults to SDXL 1.0 without a filter.

[hoary galleon]

Alrigth thank you!

[atomic wigeon]

That was the entire prompt. in default.

[atomic wigeon]

Tring again

Trying

[cedar gorge]

yes, im at 55

[atomic wigeon]

55 wehat?

[cedar gorge]

percent

[atomic wigeon]

Oh. I default to 150. Only lower as needed.

[cedar gorge]

wdym?

i could generate faster?

[atomic wigeon]

No, 150 is slower. However, on Playground it's faster than local.

Better result

Soemtimes. And other times less is more. Lower steps is much cooler in some projects.

[cedar gorge]

oh damn this is much better

[atomic wigeon]

2d Layered illustration, cartoon, deep black shadows, medieval clothing, black plague vibes, motivated light, dark comic style, fog, mist, mushroom warrior holding a leaf weapon, faint red spotlight, dungeon background, evil,

[cedar gorge]

cool! did you do this in playground?

[atomic wigeon]

Yes.

[cedar gorge]

what is prompt guidance?

[atomic wigeon]

12

On the right

[cedar gorge]

and could you please share your exclutions?

[next flint]

im not sure how you would implament it but it kinda seems reminiscent of the hades art style

[atomic wigeon]

poorly rendered face, poor facial details, poorly drawn hands, low resolution, images cut out, bad composition, mutated body parts, blurry image, disfigured, oversaturated, bad anatomy, deformed body features, out of frame, text, error, cropped, worst quality, low quality, jpeg artifacts, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, used fingers, too many fingers, long neck, watermark, signature, blurry, bad anatomy, extra limbs, poorly drawn face, poorly drawn hands, missing fingers, letters, numbers, fonts, text, words, symbols, autographs, nsfw, pink

[cedar gorge]

i meant more like, why 12 because i dont understand rihgt now what it does. only says higher value gives me something closer to my prompt. so why not make it 30?

[atomic wigeon]

More is worse. That's the amount of AI freedom you're giving to both your image to image strength and your sampling (seed variance).

7-15 is optimal. 15-20 only if the AI is being super stubborn.

[cedar gorge]

oh ok i get it

ok i really like the results now

thats pretty cool tysm

[atomic wigeon]

🙂

Gonna test it in ReVAnimated. Which is like a super high quality soup of stuff from SD 1.5

[cedar gorge]

ok i went through all the struggle yesterday to install SD just to realize that playground is much easier and faster

[atomic wigeon]

Playground is amazing. And you can remove background right at the image. Lots of stuff for edits and post production. It's my #1 favorite AI generator. Plus it doesn't do that tokens or credits bullshit.

[cedar gorge]

oh i can edit as well?

like lets say if i like the guy in my 3rd image and the background of the 5th one?

[atomic wigeon]

Yes, Canvas is their editor. Generate images there, assemble them, enlarge them up to 10k resolution, on and on and on.

Upload that image into Playground. Then crop that guy out. Now put it in image to image and spin it with the prompt.

You can inpaint in Canvas too. And in quadrants. So you can jsut remove the background or jsut remove the guy.

ControlNet (Control Traits) is in the bottom left corner of Canvas.

[cedar gorge]

ohh

ok alot of stuff i still have to discover

[atomic wigeon]

Lots of videos out there.

Before controlnet

After controlnet

[cedar gorge]

its ot really a mushroom anymore but for sure looks very cool

[atomic wigeon]

Well just keep playing with the prompt.

That's the fun part

[cedar gorge]

oh you just took him and basically put a filter on it

[cedar gorge]

thats what im doing right now 😄

[atomic wigeon]

Yes. Controlnet is an edge detect software. The AI recognizes the main subject content and locks it down so it cannot be altered. Then you prompt builds the scene around those internal lines. It's included with Playground.

[cedar gorge]

mmh right now its more like doing warriors with mushroom weapon and not mushrooms with leaf weapons 😄

[atomic wigeon]

Lol. The priority in your prompt needs to be adjusted. Move priority to the front of the prompt.

[cedar gorge]

getting closer

[atomic wigeon]

There is a Dark Comic filter on Playground (left side). That may help.

Just know that many of the filters give SDXL, but only the top 5 work on SDXL 1.0

[cedar gorge]

ive seen that. what is sdxl?

[atomic wigeon]

Ok...

• Playground V1 is a custom SD 1.5 only available at Playground. V2 is in production.
• SD 1.5 is Stable Diffusions last major update for general use stuff.
• SD 2.0 is dogshit. Dont use it.
• SDXL is a major update to SD 1.5
• SDXL 1.0 is a major upgrade to SDXL (the current best)

[cedar gorge]

i thought i am using SD 1.5

[atomic wigeon]

If you jsut got to PlaygrounAI for the first time make sure to change your sampler in the bottom right. Use k_euler_ancestral

[atomic wigeon]

You mean local or Playground?

[cedar gorge]

on the local

[cedar gorge]

ok, what does it do?

[atomic wigeon]

SDXL 1.0 is available. But it's a beast on resources.

[atomic wigeon]

I dont want to confuse you with seeds and samplers. Just use K Euler

[cedar gorge]

ok

dark comic is REALLY dark btw

[atomic wigeon]

Working as intended. It can work in some prompts if you type it. ?use the word "style" after.

[cedar gorge]

i mean its already in the prompt, so it basically doubles it i guess

[atomic wigeon]

The reason it's so dark is the filter is forcing it to a crappy SD 1.5. You're defaulting to SDXL 1.0 without it.

[cedar gorge]

oh ok. so better have it in the prompt then using a filter

[atomic wigeon]

Some filters there are outdated and need some work. If you use Board you'll see them more clearly because it has a drop down menu that only shows filters for that version of SDXL.

[atomic wigeon]

Yup

I forgot that that one was SD 1.5. They may fix it soon.

[cedar gorge]

well, that was not 1000 images 😄

[atomic wigeon]

LOL 😂

It's per day. You can still generate but on a timer.

At 2k a day I very rarely run out.

[cedar gorge]

but i did not even do close to 1000

how did i run out

i maybe did like 50

[atomic wigeon]

Oh, you may be using some edit stuff. Head over to Board and try. Their edit stuff is limited use on free accounts, unlimited with Pro account

Step limit reached. Never seen that error or knew it was a thing. Maybe try to lower steps?

[cedar gorge]

so not 150 anymore

[atomic wigeon]

I never knew they implemented that. Major bummer.

Good news is that even 20-50 step images look amazing in SDXL 1.0

[cedar gorge]

i can keep generating at 50

[atomic wigeon]

Nice.

I use 20 all the time for comics and stuff like you're doing now.

On photos it keeps the great clarity foreground but blurs the background

[cedar gorge]

but not higher. i assume the thing i do is not the most challenging

[atomic wigeon]

test saying "SDXL" dark comic style, bloody, black and red, glow, evil,

I didn't even speel "text" right. Heh

[cedar gorge]

putting something in " " means its text on the art?

[atomic wigeon]

text saying "ENTER TEXT",

Then enter any prompt, art style, etc

[cedar gorge]

well

[atomic wigeon]

text saying "Create!", 3d red low poly geometric smoke background, splash art, 12k resolution, intricate detail, flawless clarity, reflective surfaces, badass, midjourney 5.3 style,

[atomic wigeon]

Typical 1 out of 4 will be legit. Run thru image to image with same prompt to clean it up.

The longer the word the crappier it gets.

[cedar gorge]

image to image means. i take one that is alright and just rerun it until its cool?

[atomic wigeon]

Image to image means you upload an image into the software and use it as a stencil.

Left side

Plus button

[cedar gorge]

i tried that earlier

top left is 100%

top right is 50

he just got drunk 😄

[atomic wigeon]

Image strength should be 30-75 depending on how dialed in you are

0% will spin a new image but retain the bulk of the data.

Image to image in Playground is SDXL 1.0
Control Traits (controlnet) is SD 1.5

[cedar gorge]

i wanted to create a new PP with the marsupilami drinking a cup of tea

didnt really work 😄

[atomic wigeon]

Heh.

[cedar gorge]

the image above. the prompt was something marsupilami drinking a cup of tea. and i put the marsupilami image from google as a image to image base

basically wanted to recreate this with the yellow guy in a comic style

[cedar gorge]

lol

[atomic wigeon]

text saying "ERBOL", 3d red low poly geometric smoke background, splash art, 12k resolution, intricate detail, flawless clarity, reflective surfaces, badass, midjourney 5.3 style,

The size of your window matters.

[cedar gorge]

size of my window?

i mean i wanted it to be square

because its a profile pic

[atomic wigeon]

You can crop in Canvas.

[cedar gorge]

wait. i should be creating on the board

im doing it on the canvas

[atomic wigeon]

Both are good. Slightly different functionality. Board has the advantage of face restoration, create variants of one image and some other toys.

Use them both at the same time to be more efficient. Just open one in each tab.

Cant use two Board or 2 Canvas. Has to be one of each.

So 1 Board, 1 Canvas and 1 local all at the same time. : )

[cedar gorge]

i still need to learn to handle one

[atomic wigeon]

My advice is focus on Canvas. Where the industry is about to go you'll want to master that and Davinci Resolve. Pro tip.

[cedar gorge]

the industry? i just wanted to this for fun, not get a professional Ai artist

im still lost with the control traits. like i want to combine the two images and alter the style i just dont know where to start

[atomic wigeon]

Artists use AI as a medium. Because they do it doesn't make them "AI artists". There's no such thing.

If you're good at AI art you're a good artist.

[high leaf]

sd does not make art. it's a glorified picture merger. your prompts and options are the real deal

[atomic wigeon]

That's partially true

AI does assist with synthetic generation

[cedar gorge]

well i dont think that is true, just because i know what looks pleasing doesnt mean i would be able to create that with a brush or pencils

[high leaf]

i actually see sd as a 100% replacement to photoshop

[atomic wigeon]

I do as well. That's why I support SD and PlaygroundAI. Their "Canvas" is the future Photoshop.

[high leaf]

200% better then inpaint or sketch i agree

try the "make it winter" with SD inpaint

[atomic wigeon]

Canvas includes inpainting and sketching (draw to edit).

[atomic wigeon]

LOL that's quite literally one example Playground uses in their latest video. They do.

[high leaf]

sd-canvas-editor?

[atomic wigeon]

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

[cedar gorge]

dont know why he sometimes looks like a drug addict, but i guess im getting closer

[atomic wigeon]

Try "adorable eyes", "cute adorable eyes" or "Pixar style".

[atomic wigeon]

All that and you can collaborate too.

Pretty cool right?

[high leaf]

not free but I can see potential

[atomic wigeon]

It is free.

1000 images a day. Free

2000 images a day for $15 bucks a month (same as free).

How much is Photoshop?

[high leaf]

i don't see a market for 2000 crazy images a day. time will tell

i don;t pay for anything. Ps and Lr are free on torrent

[atomic wigeon]

Crazy? I can do 100% of anything I would do in Photoshop in Canvas.

[high leaf]

the face swap thing is particularly appealing

[atomic wigeon]

It's a bit rough at times. They're working on it.

It's powerful software. And I'm glad to see at least someone doing something tangible and realistic with AI art software.

[high leaf]

but it's still a niche. I know lots of people who despise this kind of art

[atomic wigeon]

People despised the combustible engine.

It's evolution.

[high leaf]

we'll see! thanks for chatting

[cedar gorge]

maybe i have too high expectations, but these are all not what i want

some of them have cool elements, but they all look weird

[void patio]

hi all, did anyone else's sdxl Canny just died?

[sullen rover]

What's a good prompt to replicate this flat shading with shadow style?

I tried this but it's meh
,a couple of tables that are outside of a building, flat shading with shadows, scene from Alto's Adventure, Alto's adventure, , industrial space, inspired by Kieran Gabriel, awnings, orange and white color scheme, empty buildings, storefront, partially operational, reconstruction, low poly, vivid colors, 4k, high quality, by Kieran Gabriel, hard shadow

Artwork by Kieran Gabriel btw

[void patio]

or maybe someone can share their sdxl CN workflow?

[twilit arrow]

Can anyone explain to me what Batch Size and grids are for? When would I ever need that? I always keep Batch Size = 1 and use Batch Count = 10 or whatever if I want to generate 10 images, and I see zero use for that extra grid image since I use Adobe Bridge to review the work. I remember reading that images generated in Batch Size go much faster because it essentially copies some of the work within the batch, but I usually want very different output. Why/how do you people use Batch Size?

[silver valley]

If you have enough vram Batch Size can generate multiple images simultaneously. So you'll be faster with it. Also if you dont set a seed you should get different outputs

[orchid ore]

@sullen rover i think those sharp shadow could be part of early morning or late evening tags

[echo oak]

are there parameters for SD promt ?There are parameters for Midjourney such as --no --iw --chaos etc

[orchid ore]

sharp? Or render in higher resolution and downsampling?

[atomic wigeon]

Sharp focus, intricate detail, cinema quality, flawless clarity. And a good neg prompt helps:

poorly rendered face, poor facial details, poorly drawn hands, low resolution, images cut out, bad composition, mutated body parts, blurry image, disfigured, oversaturated, bad anatomy, deformed body features, out of frame, text, error, cropped, worst quality, low quality, jpeg artifacts, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, used fingers, too many fingers, long neck, watermark, signature, blurry, bad anatomy, extra limbs, poorly drawn face, poorly drawn hands, missing fingers, letters, numbers, fonts, text, words, symbols, autographs, nsfw, pink,

Neg prompts effect all subject matter. But they can hinder you as well. Try with, try without.

Wider, longer, further = blurrier. There are ways to mitigate this. But it will take too long to explain. Watch some videos on photography.

[torn cliff]

Is it possible to use one model meant for humans to generate a human with no background, then use a different model meant for scenery and just edit the human onto the background?

Then why are you using prompts designed for humans?

[atomic wigeon]

Yes. Midjourney is famous for this. You'll see a lot of focus on GPU render in the foreground and blurry backgrounds. SD has built in trigger words. One of them is Midjourney. So add "midjourney 5.2 style" to your prompt.

You can use multiple images to zoom in and repair face then emboss on a background after.

[torn cliff]

Only real problem I have with backgrounds is any time I try and make an image of a snow leopard (feral or anthro) it almost always adds snow... even if they are indoors...

[atomic wigeon]

It's not the AI's fault, it's yours.

[torn cliff]

Someone suggested I try clouded leopard instead, and then every image had clouds. Lol

[atomic wigeon]

The AI results are what you tell it to produce. You have to find a creative way to be more specific and detailed.

[orchid ore]

How about white leopard in summer?

[atomic wigeon]

Some tricks include indoors and outdoors.

You can always use image to image to help guide your prompt.

[torn cliff]

I was able to somewhat get it to stop by using "wood floors", that at least got it to make ones without a patch of snow in the center of the floor. Though they still tended to have a window with snow falling

[orchid ore]

havent checked 3 and 4

[atomic wigeon]

Blurry background is steps.

Increase your steps

[torn cliff]

It's partially a result of the models anyway. I've noticed that most of the anthro models tend to have difficulty with more unique species. Like I was able to get results for "Wolf" just fine, but "Snow Leopard" or "Hyena" didn't work as well.

[orchid ore]

this is just and only white leopard, going to try snow one

[torn cliff]

Never thought to try white leopard.

[atomic wigeon]

Giraffe too

[orchid ore]

is this snow?

[atomic wigeon]

Fluff it up with Lynx. Add "soft and fluffy fur" and "add static to scene". The static electricity will lift the hair up and make it more detailed.

[torn cliff]

Is there any way to stop or limit color bleed? Like if I specify eye or hair color it tends to bleed into clothes, or vice versa depending on which comes first.

[atomic wigeon]

intricate detail and intricate details are two different mechanics. Similar and subtle, but the plural will add more junk to photography. Good for splash art.

And the bleeding into clothes thing is something I struggle with a lot.

Sometimes if I remove background and pull the character over another background image and add "3d" to the scene it mitigates that.

[torn cliff]

I did find that sometimes putting quotes around them, like "pink hair", "blue dress" did stop the bleed, but it doesn't always work.

[atomic wigeon]

Quotes dont really work. It encourages text

text saying "INSERT WORD" is how we intentionally do text.

[torn cliff]

Never saw that happen with my attempts. Least using Automatic1111 and some of the models from Civitai

Then again I also had Text in the negative prompts so that may be why

[atomic wigeon]

It's an SDXL 1.0 thing

[orchid ore]

How about flat colors, but may be as well unwanted.

[torn cliff]

Ahh I haven't tried using SDXL yet. Only just started trying SD Next

[atomic wigeon]

muted colors

rich colors

vibrant colors

happy colors

sad colors

You can also random colors or random background color

[orchid ore]

I thought flat color may not bleeding 🙂

[torn cliff]

Found that SD Next is a lot faster than Automatic1111 using the exact same size, prompts, etc. But after a few generations it tends to start claiming that there isn't enough memory and wont stop it until it's either closed and restarted or the size is changed to smaller and then back...

[atomic wigeon]

The only local I have is not Stable Diffusion so I cant help with that sorry. For all SDXL 1.0 I use playgroundai.com

[torn cliff]

Not sure if Flat Colors would help, any mention of color may end up affecting hair, eyes or clothes.

[orchid ore]

@torn cliff whats your gpu?

[torn cliff]

Radeon RX 6800 XT

[atomic wigeon]

You could also try adjusting the size of your image (has a major impact on some content) and adding a camera lens: https://rikkar69.github.io/SDXL-artist-study/cameras/

[torn cliff]

I'll have to try that when I test SDXL.

[atomic wigeon]

^ that is one of the most valuable sites you'll find for photography in SDXL 1.0

It works in SD 1.5 as well

[orchid ore]

ou, your best way is to use linux with AMD cards. I like AMD but unfortunately its support in Win is bad.

[atomic wigeon]

Try your snow leopard again. This time add this to your prompt: Sigma fp with Sigma 45mm f-2.8 DG DN

[torn cliff]

Unfortunately I will have to wait on trying SDXL because SD Next currently isn't stable for me. It's ok for one image at a time, but I cant run batches cause of the memory issue.

As for Linux, I would switch to it in a heartbeat if all programs and games would work for it. I HATE Win 10...

[atomic wigeon]

www.playgroundai.com gives you 1k free SDXL images a DAY! A DAY!

[orchid ore]

sigma has not much point on rgb monitors 🙂 I mean foveon vs RGB patterns. It is like watching color movie on black and white TV 🙂

[atomic wigeon]

I use that lens almost daily. It's amazing the results

[torn cliff]

LOL, my txt2img folder in Automatic1111 alone has over 4000 images in it 😛

[orchid ore]

oh it is lenses. I thought it is camera

[atomic wigeon]

Only 4k? Slacker

[torn cliff]

I delete the badly warped ones

And any ones that weren't what I wanted, cause I had forgotten to move SD off my C drive which I try to keep only for Windows and other programs that need to be on C. Everything else gets installed to other drives.

[tired vigil]

@torn cliff how many iteration/sec do you get?

I'd love to get a 6800

[torn cliff]

Uhh I forget what I got on Automatic, but a 500x1000 (500 is easier for me to remember than 496) image with 30 steps was taking me 4 minutes. But the same on SD Next was only taking me 40 seconds

I'll check now

[atomic wigeon]

I only use Radeon for professional work. All images and video editing. I just get much better results vs 3090-4080. And ya, the 6800 and 6850 are outstanding cards!

[torn cliff]

I was recommended the 6800XT because I was looking for one that would play most games but not cost a fortune.

[orchid ore]

on sdxl i think it must be posible to divide by 16, otherwise bad image, not sure if in A1111 as well, so 512 is very important vs 500 as well as 1024 vs 1000 for example @torn cliff

[torn cliff]

Oh I know, but 500 gets changed to 496 automatically by Automatic1111, that's why I said its easier for me to remember 500

[torn cliff]

6.64 it/s with Automatic1111

[tired vigil]

6800s are only a bit more than 3060s for me

[torn cliff]

Cost me $900 cause I'm Canadian and we get screwed on the exchange rate...

[atomic wigeon]

Honestly the last couple gens of Radeon are still really good. Get what you can afford.

[tired vigil]

I get 6-8s per iteration on my rx 580 haha

[torn cliff]

I was given a brand new AMD cpu by a friend so I was able to save some money that way.

[orchid ore]

iterations depends on resolution as well, cant be separated. As well sampler matter very.
i got 1,38 it/s with 2048x1152

[atomic wigeon]

Oh and here's the Sigma lens.

[tired vigil]

I stick to 768 max resolutions

[orchid ore]

i thought sigma camera. It is foveon vs Bayer

it is must in 1.5 or 2.1 models that res

[tired vigil]

Is that good ?

[tired vigil]

Sometimes I do "pale" instead of white when trying to specify skin color or fur

Cause of the bleeding

[torn cliff]

Uhh, wouldn't a higher it/s mean faster generation and lower it/s mean slower generation?

[orchid ore]

no.

[torn cliff]

I just tried the same prompts on SD Next and got 1.64 it/s but it only took 49 seconds instead of 4 minutes with Automatic1111

[orchid ore]

you can compare only with same resolution, and same sampler + same model and prompts maybe

also you can be confused by it/s and s/it

[torn cliff]

Oh right, it does say s/it, big difference then

[orchid ore]

yes

[torn cliff]

@tired vigil The only thing about the 6800, at least with Automatic1111, is it says that training doesn't work with AMD cards.

[atomic wigeon]

Here are some more examples of adding that Sigma lens to your prompt.

Sigma fp with Sigma 45mm f-2.8 DG DN

[torn cliff]

Not bad. I know nothing about photography or lenses so I never bothered trying with any of those.

[atomic wigeon]

That's why I recommend you save this link. The guy went thru a lot of trouble to help all of us prompt. It's only purpose in life is prompting like a pro. https://rikkar69.github.io/SDXL-artist-study/

[orchid ore]

sigma camera is great for red color, because got red green and blue not spatial but each behind other. It means each pixel got exact color. Bayer needs 3x resolution and math to get 1 pixel, which is made out of 4 pixels, tends to be green as there is red, green, blue and green.

[atomic wigeon]

You can add "spotlight" or any color spotlight to a scene to offset and alter the effect.

[orchid ore]

Only posibility to compare colors is on physical photography imo

[atomic wigeon]

This is "orange spotlight" added to my prompt. Now watch when I add "aqua and red spotlight".

That's only a spotlight making those changes. Nothing else.

[orchid ore]

nike vs Fila 🙂

looking great

[torn cliff]

I've found that narrower and taller resolutions work better for making full body images, square images like 512x512 seemed to always be thigh up. Is there a best size for full body standing poses?

[orchid ore]

yes your findings are right

[atomic wigeon]

I was just having fun with kicks. Kicks with kicks. 😆
But I wanted to demonstrate how film or camera lens is irrelevant if you have light control in the prompt.

[orchid ore]

512x768? or 512x1024? but can have issues with 1024

[atomic wigeon]

Tip: You can make a 512 x 512 and import it into playgroundai.com's Canvas. Make it 10k resolution then import it back into image to image.

[torn cliff]

I'd have to test 1024, much above 1000 will give me out of memory errors. The 6800XT has 16GB vram but not enough apparently

[atomic wigeon]

If you have the paid version of Davinci Resolve you can import your image there, make it 32k res, then import back into image to image. Easy ways to bypass GPU reqs.

The highest free that I've seen is playgroundai.com. You can drag an image up to 10k.

[torn cliff]

I've only used SD with Automatic1111 and SD Next so far. Never used any of the sites, especially the pay versions. Seeing how many "bad" images come up I can see how fast free "credits" or "images" would go while trying to make one good one.

[atomic wigeon]

I would never recommend any site with credits or tokens crap. PlaygroundAI is free 1k images a day or $15 dollars a month subscription for more toys + 2k images a day. I dont work for them or represent them. Just highly encourage it.

And Davinci Resolve is free as well. The free version is unrestricted, no watermarks or anything. It's 90% of the content that's in the free version. Paid version is $300 and replaces Premiere Pro.
Resolve free = CPU render
Resolve paid (called Studio) = GPU render

^ together they are the most powerful post production AI image software on earth.

[torn cliff]

Ahh yeah I won't consider Resolve free right now then. I need to get a liquid cooler for my cpu before I consider that. My case isn't as great for airflow as I had hoped.

[atomic wigeon]

Ya. Or take the side of your somputer off and add floor fan. Making a movie or video in Resolve will melt your CPU lol. BTW every Hollywood film goes thru Blackmagic Davinci Resolve.

For quick image edits tho you're heat will be ok. No issues.

[torn cliff]

I did that years ago with my first desktop. GPU fan died on it and I didn't know, any time I tried to play Minecraft it would freeze the whole pc. Installed a temp monitoring program and found out that the gpu temp was hitting 100C

[atomic wigeon]

ya video editors and games are brutal. But I highly recommend you guys consider a good video editor for fun alterations and post on your still images. Pro tip.

Video = really fast moving still images.

[torn cliff]

I'm not going to be doing any video or animation editing, I've got no artistic skills at all. Will take a look at playground later though, getting a rather sharp stabbing pain in my head right now though so not going to do anything that requires thinking...

[atomic wigeon]

Hope you feel better. Take care.

[torn cliff]

Thanks. And thanks for the advice as well.

[atomic wigeon]

You're very welcome. 🙂

[atomic wigeon]

And by the way, that's what I'm trying to say. Video editors edit still images.

[iron coral]

I am in need of some advice for training a scene/setup/action: the main idea is to train a LORA or embedding to have a driver in a car. However, the current models/LORAs out there do not fully have the driver on the controls (i.e. hands on steering or shifter, feet at the pedals, etc...). I have done LORA training on specifically connecting the legs and feet to pedals but my LORAs tend to over take the style I am looking for.

any good advice on how to approach this?

[glad elm]

I started getting this error now when I didnt use to before, whats wrong?

[orchid ore]

What is your GPU? and what do you have in Webui-user.bat?

It is more for tech support

[glad elm]

geforce gtx 1650 ti

also true, want me to switch to there?

[orchid ore]

hm only i can adviced is lowvram, but it hit performance. close it and restart again.

[glad elm]

I see thank you

whats the difference between medvram and lowvram?

[orchid ore]

it needed even less vram, but hit performance very much.

[torn cliff]

Does upscaling with PlaygroundAI count as generating an image and take away from the 1000 a day?

[atomic wigeon]

I doubt it. It's a method of saving an image. I honestly dont know. I subbed after like 1-2 hours. Been on sub for 6 months now.

I do know that they restrict or limit some stuff in Canvas with a free account. Pro account is unlimited.

[torn cliff]

I don't see anywhere that it tells you how many you've created...

[atomic wigeon]

It doesn't tell you. But as you get close it will display it under the Generate button in bottom left.

[torn cliff]

Not really a good setup there, no one is going to count to keep track of how many they generated, especially for 100+ Oh well.

[atomic wigeon]

I mean, it's 1k images. Unless you're using "Board" and doing "Create Variants" you will have a lot of time to burn thru those 1k. Careful prompting and adjusting and you'll almsot never hit the 1k.

But even if you hit the cap you can still generate. It starts a 30 second timer.

The timer increases in duration with each image.

[glad elm]

by performance do you mean imge quality or generation speed?

[torn cliff]

Yeah I'll probably only use it to upscale images I generate with the local install of SD. But the quality looks pretty good.

[atomic wigeon]

Wanna see seomthing cool? Check out that ReVAnimate filter. It's SD 1.5 but amazing!

Warning tho, ReVAnimate adds smooth and round edges. The AI will want to sexualize humans.

[torn cliff]

Yeah I tried that using an image I generated with Revanimate as a base and the same prompts. First one wasn't perfect but looked cool. Second one had some flaws so I tried to redo it and it ended up looking like someone twisted a person like you would wring out a towel 😛

[fast kite]

hi all, not totally sure if this is the appropriate channel but i'm wondering if anyone can offer any guidance for a particular challenge i'm having -- i've trained a couple of SD models on images from the internet of posed "outfit of the day" images. i'd like to generate images that reflect the training data as closely as possible. i'm using 1.5 and euler a which is giving me pretty good results, a bit grotesque and strange which is fitting for the project. but i'm wondering how better to get images that look closest to the sample images that the model training gave me... maybe there's a prompt that could get back to this? i'm looking for a photorealistic approximation of the training data, essentially

[atomic wigeon]

ReVAnimate is terrible with fingers, claws, talons, toes, etc. Best to medium shot or closeup.

[atomic wigeon]

Are you using a negative prompt?

[torn cliff]

Oh it wasn't the fingers that were the problem. LOL It literally looked like someone wrung out a towel

[fast kite]

the negative prompts i'm using are just to eliminate issues that the seemed to have with clarifying the subject. "boy, man" is enough for one of the models (trained on many images of the same woman) for example

i've managed to get pretty close to the sample images that i got from training the model by describing the average sample image in detail and using euler A, but i'm wondering if there's an easier way to do it?

[atomic wigeon]

It could be your prompt then. Try one of my old prompts. Remove and replace what you dont need:

3d model UE5, Young American soldier running into combat, action scene by Peter Jackson, explosions, debris flying, shrapnel, sad expression, depressing mood, realistic, dust in air, light rays, sharp focus, intricate detail, 32k resolution, Sigma fp with Sigma 45mm f-2.8 DG DN

That's from K-euler ancestral

[fast kite]

could i replace the lens info, etc with something to give an effect closer to the sample data? which are all iphone mirror selfies

interesting, i wonder if the best approach will be working up to the original data by clarifying the prompt. i just liked the images from training the model so much and wanted to see if there was a way to generate a dozen or so of those at higher res...

[atomic wigeon]

Ofcoarse. The lens is a trigger for professional photography. Adjust it and add portrait, photograph, photographic, cinematography, etc. Put your prompt in this prompt builder and it will alter it for your needs. Pick an art style and generate. The tool will give you a perfect prompt. https://codepen.io/TheCopernicus/full/mdQZGpW

[fast kite]

oh looks perfect, thank you

i'm inspired by some work on GANs from this guy a while back: https://ai-clothingdaily.tumblr.com/

i'm so curious what he's using to get these results -- basically approximations of hundreds of images of training data

[atomic wigeon]

Nice. You may also try to utilize some of the art styles at this link. It's designed for prompt building in SD. https://rikkar69.github.io/SDXL-artist-study/

[fast kite]

do you have any impression of how he might be getting the results from the tumblr link? i'm really trying to get to that level of detail/abstraction -- the images that are generated reflect the source image super closely but the weird little deviations make for such interseting results

[atomic wigeon]

It looks like he's combining monochrome with another style.

[fast kite]

https://64.media.tumblr.com/ae40d2b4e5021db9e740e5d8717435d9/e7a4ff0bc951c8e1-ba/s1280x1920/afdd66d11f49c21a9e9cf56fad7983058d0b61e7.png for example

hmm

monochrome style in the prompt?

https://64.media.tumblr.com/12a508ed034b10793c4109313b47a44c/fc9d198eaca9c02d-df/s1280x1920/339c5a87a01e64f22a50caa3b9abba85bd019db4.png these also are close to what i'm aiming for with respect to the lvel of detail / abstraction

[atomic wigeon]

monochrome or monochrome style, yes. But that image you jsut linked is not monochrome. The background would be "radiant white background" maybe. The word "Noir" might be of use to you as well.

[fast kite]

so interesting that the prompts are this important - still getting used to it all. i was under the impression that this quality of image had more to do with the bulk of training data

but it seems that it's a combination of good data to train the model on and a refined prompt

[atomic wigeon]

Yup, prompts are everything. That's the art in AI art.

[fast kite]

thanks for the help, i really appreciate it

[atomic wigeon]

Good luck on your model.

[fast kite]

thank you!

[orchid ore]

@glad elm sorry wasnt here, generation of images.

[glad elm]

So the quality?

[orchid ore]

no quality not. Quality hit can be only using less steps or lower resolution imo

speed of generation hit is big

[glad elm]

I see thank you, so setting it to low decreases the speed but makes it use less vram

[quiet zodiac]

Does anyone have tips on prompting visible aspects of superpowers? I have...
sonic scream with visible elements
But no visible elements are added to the image.

[torn cliff]

Maybe try sound waves?

[quiet zodiac]

Another one is...
precognitive mental waves

[torn cliff]

Is there and prompts, models, loras, etc that can make decent tattoos without the tattoo being just one design or in just one location? Like if I wanted a butterfly tattoo on a shoulder or a flower tattoo on a hip for example

[torn cliff]

Not sure if it will work, but maybe "purple ripples around head" ? That tends to be what signifies psychic power usage in images

[sterile lily]

Having trouble with my prompt that is causing too many "flagged" images or blank/black images. I have uploaded an image of my son in a vampire costume (see attached with face removed), then prompted "boy rising from a coffin in a basement, showing his fangs". Half the time it gets flagged, the other half it shows up blank. Tried about 20 images so far. Any help on variations of this prompt? When I remove the uploaded image, it shows me a desired result (see attached), but with a creepy grown person. Any help would be appreciated.

[quiet zodiac]

Maybe try...
a cute kid in a vampire costume showing his fangs while emerging from a coffin in a basement
It could be your mix of boy and rise.

[sterile lily]

@quiet zodiac Tried your exact prompt on 4 images, it stops processing immediately, cancels the action, and shows a brief popup in the bottom right of my browser that says "Something is not quite right with your prompt". Removed the word "cute" and the same result. Also tried adding a comma, and same result: "a kid in a vampire costume showing his fangs, while emerging from a coffin in a basement". I tested my old prompt and it didn't cause the popup.

[quiet zodiac]

Eh who knows? Use another system. The prompt works fine on mine, using MLOPS.

[sterile lily]

@quiet zodiac Good to know. Thanks.

[fast kite]

hello all, still working on reverse engineering an approach taken by a favorite ai artist of mine. if anyone has any insight on the best way to get similar results (totally different source material of course) i would really appreciate it...
basically i'm trying to figure out what sorts of prompts to use with a model i've trained on many images of the same person. i'd like the images to be as close to the source material as possible, but a bit of a corruption/distortion of them. here's what i'm going for:

https://64.media.tumblr.com/a96e27825920844bd7bc80ecdd9e980b/c59eafb06365ac12-a8/s1280x1920/038ca486869e4950d50a9c0e2b9359a6b4406541.jpg

https://64.media.tumblr.com/b26c3ef6a378578d1ddf8a881da3728f/1d5ba8f3a9157fdf-15/s250x400/17b0ea379139047793ee173841ae4163f56b3e8a.jpg

https://64.media.tumblr.com/aec77b7d65e74c59a8c66202eb3e7d7b/c809ee0f085530e0-c7/s1280x1920/980467b111e620435bcac0fb1e331fb0fa365659.png

https://64.media.tumblr.com/1ca901a493c26d0480b05695a8725940/3f4fe56903c157be-d2/s1280x1920/7faddaba5240452877cb37f6909f41ba6c06c5d6.png

if anyone has any thoughts on the sorts of prompts involved, i'd really appreciate it. i also am not sure if it's possible, but can you over-privilege the source data of the model i've trained? the images i like the most come out of just training the model

[solar dust]

I have 12GB of VRAM, these are my current commands.

If I add something like medvram/lowvram, my images will take more time to be generated

correct?

[torn cliff]

medvram has a slight impact, lowvram has a larger impact, from what I have heard. But I think it said somewhere you should consider using medvram if between 8-16Gb

[warped drift]

is there any models/loras i should know about, and could anybody kind of guide me with the prompt language, im used to midjourney and idrk how to prompt for SD

[rich crow]

sounds like youll want to use controlnet for this https://github.com/lllyasviel/ControlNet-v1-1-nightly

[fast kite]

to clarify, what I've done so far is used a google colab for Dreambooth and gotten decent results using dozens of images of someone producing images that reflect her quite accurately. but what's missing is what I see in the images I linked above - an approximation of the qualities of the sample images. i'd really like a bunch of photos that as closely resemble the sample images as possible. I'm not trying to render myself as superman which is what I've seen a lot of SD tutorials geared towards. what I'm after is more in line with what I've seen from GANs back in 2021ish.

if anyone has any thoughts i'd really appreciate it.

[flat moth]

You’re using control net on the bot? Try also setting an aspect ratio. I was getting blank images with 0 width/height and setting aspect 1:1 solved it.

[warped drift]

im just trying to create variations on simple stuff like this, anything i can use to help

[high leaf]

orcs are not that vested in aestetics

[warped drift]

i have a lot easier time modeling with a ref drawing

and i just wanna be able to create some variations but if i use img to img to all looks too close to the original

[tidal junco]

It fixed after updating and restarting. Thank you all guys! 😄

[warped drift]

.

[gentle lantern]

Does anyone have a prompt for hugging a teddy bear that doesn't result in both the teddy bear and the person facing the camera or teddy bear distortion?

Preferably for SDXL

[solar dust]

Like this?

[next flint]

Like this?

[gentle lantern]

that doesn't result in both the teddy bear and the person facing the camera

So no

That's basically every image. The first one above is technically not facing the same direction, but you can still see both faces. Basically I want someone hugging a teddy bear and ask you can see of the bear is the back.

Search results: https://st2.depositphotos.com/1794440/9209/i/950/depositphotos_92097806-stock-photo-cute-little-girl-hugging-a.jpg https://thumbs.dreamstime.com/b/little-sad-girl-hugging-teddy-bear-home-79586059.jpg https://thumbs.dreamstime.com/b/girl-hugging-her-teddy-bear-12847900.jpg http://image.shutterstock.com/z/stock-photo-beautiful-girl-hugging-a-teddy-bear-116086207.jpg

Not necessarily looking for full-body images or girls or even kids specifically.

I used to sleep with a teddy bear (still do, technically, I use it as a pillow now lol) and I never held it facing the same way as me, if I was sleeping the head was usually on my shoulder next to mine, usually with me facing the other way to avoid a face full of fur. I've got a few pictures of kids sleeping with teddy bears (under the covers works well), but they're usually facing the camera. Putting facing the camera, facing, face in neg hasn't helped much.

[solar dust]

My images are not facing the camera

they're looking away.

Almost half of her face is hidden.

[gentle lantern]

Not sure how else to word it. [In your pictures] she's looking away but she's almost facing the camera.

They don't necessarily have to be completely facing each other, but I don't want to see both faces; best if the bear is facing away from camera entirely and person is not, acceptable if it's a side view of both.

[gentle lantern]

Lol. I mean that's what's happening in the picture above, not what I want.

[atomic wigeon]

Asked and answered. I told you how to do it. I deleted my replies so that you cant use the information. You should have been respectful to the people attempting to help you and said thank you.

[heavy parrot]

Hey there! I remember seeing once a way to prompt for probability using for example different clothing types in the prompt enclosed in brackets and then SD would choose one of those at random of the image. Is this a thing? Do i remember correctly? Do you know how i do that?

[orchid ore]

@heavy parrot not sure i understand properly, don't you think this?
https://github.com/adieyal/sd-dynamic-prompts

[heavy parrot]

Yes thnx! That is exactly what i meant

[orchid ore]

Great 🙂

[heavy parrot]

Do i need an extension for that? I think it was build in

[orchid ore]

i think implemented is pay more attention is some word only. This extension is very feature rich.
(blue pants:1.5) for example is build in a1111 it make it very probably that there will be blue pants.

wildcards i think are not build in

[heavy parrot]

Thnks brother! i ll test it on default and I ll get the extension if it doesnt do what i need

[halcyon goblet]

How do I make thicc girl

[chilly karma]

Hey guys, I've made a lot of images, but most of them end up having fuzzy eyes, I've even used eye loras like detailed_eyes, detailed_skin_and_eyesand more but the eyes still end up being fuzzy

when zooming in there isn't any clear border for the iris

they happen to be melting into the eye

how do I create sharp irises

and no I don't use face restoration

[solar dust]

Use Adetailer.

[atomic wigeon]

blue eyes with black pupils

[chilly karma]

am I supposed to add that to the prompt?

[chilly karma]

I don't know how to, is there any quick way to do it? I end up making them worse

[atomic wigeon]

https://discord.com/channels/1002292111942635562/1011743094309396631 Blue eyes with dilated black pupils

If you need them larger

[solar dust]

Install Adetailer, activate it and gg

it's automatic.

[atomic wigeon]

Or, just use a prompt.

[chilly karma]

before

[chilly karma]

after

[atomic wigeon]

I get flawless results.

It's something else.

[chilly karma]

maybe it's my merge?

[atomic wigeon]

Those are good pupils. My entire prompt is: womans face with blue eyes and dilated black pupils,

[lavish surge]

Hello. How do you guys attempt to write prompt when you don't wish the character to have any hand poses, but there doesn't exist a pocket in their outfit?

[chilly karma]

could it be my gpu?

[atomic wigeon]

No. It could potentially be steps. Or steps and the prompt.

How many steps you using?

[chilly karma]

30

[atomic wigeon]

There's a problem. Detail = higher steps.

[chilly karma]

lemme try with 90 😄

[atomic wigeon]

Prompt order contributes too

[chilly karma]

it's pretty higher in the order

[atomic wigeon]

Priority to the front, secondary at the rear.

[atomic wigeon]

Controlnet

[lavish surge]

Sure. Can controlnet detect anime style images for poses?

[atomic wigeon]

It can detect any object, any character, any animal, any logo, etc. Edge, Pose and Depth are your options to lock down.

[chilly karma]

this is with 150 steps

[lavish surge]

alright

[chilly karma]

a little bit better I suppose

[atomic wigeon]

It's still not great. Man, I can only assume it's your model, lighting, or something else. Tough one.

[chilly karma]

maybe it's the model

lemme try any original model

[atomic wigeon]

Try Pixar style in your prompt. Mechanical reasons.

The reason "Pixar style" is because Pixar is famous for large round soft friendly eyes with well defined pupils.

[chilly karma]

do I need to prompt in here something?

[obtuse torrent]

most of those who use it go with the default setting..
basically that tool is an abbreviated inpaint script with your own templates... the text boxes are useful when what you are looking for in the face strays from what you described in the original prompt ... run a test with the default models...8n / 8s (only face)
I've seen people use the original description and add some... detail, details, sharp, sclera, pupils, iris, detailed (one or two extra words)... never personally tested

[solar dust]

Btw

@chilly karma What's the model you're using?

[chilly karma]

it's a merger b/w a couple of models

anything v4, endless3.5,chilledV2,meinaunrealV3

[solar dust]

Huh?

You made those images by using multiple models at the same time? @chilly karma

[crimson patio]

Guys. Is there some trick to better inpaint high res renders? e.g. I have 1500x1000 (after hires fix) and its really hard to inpaint after that... like 4 hands etc. its about res or promps?

[orchid ore]

@crimson patio shouldnt help, but are you using inpaint model for inpainting?

[crimson patio]

not everymodel have inpaint model

so in this nope

[orchid ore]

question if not better using inpaint model different. But i think you are right

[crimson patio]

so inpaint models can easy good inpaint even in higher res?

[obtuse torrent]

I am not an expert but you must work the prompts
if it's inpaint to fix bugs or modify things check "only masked" and adjust the prompt - or it will try to duplicate every single thing from the original image inside the inpaint zone ...test by modifying the D.S. someone with more experience could cooperate with good parameters for the size of the mask, blur, etc.

[chilly karma]

I merged those models into 1 and then made it 😄

Here I used the 3 ADetailer models

[solar dust]

Could you please send me a link to to the model you use?

[full flint]

Has anyone tried to create characters with cybernetic arms consistently? So far, the success for me varies from model to model, but most tend to be very stubborn. It's also never clear what prompts will work or not. I am usually trying "robotic arms" and sometimes "prosthetic arms". Most often a model can absolutely do them just fine, but ridiculously rarely. Anyone had success?

[gentle lantern]

You told me candid shot, I'm aware of what candid photography is. I'm not interested in whether or not the person is facing/aware of the camera, as I explicitly mentioned above. I'm sorry if I came off as rude, I'm frustrated by help that isn't what I want (possibly because I can't explain what I want well enough). I do appreciate the effort, though.

[maiden geyser]

I am not sure where to put this, but it does involve prompting help. If this is the wrong place, please, admins, let me know if there is an appropriate place.

Stable Diffusion Mastery - course for beginner prompters from paid prompt engineers 🙂

As I come up on my 30th year as a teacher, I have noticed that a number of people have trouble using Stable Diffusion, even though it is the industry standard for professional AI image generation (I am a paid prompt engineer). After posting in a number of places and finding there was great interest, I have gotten together with two other AI professionals and we have created an inexpensive and complete course for you to learn how to use this software and its many features.

The course is $80, and consists of 6 one-hour Zoom sessions which can be reviewed as many times as you like on a private Discord server, in addition to being able to ask questions of instructors during the instruction time and between sessions. At times the class might go over the one-hour class time. This is a cost of less than $15 per session. All three instructors work in the AI industry and have different topics of specialty, as described below:

Instructors:

• Harris Terry - 30-year educator from USA, works as a prompt engineer with a startup and has given lectures and masterclasses on various AI topics including Midjourney with more than 60 attendees.
  https://aidreams.tech/

• Geoffrey Mollet - Prompt engineer, web developer, and AI social media specialist from France with videos with 38m+ views.
  https://linktr.ee/singularitydiffusion

• Tanvir Hafiz - YouTube AI educator and tech specialist from Bangladesh, AI freelancer for corporate clients and video production.
  https://www.youtube.com/@TanvirsTechTalk

This course will be Sundays at Noon EST, starting August 27th via Zoom.
The dates are:
Aug 27
Sept 3, 10, 17, 24
Oct 1

Course Summary
Session 1 - Basics

• txt2img settings, models, prompting, styles, consistent characters, upscaling

Session 2 - Images and Inpainting

• img2img settings, inpainting basics, inpainting for upscaling, Adetailer use

Session 3 - Deforum

• installation, basic features, scheduler use and frames, movement controls, prompts, Init and Final tab settings

Session 4 - ControlNet

• installation, model download, basics of UI, types of models & uses, usage in txt2img and img2img

Session 5 - Textual Inversions & LORAs

• basics of Textural Inversions & Negative Embeddings, LORAs and various uses (characters/styles/hair/accents)

Session 6 - Additional Extensions

• installation and use of:
• ROOP - allows face swapping
• Inpaint Anything - faster inpainting alternative
• Tiled Diffusion / Multi region prompt - different prompts for different areas of the screen
• Tag Autocomplete - see popular terms and use wildcards
• Cutoff - allow separation of colors and descriptors by using commas

If for some reason a class is postponed due to an unexpected emergency or technical difficulty, the following Sunday will continue where the last session left off.

Upon sign up, you will be emailed a single-use link to join the class Discord server, and will receive a Zoom link for the class before the first session. Sessions will be recorded and available via Discord for members. There are no refunds for this course, and distribution of any Discord or Zoom links will get you removed from the course without refund immediately.

You are expected to have some sort of Stable Diffusion software (A1111, SDNext) installed before sessions begin if you want to follow along and try what you are watching. The Discord server and YouTube have install videos to walk you through the process, and the instructors are happy to help you during non-lecture time. By signing up, you agree not to distribute or share videos or materials from the course.

Let's become masters at Stable Diffusion!

SIGN UP HERE
https://sowl.co/s/1HsPX

I hope to help as many people as possible get good at prompting. Getting what you want rather than doing 100 iterations can be much more rewarding.

Additionally, this is not some money-making scheme, as all three educators make much more money with their AI jobs or other work (myself a 30-year music instructor). Rather, it is a way for us to give back to the SD community and help beginners get going with this software and what is has to offer (see the topics above).

Thanks for reading!

[high leaf]

hi, comfy users: what is the correct way to XY plot loras? I feel like the results are not the same if i plug loras in the XY versus the prompt versus lora stacker

[quiet zodiac]

@ᏦᎥᏒᏗᎩᏗ Try using close up or detailed face. Get the camera closer and the hands won't matter.

[keen perch]

Been trying to create more photos like this, particularly of a space station corridor in this high contrast, dark style. Any ideas?

[keen perch]

For now I'm experimenting with black and white outputs only, and processing through an image editor myself.

[chilly karma]

it's something I made merging the models I mentioned

I would need to upload it

[magic badger]

Any ideas on how I could fix the issues I'm having with the bowstring and arrows? 🤔

[buoyant thicket]

I am trying to get art similar to this but in greater quality for my game but I cant get any good results. Does anyone have any tips?

[magic badger]

I think the solution is just spamming keywords like 'archer, bowman, hunter' 😅

[magic badger]

You could try some kind of combination of top-down, god-mode-perspective, cavalry, mounted-knight, horse, warrior, spearman, swordman, infantry, pixel-art, sprite, video game, 16x16, 32x32, retro

Don't think I've ever seen anyone use SD for game sprites before. 🤔

You might get best results from setting the image size to precisely 32 or 64px, depending on your preference 🤔

[buoyant thicket]

thanks

[gentle lantern]

If you're using A1111, there's a pixel art extension, if that's what you're looking for.

[buoyant thicket]

im not necisarily after pixel art. Just the top down angle

[atomic wigeon]

Look up at all the recent questions. Doesn't it feel like they're all directed at anime porn?

In the future I will not be answering any questions pertaining to anime.

[proven jasper]

but i will

[atomic wigeon]

That's a sigh of relief.

[orchid ore]

@buoyant thicket what model and resolution are you using?

[atomic wigeon]

Prompting tip:

^ The size of your image will orientate your character (standing, sitting, laying). Add the word chair for example, and a pose, and the AI will build a scene around that character posing in a chair.

[buoyant thicket]

I have just been using the discord because my computer isnt powerful enough to run it locally

[orchid ore]

o.k. @buoyant thicket

[quiet zodiac]

Does anyone know of a model that doesn't make oversized women's breasts?
I've tried prompting them away, both positive and negative. Across several models, they can only produce these ample bosoms.

[lavish surge]

In my experience, just specify the size as medium or small.

[chilly karma]

@solar dust I tried different models and got no well defined border specially for eyes

[magic badger]

@quiet zodiac You can see the piece I've been working on in #1019361238234443776 today. I specified an elf, young adult, female, feminine and got a pretty conservative bust

[lavish surge]

Is it that the way to hunt seed is to mass making images without highres?

[solar dust]

You can still use Adetailer.

[chilly karma]

ya I used that too, @solar dust here

I gotta try it on a different machine

can you try it on yours?

you must have endless_renatus

[quiet zodiac]

I have specified (((small breasts))) but the models ignore the prompting.
I feel like it's a built-in part of the model. That's why I wonder if there is a mdel that excells at making regular women and not fashion models.

[chilly karma]

try (small breasts : 1.8)

[quiet zodiac]

It doesn't have any effect, I tried adding weighting as well as parenthesis. I really have tried everything, including breast implants in my negative.

[chilly karma]

then you must use a lora, it's in the model, it can't me it any smaller

[quiet zodiac]

Thanks for the tip. I discovered you have to place it very first in your negative list, then it starts working.

[hollow igloo]

Guys do u know how is he doing this and what model is he using

[pine mantle]

Hi guide please I need help refining a prompt concerning mobile app icon

[atomic wigeon]

"Small upper chest" works well. I use it sometimes in Playground because they have filters like you're describing.

[tired vigil]

Hey how do I avoid unnecessary body parts

Like four legs or body parts randomly being at places they shouldn't

[atomic wigeon]

1. A good negative prompt
2. The size of your image matters. AI fill.
3. SDXL 1.0 -vs- SD 1.5 = huge difference.

[tired vigil]

Which one do you recommend?

[atomic wigeon]

SDXL 1.0

[tired vigil]

What does a good negative promt look like

[tired vigil]

Ok

[quiet zodiac]

Kind of tricky, but specifying a pose can help, try "casual pose" or "action pose" or "natural pose". Sometimes duplicates are the result of choosing the wrong resolution. Stick to common values, like 512x1024 for portrait or 768x768. Also give another model a try.

[atomic wigeon]

I don't know if this is "good", but it's the only one I've got. Some of you guys probably have better.

poorly rendered face, poor facial details, poorly drawn hands, low resolution, images cut out, bad composition, mutated body parts, blurry image, disfigured, oversaturated, bad anatomy, deformed body features, out of frame, text, error, cropped, worst quality, low quality, jpeg artifacts, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, used fingers, too many fingers, long neck, watermark, signature, blurry, bad anatomy, extra limbs, poorly drawn face, poorly drawn hands, missing fingers, letters, numbers, fonts, text, words, symbols, autographs, nsfw, pink

[tired vigil]

Thanks

[atomic wigeon]

Disclaimer: That has been my neg prompt since SD 1.5. I think it needs updating to better coincide with SDXL 1.0.

[torn cliff]

Anyone know good prompts for portraits and for chest up shots?

[high leaf]

hello

trying to XY plot loras. what's the correct way?

[pulsar isle]

hello what should i use for anime pictures in sdxl comfyui

[high leaf]

what's the question? you can gg for prompts

[atomic wigeon]

Medium shot is a waist up portrait shot, and the most common. Closeup is a shoulders up shot.

[torn cliff]

Thanks. I tried just "Portrait" but it gave me a shot laying on a bed...

[atomic wigeon]

Portrait is universal pose term for a painting.

Medium shot is what you're looking for.

[torn cliff]

Ahhh crap... Image froze again... v.v

[atomic wigeon]

medium shot chair

^ pretty obvious what that will do. Can also include face pose or random face pose.

[torn cliff]

I'll try that when I start it up again... Sometimes when using SD Next when I try img2img it just freezes up for some reason and I have to close the whole thing and reopen it. Nothing shows in the logs when it does that either...

[slow cedar]

Any tricks or tips for making good prompts related to the next images?

I've been experimenting with different prompts, but sometimes just two extra words seem to disrupt the entire meaning or, at least, how SD interprets it. Many times when I encounter such alterations, I wonder if it has to do with how I structure my prompts.

Take this image for example: My intention was to create a female character in a medieval world with a 3D anime style, similar to the recent Final Fantasy games. When I generate the image using the prompt without specifying the full body, it produces a rather beautiful image that closely resembles the original idea.

Prompt used: female adult cute face with light pink hair wearing medieval armor, anime drawing style, medieval fantasy city background

I haven't used any negative prompts or textual inversion for this ones

My intention was to generate a full-body image. However, as soon as I add the words 'full body' to the prompt, the character's quality deteriorates significantly, and the face becomes rather unpleasant compared to the other image

[torn cliff]

May be because of the image size

[slow cedar]

Both have the same size, they get upscaled later

What do you suggest?

[torn cliff]

If you have a close up shot and a full body shot the quality if of course going to differ

I mean if you have them both using the same size.

[slow cedar]

You're right

[torn cliff]

Square sizes like 512x512 tend to work better for headshots, while rectangles like 512x768 tend to work better for standing shots.

[slow cedar]

But to expect such quality loss when streching the charcater? The face is really badly done in the full body one

[torn cliff]

Mind you thats just example numbers that are easy for me to remember.

You could try to run it through img2img again with just the same prompts, or use a more detail thing to help

[slow cedar]

by just making the rendered image resoultion bigger?

[torn cliff]

Not even a need to make it bigger.

This is what I mean https://civitai.com/models/82098/add-more-details-detail-enhancer-tweaker-lora

[atomic wigeon]

Further from the camera = more steps needed.

[slow cedar]

Thanks for the resource, I'll check it out later

[slow cedar]

Damn, even cranking it to 70 steps doesn't performs well...

[atomic wigeon]

You might wanna put some underwear on her.

[torn cliff]

Air conditioning...

Can also try inpainting with a mask over the face, possibly changing prompts to focus on the face alone

[slow cedar]

Dont worry, at 230 steps the skirt goes down again aaaand... the face doesn't fix. Or at least, have the 3D anime style I wanted for it (like the half body image)

[torn cliff]

Next time tell the AI to take the picture when the wind stops blowing... 😛

[slow cedar]

I might try that, I should save the pics without scaling first, cos my gpu wants to kill itself when rendering a 1024 x 2048 image

[torn cliff]

Yeah

[atomic wigeon]

I thought max cap for SDXL 1.0 was 150 steps?

[slow cedar]

Not using SDXL ;)

Altough, I'm pretty sure that you can surpass that limit, but wont generate any more detail

[atomic wigeon]

Mind sharing your prompt? I want to test it.

[slow cedar]

here it is

Noting complicated, but I think maybe its the way the prompt is written? I don't know

Maybe its the model as well, but its really weird that both pics came from the same model and have such a big difference in the face regions only because I added "full body" to the prompt

[slow cedar]

forgot to say it, the 2nd image prompt is full body female adult cute face with light pink hair wearing medieval armor, anime drawing style, medieval fantasy city background

[atomic wigeon]

ReVAnimated at 150 steps. Bad hands.

[slow cedar]

Altough the faces are really sharp, they do look like anime ones, not like mines

I'm using DreamShaper v8

[atomic wigeon]

Spin #2 pure prompt ReVAnimated (SD 1.5)

[torn cliff]

Ahh didn't know there were new versions of DreamShaper, think mine is 6

[slow cedar]

Could you try to make at 640 x 1024 dimensions?

[atomic wigeon]

Raw SDXL 1.0

[slow cedar]

I might have to download RevAnimated

[atomic wigeon]

Yes. Next spin will be 768 x 1344

ReVAnimated isn't a download. It's a filter at playground

[slow cedar]

wow, what an awesome raeson to dowload SDXL

So this might be the models fault and that's it? Its a shame because the half body pic looks exactly as I would want for a full body image

[atomic wigeon]

Spinning 640 x 1024. Waiting

[torn cliff]

No RevAnimated is a model on Civitai too

https://civitai.com/models/7371?modelVersionId=46846

[slow cedar]

ik ik, I just never downloaded it

[atomic wigeon]

Oh? Didn't know that

[slow cedar]

Is there any difference from hf? I don't think so

[atomic wigeon]

Soemtimes ReVAnimated actually makes nicer stuff than SDXL 1.0.

[slow cedar]

Then I guess this kind of generation is the models fault only...

[atomic wigeon]

That's SDXL 1.0 default at 640 x 1024

[slow cedar]

I really like this cartoon style more than the anime one, but it triggered with the 'anime style' prompt right?

[atomic wigeon]

That was your exact prompt. I jsut rebuilt it to test something.

[slow cedar]

Alright

So I'm guessing that the DreamShaper model just reached its max capabilities and doesnt have the full body version of that type of face I'm looking for

[atomic wigeon]

It was a super simple prompt. It jsut needs to be dialed in with details.

3d concept art, female, medieval armor, pink spotlight, Final Fantasy style,

[slow cedar]

My main goal for this is to have a face like in this pic

In a full body armor with the same style

[atomic wigeon]

Your pic failed

[slow cedar]

Yeah, and the detail enhancer tweaker lora doesnt help at all

I might have to outpaint the image or something, I don't really know what to do here

[slow cedar]

Damn, this ones are kinda what I'm looking for, what did you use?

[atomic wigeon]

3d concept art, female, medieval armor, pink spotlight, Final Fantasy style, medieval fantasy city background,

That's REVAnimated. Free at playgroundAI.com

10k resolution

[slow cedar]

10K? holy shit