📝｜prompting-help

[slender pasture]

SDXL

I uninstalled whatever was before that then reinstalled all of stable with XL

huh ok ❤️ well I appreciate everything anyways thank you n_n

i did find the sdxl canny, just need to know what folder to boop it into

[rich crow]

i hope someone made an sdxl controlnet extension for a1. from my limited knowledge, it was only on comfy (at least at the time) that people used the controlnet. there's no need for an extension on comfy, just need to plug the controlnet node in and select the sdxl canny model

[slender pasture]

Talking about this stuff sounds like gibberish to me tbh. lmfao Im lucky I can use SD at all, lmao

[silver valley]

I linked you the correct models yesterday. The 700mb files.
They go into models/controlnet folder.
Link here: #🤝｜tech-support message

[slender pasture]

oh oopsie thank you~ ❤️

Maybe id missed it

I did!

thank you ❤️

CS1o to the rescue yet again!

Got it working!
Thanks n_n

[silver valley]

No problem 🙂

[quiet raft]

I am testing some "real" models and I tend to get "fuckt up" faces.

Any recommendation or suggestions?

[fallow rock]

Can anyone help me solve it? I'm trying to use animate diff

[red hemlock]

does anybody have a working workflow for sd3 that can do a photography portrait ?

[spiral jungle]

What are you looking for specifically? Do you want e.g. openpose control to get a specific positioning? If not, you can probably get away with writing "portrait" in the prompt

[red hemlock]

i would like an example workflow that i could drag&drop into my comfyui that will make a portrait photography

(not the prompt but the whole workflow)

[harsh plover]

lLLllMLm:pkojoijp)u()jlml<<mlM.M. K/PI'L LL;L;L

'L';L';L'L'LPKP'; L

[twin yoke]

You okay, bro?

[harsh plover]

Yes

[zenith echo]

你好

[reef axle]

I'm creating coloring books, mostly cars and with one item I can't deal with stable difusion makes my tires white, it makes them black or heavily gray, I'd like to get them white with an outline that you know it's a tire.

[frank ivy]

Does anyone know how to make disney pixar poster style images from a base image?

[nocturne gale]

I really seem to struggle with crafting prompts

[nocturne gale]

ugh I add like a few things to my prompt and now everything is completely screwed like wtf

[remote yacht]

Hi guys, anyone know how to prompt a slight belly pooch? Anything including fat/chubby leads to morbid obesity and belly/stomach/navel leads to perfectly flat abs.

[versed walrus]

Hello, I'm trying to use 3D renders of a character to train LORA of that character. Unfortunately the results are becoming this weird half-3D half-2D mess.
Is there a trick to captioning the training images when using 3D renders? I tried writing "3D render of X character" instead of "X character" in the captions, but it didn't seem to help much.

[vital remnant]

When you're starting a new piece, what are some keywords you always add to begin with? Stuff like 'best quality, highly detailed' etc?

[frigid kernel]

Hello everyone, can anyone suggest me why my generation of mask is not applying on the left side of the image. I am keeping the dimensions - 1016 x 504 but there is this left black patch that is coming again and again

[atomic wigeon]

I'm speculating but it looks almost like a 1920x1080 issue. Or 1440 maybe. Strange.

[atomic wigeon]

For me personally I like to start with 3D and hyper realistic. That's where the majority of my work takes me. But it's very specific to your art style and the project you're currently on. Words toward the front of the prompt take priority over everything else. So if you need more of something, move that word closer to the front of the prompt.

[atomic wigeon]

Pregnancy. Describe how far along she (or even he) is. Beer belly is another.

[atomic wigeon]

Add "Disney" and/or "Pixar" "3D" to your prompt. Cartoon style or hyper realistic. Then describe the lighting and scene. If you want 2D try to use "sticker".

[crystal gate]

Hi, I've been lurking for a while but this is my first post. I'm trying to recreate an album cover, is this the right place to ask for help?

[atomic wigeon]

Hi. What specifically do you need help with? Are you trying image to image or purely using a prompt?

[crystal gate]

Just with a prompt but I'm doing it via a companion AI and that limits me to 500 characters.

[atomic wigeon]

Try "album art", "movie poster style" and a more specific design aspect. For example you want a contempory or modernistic and minimalistic scene using vibrant colors that are centered and symmetrical.

If you know of a band, for example Pink Floyd, you can say "pink floyd style". Works with any band.

[crystal gate]

Thanks for your help. I'm trying to recreate this.
I have the yellow tones but I'm struggling to get two people sitting across from each other symmetrically.

[atomic wigeon]

"sitting at a table passionately conversing with each other". Or use a reaction from an emotion. For example, "mother telling daughter she's dying". The AI will illicit a response tied to the emotion. Same goes for common human activities that require eye contact.

[hazy kiln]

Hey guys. I'm using "AUTOMATIC1111/stable-diffusion-webui" with Pinokio 1.3.4 on Ubuntu, and I'm able to generate images, but I need photorealism and it's not.
I tried the exact same prompt as in prompthero examples, with same parameters, and it's totally different.
Am I missing something ? I use v1-5-pruned.safetensors as most people on prompthero

[crystal gate]

Thanks very much, I'll try that. Much appreciated!

[atomic wigeon]

Use "hyper realistic" and describe a camera and lens. Try this prompt to see what I mean.

Extreme close-up portrait of a boy embracing a bear, utilizing the rule of thirds, with engaging, friendly eyes directly gazing into the camera, set against a stylized interior background, melancholic ambiance infused with grunge elements, ambient fog embracing the scene, executed in 3D with subtle motion blur and a 25-degree tilt-shift perspective, captured in the essence of Peter Jackson's cinematographic style, lensed through a Sigma fp and Sigma 45mm f/2.8

That same lens works great for any realistic subject matter. Edit: the below image original was amazing. But this pixelated one is the result from pulling it from a social media site.

[hazy kiln]

Yeah that's what it give me 😅 I don't understand

[atomic wigeon]

Oh man. Is that sdxl?

Increase your samples.

[hazy kiln]

Nah 1.5. Top left corner

SDXL gives the same result

Basically each time I try to copy a prompt, the result is totally unexpected

[hazy kiln]

I can set it to 100 the issue remains the same

[atomic wigeon]

That's really strange man. It's not the prompt. I promise you that. It's gotta be a setting. Try and ask in: https://discord.com/channels/1002292111942635562/1004159122335354970

They can help you far better that I can. Sorry.

[hazy kiln]

I know it's not the prompt, as I said I tried to reproduce many images by copying prompts online

[hazy kiln]

Ok thanks

[atomic wigeon]

Alternatively I would recommend trying the prompt on a cloud site like Playground. It's free. Then import the result to image to image. See if it still happens when you prompt changes to the image. That's if you're in a crunch. Until you figure it out. Good luck!

[atomic wigeon]

You're very welcome. 🙂

[carmine spear]

Anyone got some nice ways of avoiding generating this little onlooker guy? Currently "deserted, abandoned, desolate" are in the positive and "person, figure, man, woman, robot, onlooker" are in the negative

[atomic wigeon]

Maybe you could describe the landscape after the fall of mankind. I never tried because I haven't had that issue. But if you can describe "after" the apocalypse" it might work. Tough one. Good luck.

[carmine spear]

I'll try experimenting with natural language down that kind of route, thanks for the suggestion

[plush flume]

hello Im trying to remove speech bubbles from images. Is there a way to do this with in-paint and what prompt it might be? Thank you and have a nice day

[atomic wigeon]

Are you making comics? There are a ton of art software programs that can remove them. Playground is a cloud based AI art website that has a very powerful image editor that can remove those with a simple brush stroke. Other than that you'll need to share your prompt and I can advise after.

[wispy tapir]

Anyone knows how to create Brazil phonk style pictures?

[trim path]

I tried (with the help of gemini), i will share it with you, tell me if its similar

model: sd3 medium
prompt: Dark anime edit: Extreme darkness. Red-black shadows swirl behind a menacing, completely dark figure. Only glowing red eyes and red outlines on its edges pierce the void. Speed lines and a digital glitch effect distort the scene. Deep blacks and neon red dominate the palette. It has to be slightly blurry
Negative Prompt: deformed, distorted, disfigured, poorly drawn, bad anatomy, incorrect anatomy, extra limb, missing limb, floating limbs, mutated hands and fingers, disconnected limbs, mutation, mutated, ugly, disgusting, amputation
guidance scale: 7.5

The results are attached. Just tried helping

[quiet raft]

Should I use this when doing Upscale on Img 2 Img ?

[valid violet]

• Im working on a project to make self evolving images. so the prompt i wrote to make these... I havent been able to modify it to get a similar result yet.
• I made some interesting evolving images from doing a feedback loop(resending the output back to the generator over and over)
  -1st image I generated seperatly then uploaded it and sent to stablediffusionXL and then the corresponding images are in order and you can see things start to evolve. its gets more and more complex as long as i use the same prompt

[quiet raft]

Hey guys.. If I have an character that I've made and I wanna try and keep that pose and clothing as much as possible.

What Controlnet is it that I should use?

Is it IP-Adapter?

[twin yoke]

If you're keeping the same character in the same pose and same clothing, that seems like you're just trying to change the background and other elements. You're probably better off sticking with inpainting if that's the case.

[quiet raft]

NAh more like, wanna try to keep the pose and such very similar but change the pose a bit.

[twin yoke]

The only way to guarantee that you can stay as close as possible is to train a LoRA. Outside of that, IPAdapter+ along with ControlNet can get you close, but it's not a guarantee. The more you work on it, the closer you can get, though. You might be able to retain the elements you want, but that can take some time if you don't get lucky. It's really a matter of balance between expectations vs. results. The more flexible you are with your expectations, the more likely you will receive results that are satisfactory, but the more specific your expectations, the less satisfied you'll likely be.

[quiet zodiac]

I am crudely blending two prompts together by surrounding each one in parenthesis and giving them weights. They're concatenated with the word AND. Is there a better way to emphasis one part of a prompt over another?

(A renaissance fair set in a giant,  ancient tree whose branches are home to a thriving community of tiny humanoid creatures,:0.863636) AND (A serene village on the edge of a mirror-like lake,  with biomes mimicking Earth's environments,:0.136364)

[atomic wigeon]

Whichever words you place closest to the start of the prompt will prioritize the results.

So if my prompt is 1, 2, 3 , 4, 5 the emphasis will be on #1 and least on #5.

Then the words themselves matter. By you using the word "giant" near the front of the prompt you risk some results displaying a gaint at a fair.

It's a constant struggle for all of us. That's the art within the art.

[hollow echo]

Hi everyone, I'm learning stable diffusion, I'm trying to change the background of a photo, but the results was not as expected!

Can you give me advice?

[silver valley]

Hey dont prompt for a girl if you only want to generate the background

Then check your preview image if the mask is good

[hollow echo]

I tried it, but it wasn't working as expected @silver valley

[night ember]

can anyone help me?For some reason all the loras i put in the folder without 2 are not showing up does anybody know why?

[edgy river]

Hello everyone, everything good? I'm having a problem that I can't solve or find a solution for. In certain renders, especially larger ones, the reconnecting box appears and freezes completely. It doesn't show any errors in the command prompt, and as soon as I type any key it simply closes. Can someone help me?

[twin yoke]

You can't say "no girl" in the positive prompt. The encoding doesn't work that way. Literally every word in the positive prompt box is taken as a positive to be included in your image. Your positive prompt should only be a description of the background you are trying to create.

[silver valley]

"No girl" in positive prompt is wrong.
Either remove no girl completly or add girl to the negative prompt

[silver valley]

What's your GPU? Maybe the workflows is to demanding
(Needs to much resources)

[edgy river]

4090 😦 shouldnt be a problem. I've rendered more complex workflows. It started recently. I'm desperate, have to get this job done by tomorrow, but cant find a solution.

[twin yoke]

If you're seeing "Reconnecting...", the web service for ComfyUI has either died or become unresponsive. Since you said there's no errors in the console and you hit a button and it closes, that leads me to believe that ComfyUI died before you got to that point. If you look through the console on startup of ComfyUI, you should read every line to see if there is a node pack that has issues.

If nothing shows up, then you need to try a fresh, basic workflow to see if you can generate anything. If you can, then you likely have a node pack that's causing issues. Typically, at that point, I start by disabling half of my installed nodes (don't uninstall, just disable via Manager) and try again. If it still fails, then I again disable half. If it works, then the broken pack is in what you disabled; flip the disables and do it again until you've narrowed it down.

[edgy river]

Thank you very much. I'll do some tests. The strange thing is that I can generate smaller renders using the same workflow. But when I try to render the entire section, it breaks.

[twin yoke]

If you can do smaller renders, then that's an indicator that you may have a node that can't handle what you're trying to pass in. You might consider doing groups of bypasses to see if you can isolate which node is having the issue.

If you can see where it's breaking and if it's consistent, it's probably not the highlighted node when it crashes and is likely the node right after that.

[edgy river]

I think it is the highres fix. It always crashs on in

[twin yoke]

See my last post.

Though highres is possible because that's a multi-step node.

[edgy river]

thas awesome! Never thoght about it

[twin yoke]

But I think you can take some steps here that I've outlined to see if you can isolate it. Also, if you do find that there's a specific node that's bombing out, you can probably accomplish the same task in other ways. There are tons of node packs out there and lots of ways to skin a cat.

[edgy river]

Thank You man!! Realy apprciete it

[twin yoke]

Best of luck.

[vernal ice]

Looking to train an anime style LORA for Pony XL or Autism Mix and was wondering what are the best practices when it comes to manually tagging a dataset, such as using natural language, tagged, or a mix of the two.

So far I did test runs with "Natural Language" and got decent results giving long descriptions. 30 images trained for testing, but the dataset includes 70 images.

[stable valley]

what prompts could yo uuse to get multiple images of same character in 1 image, like a character sheet? I tried loras but they are anime based so a bit shit

[open solstice]

What tools can I use to keep my character's appearance other than using Lora.
Can you recommend a dipfake or an analog
anime/2d style

[silver valley]

You can use Controlnet openpose for that there are character sheet style poses available online

[silver valley]

Controlnet IP-Adapter can do this

[acoustic anchor]

test

[worthy crypt]

how to get

this on swarm?

[worthy crypt]

Failed to download CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors from https://huggingface.co/h94/IP-Adapter/resolve/main/models/image_encoder/model.safetensors: One or more errors occurred. (Unable to read data from the transport connection: An existing connection was forcibly closed by the remote host..)

[silver valley]

Controlnet is an extension for thats also available for comfyui and Swarm

[vernal ice]

Gonna bump this since I still haven't been able to get much input

[pallid vault]

how many images in your dataset for lora training? manual tagging would give more control. But personally I just upload to civitai and let it autotag.

the last lora I did manually remove some tags after auto-tagging on civitai

[quiet raft]

I wanna know a bit more as I still cannot get it to work 100% or well good.

If I find an image that I wanna convert to another style.. So I do

1. Put the image into Img2Img and use prompt helper
2. Put it into Txt2Img ( I don't wanna use Img2Img )
3. I activate controlnet and add the image there
4. What Control type is best to use? Is it IP-Adapter or NormalMap?

[vernal ice]

I have 70 images, I manually described 30 images for a small test Lora. I get decent results with medium to long descriptions - but not to my liking.

Might be a bit heavy for a discord chat, but here are some examples. Currently 30 images where described with natural language as follows:

Natural Language
Artstyle-Here, An anime-style girl with short blue hair and bangs and piercing blue eyes, exuding elegance and strength. She wears a sophisticated white dress with long gloves ending in blue cuffs. The dress features intricate blue and gold accents, ending in white frills just above the thigh, with prominent blue gems at the neckline and stomach. A flowing blue cape with ornate patterns complements her outfit. She holds an elegant blue sword with an intricate golden hilt in her right hand. Her outfit includes thigh-high blue boots with white laces on the leg closest to the viewer and a white thigh-high stocking on her left leg, ending in a blue high heel. Her headpiece resembles a white bonnet adorned with blue and white feathers, enhancing her regal appearance, with a golden ribbon trailing on the ground behind her. The character stands poised and confident, with a golden halo-like ring behind her head. The background is white, and the ground is slightly reflective. A full body view of the character looking at the viewer.

Mostly Tagged
Artstyle-Here, anime girl with short blue hair, bangs, and blue eyes. Wearing a white high dress that ends in a v shaped bra. White frills, Intricate blue and gold accents, blue gem on stomach and neckline. Blue choker, long blue gloves, flowing blue cape with ornate patterns and a trailing golden ribbon. Holding a sword with a blue blade and a intracate golden hilt. Thigh-high blue boot with white laces on one leg and thigh-high white stockings ending in a blue high heel in the other, exposed thigh. White and blue bonnet adorned with white feathers. Confident pose, elegant, golden halo-like ring of dots behind her head, white background, reflective ground, full-body view, character looking at the viewer.

Natural + Tagged
Artstyle-Here, an anime girl with blue eyes and short blue hair standing confidently in a white dress with a blue cape and blue gloves carrying a sword, elegant look, gentle expression, thigh high boots and stockings. Frilled dress, white laced boots and blue high heels, blue sword blade, golden hilt, blue bonnet with a white underside and white feathers, blue choker, white background, golden ribbon flowing behind, golden halo, reflective ground, full body view, character looking at viewer.

[pallid vault]

Less is more (less words means less linked to weight during training required, so probably faster results with less words). Also better for prompting later on, as you need to type less words into the prompt. Also, most current models cannot understand sentences anyway (maybe sd3 would)

[wintry topaz]

Want to create bad quality, cell phone like photo, drive by photo of fake UAP / UFO or distant Aerial Objects with A1111. I actually used v1-5-pruned.ckpt (SD 1.5) LORA: instant Photo, Eula A, Sampling 20, CFG 4-6. Prompts:
year 1990, little city, a photo of an (distant:1.3) (oval:1.2) unknown flying object in the air, looks like an flying alien disc (in the sky:1.3), (flying saucer:1.4),photo focus on raw, military thrusters, mechanical AND symmetric object, rimlight, dramatic clouds, behind the clouds, (far away:1.2), disc shaped form, dusty air, blurry lights, surface scratches, sun bleeched, smudges, rugged, used, ((distant object)), (raw quality cellphone photo), <lora:InstantPhotoX3:1>, drive by photo, motion blur,

neg:
<bad_prompt_version2_neg:0.95>, by <fastnegativev2:0.45>, tiling,out of frame,cropped, worst quality, painting, drawing, crayon, sketch, graphite, impressionist, cartoon, anime, JPEG artifacts, airbrushed, cross-eyed, low quality, 3D render, noisy, blurry, soft, deformed, ugly, lowres, low details, semi-realistic, CGI, render, Blender, digital art, manga, amateur, mutilated, distorted, bad anatomy, comics, drawing, crayon, sketch, graphite, impressionist, cartoon, anime, distorted, ballon

Do you have some suggestions for prompting? My actual Results are mostly too large in the center of the generated image, not very bad but too offensive. Thanks

• I got little better results with DPM++ 2M and Aerial Style i modified a bit.

[vernal ice]

Does this apply to Pony XL though, I know most models it does but a lot of users mention how it was trained on natural language and you should use sentences instead of tags?

[quiet raft]

Hey @silver valley , you use img2img upscale option a lot right? What is the best Tile overlap and what is really the change going up or down? Beside it takes longer and it like overlaps morE?

[pallid vault]

the more words you will use, the more the model will forget about it's base. So therefore, choice your words carefully. Maybe even only use unique words (trigger words)

[supple finch]

Hey, I'm getting an issue of my recreation using png info from a civitai example coming out just looking underdeveloped? Any ideas?

[slow mist]

send me the link to the original picture, and show me what you have under the picture after genration

this thing

all params must be same steps, sampler, schdeuler, cfg, seed dimensions, models, loras, promp and GENERATOR

[supple finch]

Original: https://civitai.com/images/13828448

Info: score_9, score_8_up, score_7_up, BREAK, cute creature, furry, adorable, puddle, forest, portrait, highly detailed, detailed skin, depth of field, film grain
Negative prompt: score_1, score_2, score_3, text
Steps: 30, Sampler: DPM++ SDE, Schedule type: Karras, CFG scale: 6, Seed: 2513856529, Size: 768x1280, Model hash: 571029a6ac, Model: ponyRealism_v21Lightning8SVAE, VAE hash: 63aeecb90f, VAE: sdxl_vae.safetensors, Clip skip: 2, Downcast alphas_cumprod: True, Version: v1.9.4

[slow mist]

if they are using NVIDIA and you have AMD, they will never be the same

[supple finch]

That's a very interesting point, I wonder what GPU they're using... I'm on nvidia but that's arbitraty in relation to them

[slow mist]

when i started using sd last year i wanted to kill myself

examples were made on nvidia and i am using mac

than i tried on my sons pc and i get almost identical image

are you using the same version?

your model hash looks diffrent

that is your issue

you are not using the same model

images are generated with main

you are using ponyRealism_v21Lightning8SVAE

[supple finch]

I'm amazed that you've been able to get it to work on a mac!

[supple finch]

Ohhhh you're right, I'm using lightning 8s + vae

[slow mist]

it works much faster on my m3 pro, than on my sons 1660

[supple finch]

That's incredible - I guess I really shouldn't be comparing my 2018 mbp

[slow mist]

do not try 🙂

i can make it work there if you want, but do not complain about speed 🙂

i made it work on iMac 2015 😂

did you get nearly same results with same model?

[supple finch]

That's astounding

[supple finch]

Much more similar, but still similar patchwork effect if you look close, and less developed facial features. I wonder if the example image was post processed

Though it is one of the most downloaded models on civitai, so I can't imagine they're faking it?

[supple finch]

These are the sort of artefacts I'm still getting. I'm following all the recommended inputs from the creator at https://civitai.com/models/372465/pony-realism?modelVersionId=534642

My input: Positive Prompt
score_9, score_8_up, score_7_up, BREAK, male, strong, leather jacket, sitting on a motorcycle, city, depth of field, film grain
Negative Prompt
score_1, score_2, score_3, text
Steps:
30
Sampler:
DPM++ 2S a
Schedule type:
Karras
CFG scale:
6.5
Seed:
1663570276
Size:
1280x768
Model hash:
571029a6ac
Model:
ponyRealism_v21Lightning8SVAE
VAE hash:
63aeecb90f
VAE:
sdxl_vae.safetensors
Clip skip:
2
Downcast alphas_cumprod:
True
Version:
v1.9.4

[slow mist]

find images created with ponyRealism_v21Lightning8SVAE and see parameters for them

[proven jasper]

try to change the sampler

[supple finch]

fair call

[slow mist]

this one for example

https://civitai.com/images/14370208

[blazing path]

Hey guys... i cant get the burger and elemts shown inside the frame.. no matter what kind of prompt i try to use:

any ideas how i could improve the prompt to not have the burger, decorations, whatever extend the frame? :/ thank you so much

also tried to use control net with a decent shot which i got from dalle.. but no success

[supple finch]

That's where I started from

[slow mist]

i have never used pony, so i cant help more, but i am sure someone else will

[supple finch]

I think I might be asking the wrong question, tbh. I'm been trying to get this model to work, since it seems to be the best for faces and generally realistic scenes.

But if you have a recommendation for a better model for general use, specifically to train my face on, then I'll just swap over to that one, instead of breaking our heads trying to get pony to work

[proven jasper]

i dont think its a model issue,looks like you are using wrong settings,maybe hires fix enabled or refiner enabled

[supple finch]

I totally agree, though if there's a model that's more effective 'out of the box', I'd move to that. Eg. I've previously used dreamshaper and juggernaut, I'm just not sure whether they're still current, and also feel like since they're massive checkpoints, the may take a long time to train on my face as a prompt

[proven jasper]

yea juggernaut its great if you want realism

even though there are pony realism checkpoints out there,they are just ok at producing highly realistic imgs better to use a model that was exclusively finetuned for realism unlike pony which was created for 2d and 3d imgs

[supple finch]

Since I'm 'only' on a 3080, I'm also trying to aim for a model that is speedy. Do you have any experience with the quality of Juggernaut XL V 7 Lightning 4s?

[supple finch]

Gothca, given the name of it, that makes sense, I'd just never thought about its origins, given the very nsfw curent state of the model (which is not my intended usage, not that it matters)

[supple finch]

I happen to have V9 Rundiffusion photo 2 already installed, just scared of it taking aggggeeees to generate a model using dreambooth?

[proven jasper]

the only issue i had when i tested other lightning models was that some loras dont work very well with it,but other than that they were great for producing fast images

[supple finch]

Oh yeah, now that you mention it, I did have some LORA issues, don't remember exactly which ones. Thanks! I think I might give lightning a go!

[proven jasper]

now if u want a realistic model that can do nudity you could also try with leosam helloworld xl , it has a normal version and a lightning ver

[supple finch]

Thanks, how's its overall versatility in comparison with juggernaut? Would love as much of a one stop shop as possible, though I know that doesn't exist

Oh, just checked it out, looks like it seriously rivals juggernaut! Some amazing artistic output

[proven jasper]

yea plus it was finetuned with gpt4 in case you use an llm to write prompts

[supple finch]

Though it seems that the lightning model is a few behind? 5.0 euler lightining vs XL 7.0

[proven jasper]

since you can make any sdxl model lightning by using the lightning lora you could try to do that,only problem with that approach is that you have to play with a lot of sampler settings+prompts to make it work

[supple finch]

Have you done much training on faces / objects? I'd imagine lightning would be much quicker for that, but I'm absolutely shooting in the dark, so would be very happy to be corrected and just use the SDXL version for training, if that's the case

[supple finch]

Right, which I'd imagine is a lot of seemingly random trial and error?

[proven jasper]

yea

[supple finch]

I didn't even realise lightining works based off a LORA, I thought it was a descriptor for a new type of checkpoint that's much quicker

[supple finch]

What do you reckon the best option for this would be?

[proven jasper]

sorry i wouldnt know i dont really train faces but i do know that you cant really train on lightning mode,just train with base sdxl or juggernaut

[supple finch]

Fair enought, that's very helpful in itself. Thanks!

[wintry topaz]

Where this issue comes from? Is it Promt or Setup?
some objects in the Picture has purple blue diffuse artifacts right next to it....

Steps: 25, Sampler: DPM2 a, Schedule type: Karras, CFG scale: 7, Seed: 975717070, Size: 512x512, Model hash: e1441589a6, Model: v1-5-pruned, Clip skip: 2, RNG: NV, Lora hashes: "InstantPhotoX3: 7036b103fef3", Sigma noise: 0.8, CodeFormer visibility: 0.73, CodeFormer weight: 0, Version: 1.9.4

[supple finch]

Wow, just tried it and my test came out even slightly nicer looking than the original! Very happy with this model, thank you!

[supple finch]

What were your prompts and model used?

[edgy river]

Hy guys! Having problem with the ipiv's Morph - img2vid AnimateDiff LCM / Hyper-SD workflow, recieving the following error: Does anyone had this? Cant figure it out :/

[faint ocean]

hey, its me again, dose one of you remember the prompt for a transparent background? it was something RGB ... alpha bla bla i just cant remember. it would be rly nice if some of u could tell me ❤️

[wary hare]

Hey all, sorry if this is a super dumb question. Whenever I try to generate any kind of object that usually has some sort of label it generates with text. I'm aware that this text is bound to be mostly unusable, however no matter what I seem to try, I can't adjust the prompt to exclude text and or labels from being generated in the scene. That would be fine since I could just go through and tweak them in photoshop, but I'm trying to generate a lot of individual images so it could get time consuming fast.
Right now I'm working within a simple gradio project that I setup with realvis4 (but I've also had the same issues with SDXL and SD3 Medium when i've tried them in a spcae setup the same way) I have been using the same seed for all of my generations up to now so I can focus on debugging

Here is my current prompt

"A Realistic photo featuring an open kitchen cabinet with simple solid color containers holding baking soda, showcasing its granular texture, If faces are shown they are realistic and good looking, No Label, No labels, clear blank bottles, clear blank containers, clear blank bag, clear blank jar, unlabeled jar, unlabeled bag, unlabeled container, unlabeled bottle"
Seed: 300000000
Inference steps: 100
Guidance Scale: 12
Sampler: DPM++ SDE Karras
1280x720
Negative prompt: "(worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch, text, writing, watermark, signature, sign, brand, fonts, anything resembling text, bottles with text, bottles with logos, logos, labels, written, written labels, container with logo, ugly hands), open mouth, oversized, too big"

Help would be greatly appreciated

[wary hare]

Nevermind, fixed! If anyone runs across this with a similar issue, it's related to how the models are trained. When I say baking soda it thinks of the box because that's it's reference, however when it's described as a textured white powder it understands that and works perfectly. Essentually "baking soda" implies a brand, label or a box, not a substance

[tropic dome]

anyone knows how to get these small doodles depicting shaking throbbing movement emotions and stuff in mangas into stable diffusion

[quiet raft]

How can I do faceswap with Stable diffusion?

[silver valley]

The overlap is like a seperate task that gets generated after the tiles.
Bigger overlap can melt the two tiles better together but also to big could cause deformations. I mostly use the normal overlap

[silver valley]

For realistic images you can use the Reactor extension.
For Anime and semi realism you need to use Controlnet with IP-Adapter Face

[quiet raft]

You know a good video on how to use Reactor? 🙂

[silver valley]

nope, but for the start its pretty simple, installing it (a bit tricky), then enabling and dropping in an image of a person you want to use the face of it

[quiet raft]

I tried haha 😄

I made a "model"

[silver valley]

was the bottom image the input?

[quiet raft]

This?

Does ReACtor only switch the face, not the hair etc?

Because when I do this and add a prompt..

There is no blue eyes..

[silver valley]

yes reactor switches only the face

fpr hair you would need to use Controlnet IP-Adapter

[quiet raft]

But where is the blue eyes from the photo I provided? Did I do something wrong?

[silver valley]

your using adetailer too?

[quiet raft]

I got it installed but not used

[silver valley]

make sure its not enabled when using reactor

[quiet raft]

So how can I get similar face from the image I put in?

[silver valley]

check the cmd

it should say if it detectet a face

[silver valley]

hmm

[quiet raft]

The image from prompt and the faceswap.

It does work, but not adding in those blue eyes

[silver valley]

and if you prompt for the eyes?

[quiet raft]

I tried without eyes and adding in blue eyes.1.1 to focus a bit more on it. But not the same eyes as the model 🙂

But if I want to keep he face and the stuff.. I always need to have prompts like Blue eyes, long blond hair and that then?

[silver valley]

that would help yes

[atomic wigeon]

I typically use something like "centered and symmetrical bright blue eyes and black pupils". Try that and see if it helps your prompt result.

It should also be noted that in many cases where you add a color to your prompt it will alter the entire scene. The clothing, the background, the lighting conditions, etc. So sometimes just adding something like "blue spotlight" will result in a much shorter and direct alteration of your image to image result. That includes eyes. But it will also associate the eyes with the typical skin color and race that has those eyes. So Americans and Europeans. You will lose the Asian race and racial traits typically associated with Asians. Which in some cases works as an fantastic intentional tool to modify your image.

[odd mountain]

I am looking to make some paper cutout characters to glue onto foamcore for use in a paper theater. I found a lora in the style of Akihiko Yoshida (Final Fantasy), but I'm having trouble with flat colors, warped faces and random background elements appearing. I also have no idea on what sampling method to use. Any tips?

[atomic wigeon]

Add the word "sticker" to your prompt.

Near the beginning of the prompt, not the end.

For colors use Davinci Resolve (free).

And/or you can add a color scheme like "soft colors", "earth tones" or "pastels" to your prompt. For brighter colors add "vibrant colors", "happy colors" or "bright colors".

[odd mountain]

Awesome, thanks!

[tribal skiff]

how can i use these models to generate something?

[atomic wigeon]

You need the equivalent of a doctorate in astrophysics to put it on your computer. Or you can just do cloud websites.

[marsh echo]

Do you guys know how to avoid half of the head always being cut off?
No matter what size I pick, square or rectangle, it always cuts off the top of the head.

[atomic wigeon]

Words like "centered and symmetrical" help place your content right where you need it to be. Additionally you can add words that will trigger the AI to keep it in frame. Words like "forehead", "hair", "hat", "shoelaces" or a type of hair style like "braids". Doing that will tell the AI you want to see that specific part of the body.

So if I add "wearing a Rolex watch" it will show his or her wrist(s). Be careful with these words tho. Sometimes they will give you a negative result. For example "ponytail". You cannot see a ponytail from the frontal camera view (unless it comes over the shoulder). So you force the AI to turn the character. Describe items or things that would only be found in that specific location and from your desired camera view. Bangs, balding, hair extensions, hair roots or hair dye (commonly blends a color near the skull), hair clip, burette, etc. You can also use common professional camera terms like close up, extreme closeup or half body shot.

[quiet raft]

Hey

I got a friend who just installed Stable and has this issue

[quiet raft]

I think it was the VAE .. It was on automatic and he had none VAE

We did another model with VAE set to none and it was better.

Second thing.. What is the best upcaler for a bit more realistic images?

I use R-ESRGAN 4x+ Anime6B for animated / cartoon

[silver valley]

For realistic stuff its 4x Ultrasharp and for semi realism NMKD Siax 200k is nice

[quiet raft]

I don't have those I think

[silver valley]

You can download them here:
https://openmodeldb.info/

Put them into models/esrgan

[quiet raft]

Alright, I will test

[stuck heron]

am i allowed to send nsfw images or no its fine if not the leg is like duped and i have alot of negatives any one recommend a short neg list?

clutter, poorly drawn face, poorly drawn hands, poorly drawn eyes, boring, (deformed, distorted, disfigured), clothes, bad anatomy, mutation, low contrast, low-res, FastNegativeV2ugly, (worst quality, low quality:1.2), EasyNegativeV2, more than 1 person, fog, grain, duplicate heads(deformed, distorted, disfigured:1.3), ugly, noisy, (worst quality, low quality, normal quality, bad quality:1.4) , easynegative, background (deformed, distorted, disfigured), ugly, blurry, crossed eyes, (worst quality, low quality, bad quality:1.4), out of frame, off-center, pastries, clothes, piercing's,

[atomic wigeon]

No NSFW stuff allowed, no.

[stuck heron]

im pretty sure this is ok to show if not delete it but why is this happening please somone help i installed a vae and out of the blue this happens it worked fine b4 and i even have the vae on none first

[quiet raft]

Uhm this is weird.

Using Reference and ADetailer..

Without ADetailer

But I wanna use adetailer to fix faces

As soon as I put onj ADetailer It goes like this

[silver valley]

can you show all settings ?

[quiet raft]

Worked fine earlier today

But now.. all get messed up

Changed nothing beside prompts

[silver valley]

load an image into png info where it worked, then send it to txt2img and change the prompt

[quiet raft]

I went to extensions and did update and restart ui.

Now it worked :/

Same prompts

[silver valley]

okay nice

so it was an extension bug

[quiet raft]

But how? :/

Did nothing ..

[quiet raft]

Wait.. I turned on Hires.Fix

[silver valley]

dont use 0.7 denois

use 0.5 or lower

anything above will output deformations

like in your image

also 10-15 hires steps is enough

also ultrasharp is a bad upscaler for semi realism

[quiet raft]

I should use 4x_NMKD-Siax_200k instead?

[silver valley]

yea thats good or lolipop or maybe esrgan4x or an anime upscaler

[quiet raft]

:/

[silver valley]

try 10 hires steps

[quiet raft]

Same :/

Somethingh with hires

Without highres activated.

[silver valley]

is your webui updated?

[quiet raft]

Yeah

Hires.fix

Maybe it was something in the prompt that made it fuckt up.

I took it to img2img as you wrtoe and just did simple prompt thing

[silver valley]

nice

[quiet raft]

Is it not better with higher highres steps?

[silver valley]

not much

10-15 is enough

[quiet raft]

Any idea why one is more detailed and the other one is not?

I used these

They use the same VAE

[nocturne gale]

Hi so I am using PonyDiffusionXL V6 and I'm trying to make an image of a female anthropromorphic dragon character that has black scales with a red underglow, but the black scales hardly ever show up and my main issue rn is that I am getting a full dragon head instead of a more anthro/human dragon head and face and I can't seem to correct this. Here is part of my prompt:

(human, anthro:1.5), (Anthropomorphic female dragon:2) with a detailed human like face on a dragon like head, (black scales glowing red from beneath:1.25), and (dragon-like body features:2), laying on back in the water of a cherry blossom circled lake

I've also tried to use the PXL Human-Anthro slider lora but that doesn't seem to help me. I either get full bipedal dragon or a human body with a dragons head or a full on human.

Please tag me in replies if you would be so kind.

[nocturne gale]

I have a feeling it's the automatic schedule type. try tweaking between the sampling methods and schedulers

I don't use WebUI so I can't give specific examples of what to try using. You could also try using a lora to add more detail. There are quite a few you can try.

[nocturne gale]

I'm not sure what you mean buy score and source tags

[silver valley]

Oh then you didnt read on how to use Pony models.
Check the civitai page of PonyV6.
In the description you find what pony needs as prompts to get good outputs

Like source and score tags

Then you can mix them up.
I would go for source_pony source_anime, female dragon girl, in positive,
And source_cartoon, 3d, realistic, in negative

Score 9, 8 and 7 into positve and 5 4 3 in negative

Then change the resolution to a portrait format like 768x1024

[nocturne gale]

I am using portrait I just only posted a cropped image cause of NSFW features

[silver valley]

Ah ok

[nocturne gale]

I redid nearly my entire workflow for SDXL and adding rgthree's stuff

[silver valley]

Oh wow

[nocturne gale]

yea that took like three hours to complete because of a stupid bug

[languid pelican]

типичный завод в Пскове

[nocturne gale]

cause apparently Model Merge cannot be first in the muter unless it's enabled. Otherwise it throws an error...

took SOOO long to figure that out

also apparently CR VAE Decode's Tiled option sucks so had to switch to this and now I'm never running out of VRAM while decoding the Hi-Res Pass

altho I really am not getting great results with Hi-Res Pass

[nocturne gale]

This kinda destroyed the output ngl

and didn't help with the dragon face

[silver valley]

Destroyed ?

[nocturne gale]

yea like the hands got all blotchy and the water around the character was this nasty deep blue unlike the rest of the water and stuff. it just kinda nuked the generation for some reason

[silver valley]

Very strange

[nocturne gale]

yea idk I feel like a total idiot when it comes to prompting and stuff

[silver valley]

If I'm at home I can test it again.

You want to get a dragon body + human face right?

[nocturne gale]

yes, preferably with like horns and stuff on the head but a more human shaped facial structure.

I'm also really struggling to get it to make the dragon scales with red underglow like I want

[nocturne gale]

something like this, I hope this isn't too nsfw to post

[silver valley]

I see here you prompted for "with (no dragon snout)"
In positive prompt.
Thats not how it works.
The ai will try to generate everything mentioned in the positive prompt.
It will try avoid it only if you put dragon snout in the negative prompt

[nocturne gale]

I've changed my prompt a lot since then.

Female anthro dragon with a human-like face, (black layered dragon scales with a red glow:1.4) under each layer, laying on her back in the water of a lake, surrounded by cherry blossom trees, (spread legs:1.2), scene showing depth from a perspective allowing depth,  (human-like face:1.3),

tried to use a GPT to help me get a better prompt

still getting this tho:

[silver valley]

Hmm okay

[frigid kernel]

Hello everyone, what should be the input prompt to extend the below images. I am trying to automate the extend feature so I require one specific prompt for the extension of both the images.
for patterned images, it's working absolutely fine. But for plain background images, it's adding some background which is not matching with the original image bg.

Prompt used: Generate creative background scene matching original image. Environmental scene, city life, dressing style, nature, building, non-living objects.
Neg prompt used: Blurry, bordered, zoomed, solid color, monotonic background, disfigured, human figure, living objects, gore, dead, hazy, dull.

[atomic wigeon]

The best method is to cut out your character and apply it over the background style you like. Then image to image.
However, to answer your question there are some prompting words that may help.

1. Gradient background (in the color you want). So "Green and black gradient background" for example.
2. Spotlight (in the color you want). So green spotlight for example would alter the entire scene in shades of green. That includes clothing, background, highlight items and even eye color.
3. In the forest. In the city. In the office. In an ornate wooden corridor. Describe the location where the character is standing. The AI will fill in the scene. Then you can image to image and apply the "spotlight" as mentioned above.
4. Drag your image into a video editor like Davinci Resolve and apply a color background/generator over top of the image. Now use the filters or composites to integrate that specific color into the scene. Now save the image.

[atomic wigeon]

.

Remember that this is AI. It's a "ballpark" estimate of your finished product. In most cases post production in third party sofware may be required.

For creative backgrounds I recommend "desktop wallpaper". So your prompt would look like: desktop wallpaper background of swirl smoke with wire wheat highlights

[normal belfry]

Hey guys! Do you guys know how I can create something like this using SD? Any tips would really be appreciated :)

[atomic wigeon]

retro or retrofuturism

Then the rest of your prompt

[normal belfry]

Yes, here's what I was trying to generate images with:

Retro futuristic illustration of neurotechnology, cyberpunk style. A scientist wearing a BCI (brain-computer interface) headset sits at a control panel, facing a large holographic display of a glowing red human brain. The scene is filled with neon lights, circuit patterns, and data visualizations. Dark background with bright blues and reds. 1980s sci-fi aesthetic. Highly detailed, subtle, sharp edges.

Issue is all models that I used can't seem to get the headset right -- they are creating either headphones or helmets, nothing in between

[atomic wigeon]

Sometimes less is more. Just use default SDXL.

You can also add helmet and headphones to your negative prompt.

[normal belfry]

We're getting there :) I added Highly detailed, 1980s magazines artwork to the prompt

[normal belfry]

oh yes, seems like a good idea. I'll try that

[atomic wigeon]

A couple other potential creative ideas that you could add:

1. Psychedelic
2. pop art illustration or just PopArt (one word)
3. Musicalism
4. dreamlike Expressionism

Adjust the strength of these additions by moving them closer or further from the front of the prompt.

[normal belfry]

Oh, does the order of words in the prompt affect their strength in the image? I didn't know that, thanks!

[atomic wigeon]

Yes it does. So if you're getting too much of something don't change your prompt, change the order.

It's the art within the art.

[normal belfry]

Very interesting

[tiny frost]

anyone knows how can I produce diferents images about the same product?

[nocturne gale]

Here is my current Workflow so you can recreate my setup: https://www.nodecafe.co/workflow/tZC74mY6sh6baSIvpQDtH

[atomic wigeon]

Controlnet

This is a gaming community in the game Ark Ascended. In this case all I did was take the red and white "ArkVader" and put it into controlnet so it stays the same. And then I applied a prompt.

That's how it works. Use it with anything. Results vary.

[chilly pecan]

hELLO!
I generated a chair using stable diffusion, but now I would like to change the color of the chair to black, do you know how I can go about it? THANKS

[quiet raft]

What's is your current prompt?

Have you tried color: black or black chair?

[chilly pecan]

I have exactly the result I want and I would like to simply convert the chair to black, because when I try to transform the chair into 3d, instantmesh does not recognize the 'white of the chair'.
My solution is to tinker with the colors in Photoshop, but this takes a very long time because I have several chairs...

[atomic wigeon]

Davinci Resolve. Send me the image (or any image) and I'll give you an example.

@chilly pecan

[chilly pecan]

mp

[atomic wigeon]

That was obviously a jpeg. But if you had a png it would alter only the chair.

[chilly pecan]

yeah i can do the same thing in photoshop

i wanted to know which prompt use for change color of my image

[atomic wigeon]

black chair

or controlnet

[chilly pecan]

in img2img ?

i already test and nothing happening

[atomic wigeon]

Controlnet is image to image yes, well sort of. Then you adjust a strength slider. Apply "black chair" to your prompt. Done.

Example: #📝｜prompting-help message

[chilly pecan]

'black chair'

[atomic wigeon]

Image to image and controlnet are two different things. Controlnet uses the same method but locks the image down and allows you to prompt around it. Another alternative is inpainting.

[chilly pecan]

yes i test inpainting too but it not working for me

[atomic wigeon]

Ya controlnet or davinci resolve with a PNG are my recommendations.

Image 2 image and inpainting are lottery.

[chilly pecan]

Ok 🙂

[atomic wigeon]

If you could remove the background your life would get a million times easier.

[chilly pecan]

I put 'black chair' in the prompt

[atomic wigeon]

To get the identical chair you need to use controlnet.

Or invert the color.

[silver valley]

here is my try, but i used SDXL_autismmixPony model because its not that focused on anthro.

[nocturne gale]

is stock pony like more focused on anthro?

[silver valley]

i would say so

other pony anime models mixed more anime images in

[nocturne gale]

hmm maybe I'll try working with another model

do you think you could help me with another thing as well

I was told to try using Ultimate SD Upscaler instead of a second ksampler for Hi-Res Fix but I couldn't get it to add any new detail to the image

[silver valley]

here is a dragon girl again with this prompt:
score_9, score_8_up, score_7_up, source_anime, masterpiece, best quality, 1girl, female dragon hybrid, teenager, dragon girl, scaled body, scales, black scales wth red glow, in castle, medium breasts, dragon body, scaled skin,

[nocturne gale]

still couldn't properly capture the red underglow of the scales I see

[silver valley]

im not a comfyui main user, bette r ask that in #🧣｜comfy-ui
but i can look into it

[nocturne gale]

I did

it's just so dead atm

rn I am just using upscale image with no second pass so I can do some tests with prompts and lora weights, but I have a pretty powerful workflow setup atm.

[silver valley]

got it lucky:

[nocturne gale]

HOWWWWW

[silver valley]

black scales with red glow accents,

[nocturne gale]

Her (skin is covered in diamond shaped dragon scales:1.5) painted in a striking red and black color scheme, with obsidian black scales with reddish orange glowing accents under the scales.

literally have this in my prompt

in prompt_g

and this in prompt_l which is for more tokenized prompts:
black scales, red accents,

[silver valley]

have you tried an other model?

[nocturne gale]

No I've tried to stick to stock pony for now since everytime I switch models I have to completely nuke my prompts and start from scratch otherwise I just get a blob or screwed up generation

[silver valley]

ah okay

[nocturne gale]

rn trying to see if there is a LoRA, DoRA, or LyCORIS that would allow me to adjust the head

[silver valley]

the base pony model works not good for me, all other pony based models performa much better

better results, better quality

[nocturne gale]

these are like the closest I've gotten to a normal face sadly

[silver valley]

here with prompt

[nocturne gale]

guess I'll be changing models soon

[nocturne gale]

what selfbot are you using btw? I used to use nightly but haven't in ages

[silver valley]

selfbot?

[nocturne gale]

your changing status is usually a feature of a self bot

[silver valley]

ah, thats a done with an Plugin for the Programm called BetterDiscord

[nocturne gale]

ah I see, I had to get away from betterDiscord for various reasons. Now I use Vencord

sometimes not even that, just OpenASAR which also has theme support

[silver valley]

yea vencord is awesome, but i guess it misses some plugins like that status thing

but is vencord less resource hungry than betterdiscord or do you see no difference there?

[nocturne gale]

well since all the plugins are added by the creator himself it makes sense. But I also like it that way cause it MASSIVELY reduces the risk of bad plugins

oh much less intensive on resources

especially with OpenASAR as well

my discord launches in like 4 seconds

[silver valley]

does openasar work with vencord ?

[nocturne gale]

yea it's part of their installer

[silver valley]

oh thats nice!

[silver valley]

guess ill have to try it out ^^

[nocturne gale]

part of my theme is this little cat that follows my cursor around, perfect for my ADHD ass

[silver valley]

haha cute

does vencord has a "no telemtry" plugin?

[nocturne gale]

I believe so

[silver valley]

nice

[nocturne gale]

Honestly I haven't checked all the plugins in a while

I have like 80 enabled or some shit

honestly I love comfyUI but I wish the Hi-Res fix was easier and worked a bit better, other people have had decent luck with it but I seem to struggle with it deteriorating the quality of the image a fair bit

I even use controlnet to apply lineart in order to preserve detail but I can't tell if it's also hindering the ability to add more detail to the image.

[silver valley]

i only tried upscaling SD3 today in comfyui, used Ultimate Upscaler, but i wanted to use "Hires Fix"

need to build a new workflow for that tomorrow

[nocturne gale]

I can start adding SD3 into my workflow and then send that to you. That way you don't have to start from scratch

with rgthree's custom nodes I've been able to add basically all my models from sd1.5, sdxl, and pony into 1 workflow with no issues

[silver valley]

no its fine, i need a bit more practice in comfy ^^

[nocturne gale]

aight

the SDXL part of my workflow

[silver valley]

very nice

[silver valley]

whats CTX and CTX hires?

[nocturne gale]

CTX is the Context nodes added by rgthree's custom nodes. Think pipes but a LOT better

I've just renamed some of the nodes in my workflow for better management

[silver valley]

ahhh okay

[nocturne gale]

example

Fast muter also is what is allowing me to have mutliple types of models in the same workflow

[silver valley]

thats really clever, i wondered how it works in your wokflow seeing them all together

[nocturne gale]

yep

I love it

it's so handy

like my entire Hi-Res fix portion of my workflow looks like this

Super compact

btw I found a super cool set of LoRAs for Pony you should check out

https://civitai.com/models/352581?modelVersionId=628657

[silver valley]

Yea Ive seen them before but still haven't got the time to try

[atomic wigeon]

Is there no other thread to have this discussion other than "Prompting Help"?

I'm genuinely trying to help people with their prompting. If we don't have a thread for lora's and whatnot can we please create one? Can we please get a thread called "loras-help"?

@hard elm

[silver valley]

Sry it was prompting help at first but we got into other themes.
Should have moved to general chat or comfyui.
No need for a lora thread I guess

[atomic wigeon]

No I apologize, I wasn't intentionally trying to sound like a jerk. It was a genuine question. Loras are a major discussion topic worthy of it's own independent thread (assuming we don't already have one). I have many threads turned off, that's why I asked.

Lora-Help <--- good name

Put it right below this thread

[silver valley]

Ahh okay, there is no lora topic rn but we have #🔧｜finetune where loras could belong to (if its about training, and not using)

[atomic wigeon]

That's not very intuitive tbh. I assumed "Finetune" was one of the gajillion StabilityAI names. lol

I do think we need one dedicated to loras specifically. It's a massive topic and deserves a rightful home.

[silver valley]

Yea, thats an good idea I can give to the team to discuss

[atomic wigeon]

Thank you. I would speculate that about 50% of everything in this thread relates to loras. Def worthy of a home thread.

And I apologize about my question. It sounded rude but was not intended to be. 😆

[silver valley]

No worries ^^

[normal belfry]

progress :)

[glacial jolt]

Hi guys which SD plugin ppl are prefering to use in associate with photoshop nowsaday?
I'm still using Auto-Photoshop-SD atm but i think its outdated

[atomic wigeon]

Everyone quit using Photoshop for Davinci Resolve.

Creative Cloud is $90 per month now.
Davinci Resolve is free. Or for a better version it's a one time payment of $295.

Resolve has a steep learning curve. Especially for Fusion (a node system). But for purposes of this discussion, and AI art, Resolve is better on many levels. I do think both Photoshop and Premier Pro have more intuitive fun toys at intermediate levels. But to hell with their greed.

[edgy tundra]

...or Krita

[atomic wigeon]

Yes I agree. Krita has some fun toys too.

[graceful sapphire]

If I want a image with two very different persons (in clothes, face expressions, age, height, hair colour, skin colour etc. ), how to prevent the AI mix the characteristics up?

[silver valley]

By using the regional prompter extension

[nocturne gale]

anyone know how to properly use 4th Tail model? I keep getting really crap generations but people say it's like the best pony based model there is but I can't seem to get it to do what I want.

[nocturne gale]

/rules

!rules

uhh

@hollow tapir Hey not sure if I am free to ping you community guides but by chance have you worked with 4th tail model?

[hollow tapir]

Nah never tried it, but it's a pony model so you can look for some pony tutorial / prompt guides / examples and learn from them.

[nocturne gale]

I've been working with stock pony for a few days now so I'm okay at prompting with it

[hollow tapir]

also rules are explained in #✍🏼｜rules-and-tos

[nocturne gale]

that's why I'm so confused while using this model

yea I checked that before pinging you

to see if there was a rule about not pinging you guys

[hollow tapir]

Not sure what else to tell you then. If you can share some TASTEFUL examples of your prompts/settings/outputs then people might point out what you're doing wrong.

(in case it's not clear, keep the porn stuff out of this discord. Thanks)

[nocturne gale]

yea unfortunately I'm attempting to do NSFW generations so it's difficult to share those things without editing the prompts heavily. hard to get help with the NSFW stuff and I haven't found a good server for discussing this stuff yet

I'm surprised that this server doesn't have NSFW dedicated channels considering the absolute MOUNTAIN of NSFW related models and resources you find on civitAI and huggingface

it feels like there is at least 3-4 times the amount of NSFW stuff compared to SFW

[atomic wigeon]

Because most people that make NSFW AI art are pedophiles.

They ruined free expression.

[nocturne gale]

mm

[gaunt meadow]

Hey do we have some promting experts here? If so feel free to shoot me a DM, I am facing some issues with my prompt. I use the "ultra" model.

Here is a notion sheet explaining the issue https://furtive-bobcat-8ed.notion.site/Stable-Diffusion-Prompt-issue-b6138599b165437382e1ead2725b9175

[graceful sapphire]

Where can I find regional prompter?

[silver valley]

In the extensions tab, available, load from

[gaunt meadow]

@silver valley any idea on how to solve this issue?

[silver valley]

Hey, this is just random. If you generate multiple images without the same seed you will get a guy in a car eventually

It also could be the model understanding or the wording issue

[gaunt meadow]

so i need to randomize the seed?

[silver valley]

Try for example "sitting in a sportscar" or "driving a Lamborghini "

[silver valley]

That helps yes

[gaunt meadow]

ok why is that?

i thought having a fixed seed causes same results

[silver valley]

Yea and thats the issue. Fixing the seed will cut the creative part of it.
It will be to similar to the other image cause you only changed 2-3 words

A random seed would give you a full random image

Not tied to anything

[gaunt meadow]

if i use the same person description would it still be the same looking man?

cause thats the goal

having the same person in different scenes

[silver valley]

Not exactly. Thats why most people use extension to Faceswap or train loras for a character

[gaunt meadow]

ok ok i see

[silver valley]

I would recommend using Reactor to faceswap or using Controlnet IP-Adapter

With Controlnet you can also set a fixed pose

Like standing or sitting etc

[gaunt meadow]

can i use all of this through an API?

[silver valley]

Hmm idk

[gaunt meadow]

ok ok

randomizing the seed was the solution

i would say its 95% the same person if i just change the scene

[elder marlin]

hey im having problems installing ReActor, it's not showing up when i installed it. Any way you guys could help me out?

[rocky belfry]

Hey, need someone's help with gen. that has stable installed on a local machine

Long story short im away at work and will be home in like a month. Need someone with "stable" running locally (non XL, can be makeayo). Had been bored and wrote a prompt in my spare time. Wanted to see how well it performs, but it needs to been locally since its long/weighted and i wont get the same results with cloud based gen. (like never). Side note its nsfw(nothing hardcore or lolicon). I would send both positive and negative privately. Thank you

[silver valley]

Come to #🤝｜tech-support
And show your cmd log when launching the webui-user.bat

[quick thicket]

I am curious if there is a best practice of when to use BREAK and AND in prompts. Are there situations where you should and should not use these?

[marsh fractal]

is there any prompting guide that i can refer for SD3 ?

[dusky stream]

How can I make sure my images are in black and white? Is there any good way to do it through img2img?
I'm trying to create sprite art to be colored in later. All the resources I'm finding are on how to make uncolored images colored.

[left prism]

That's a bit of an odd assertion. I would both expect and have found that pedophiles make up the minority, just as they do in broader society. Their mere presence is awful though, but groups I've been in that do NSFW report and combat these monsters whenever they wash up like rotting jellyfish on a beach. Not that I'm a huge supporter of teh pron, but I've definitely gotten more help and learned more in those communities than I have anywhere else. That said I've also seen some remarkable and tangibly brilliant art in these communities. It just usually gets steamrolled by teh pron.

[calm gale]

does anybody know how i can make my images not look like this??

[modern bridge]

Hi anyone knows how to make exact head angle pls

[calm gale]

nahhh answer my quesion first dawg

[silver valley]

Hey, what are your txt2img settings?

[calm gale]

probaly shit

[silver valley]

What's the model you use?

[calm gale]

a1111

or no like

uhh

[silver valley]

Check the file size of it.
A model is always 2gb or larger.
If its smaller its not a model

[calm gale]

alr

[elder tulip]

hello there, i basically want the same crown just with black and gold as colours, but i cannot find the correct settings

[atomic wigeon]

I was an official tester for an AI company. My experience is quite the opposite. The honest ones never even made sexual content at all. Their interest in NSFW was free speech in politics and/or a nude model to sketch on real life canvas.

The most common offender was those that make anime. And then some, maybe 3%, are just straight up evil. When we started to be able to search images by words we saw "8 your old nude" type prompts. Lots of them. It gets worse than that. Much worse. But I'll divert that to law enforcement and the Center For Exploited Children.

[atomic wigeon]

For something like that you want to create JUST the crown. In order to do that you must remove the rest of the image. Or create it from scratch and apply it over the image. The fastest and easiest way to accomplish what you want is to throw it in Davinci Resolve, remove everything but the crown in the Color tab (masking) and then color it right there in that same tab. Or export it and controlnet that sucker.

Big project. Hours. AI inpainting just isn't to that point yet. There are exceptions but they're not found here.

[atomic wigeon]

Here is a potential workflow for you:

1. The prompt - hyper-realistic crown (then add your specifics).
2. Remove the background. Run it thru image to image again to clean up the background removal.
3. Make a greenscreen image or find one in google/youtube.
4. Put the crown on top of the greenscreen in an image editor or video editor. Rotate as needed. Zoom in as needed. Davinci Resolve or Photoshop are best. When you apply a composite the green will disappear over the entire scene. Only the crown will remain. But it's now an editable layer for special effects. Now your entire scene will composite to the crown layer evenly. This is a super easy way to animate AI images as well.
5. Layer the project with your character at the bottom and apply the greenscreen crown on the top.

Done.

[zinc zealot]

Does anyone have advice for Logo prompts and how high the CFG should be?
Been using Redmonds Logo Lora
https://cdn.discordapp.com/attachments/1026382406279770152/1261689478318260347/image.png?ex=6693df64&is=66928de4&hm=a98a712a7a8092bb2102f01e98033773a4d6bac50bd6bb1799b988590f0d905e&

with prompts like this
logo, claw reaching out of the screen (animal paw with claws), neon claw, 3d realistic <lora:LogoRedmondV2-Logo-LogoRedmAF:1>
but i get mostly shit
using fp8 to not run out of vram idk if thats an issue
and sdxl
https://cdn.discordapp.com/attachments/1026382406279770152/1261689763975532594/image.png?ex=6693dfa8&is=66928e28&hm=240ffa45396b89fcc0dfdb5aca244dcc1c75f767eaac3dcc98a534580d3a4d52&

[elder tulip]

thank you very much, as i understood, the fastess way would be, to just throw the img in an editor and remove the colours from the crown :D

[atomic wigeon]

Haha. Sorry. Short version - AI just gets you in the ballpark. Post production is where the work starts.

[elder tulip]

👍

[opal jetty]

the first 20 frames where normal after that the rest turned into a neon effect i dont know why and i specialy wanted the background to have more stylized background but i always fails i need some simple explaination to fix the little problem

[opal jetty]

sometime it can happen if deppending of the model you using this type of problem to me when i was doing text2vid]

[surreal rose]

the AI creates a frame, then modifies it for the next frame. as more frames are added, artifacting creeps in and color shift happens. remember it is drawing one image for each frame, and in order to give you a video it has to pretty much copy the previous frame and modify it for the second one, then copy the second frame and modify it for the 3rd one and so on. you can't prompt that away. what you can do, however, is pull your clip into something like Davinci resolve and color correct

[vocal dew]

I keep getting too much of the "warm hues" .. what is causing it among my prompts ?
Current Model: Dreamshaper-inpainting V8

Positives

bedroom, bed headboard, wall mounted headboard, headboard with wooden frame, ultra modern, 8k, insanely detailed, (led lights: 0.9), (shelves:1), wall paintings, interior design, (tufted headboard:1.1), (spotlights:1.1), transitional, bedroom design, (wallpaper:1.2), (wall covering:1.15), architecture design, soft cinematic light, (cinematic look:1.2), flower wallpaper, vibrant colors <lora:add_detail:1.2>

Negatives:

plants, succulent, side table, table, curtains, easynegative, bad proportions, low resolution, bad, ugly, terrible, painting, 3d, render, comic, anime, manga, unrealistic, flat, watermark, signature, worst quality, low quality, normal quality, lowres, simple background, inaccurate limb, bad composition, error, cropped, low res, worst quality, low quality, normal quality, jpeg artifacts, trademark, watermark, artist's name, username, signature, text, words, human, leather, switches

[nocturne hinge]

My scrapped/broken futuristic items are blurred and I need help fixing it

Positives

Highly realistic desert landscape at sunset with scattered futuristic structures, subtle neon accents blending with the environment, hidden Fallout 4 Easter eggs, and detailed sand textures

Negatives

No medieval elements, no unnatural landscapes, no daytime scenes, no modern vehicles, no abstract elements, no low-resolution textures, no blurry areas, no overcrowded structures, no excessive lighting

[opal jetty]

i see now

[quiet raft]

Hey, I am having trouble to use AnimateDiff .. it keeps crashing on me.. I am not able to do anything. There is a huge log of errors in the console so not sure where to begin problemsolving it.

Anyone else having this issue?

I use these settings

This is the entire error

[silver valley]

What's your GPU?

Try without enabling FILM.
If your GPU has 12gb vram, try set the context batch size to 8

[wicked basin]

hello, i want to make a picture of 2 original characters. both appear very similar to each other . any ideas how to prompt? how can i seperate prompts for individual characters?

[quiet raft]

I have a 2080ti

[silver valley]

Hey, that can be done with the Regional prompter extension.
There you can define what persons are shown and individual describe them. Also it supports loras

[opal jetty]

film takes extra memories

[rain tusk]

I've got an issue (using Automatic1111):

I want to generate some variations of an image, and for that I'm putting various loras into a dynamic prompt.
So my prompt looks something like this:
tag1, tag2, tag3, {lora:loraname:0.7 | lora:loraname2:0.7|lora:loraname3:0.7}

The issue is that this results in horribly cursed images.
I think I even identified the culprit:
highres fix pass doesn't use the dynamically generated prompt, but the full prompt, combining all loras for the upscale step.
How do I fix this?

[opal jetty]

some errors can be related to your model checkpoints

or try the same settings with 1 lora than add the others to see if its persisting

[quiet raft]

I tested this and same result 😦

Okay, I've got it to kinda work? But once I turned on my negative prompts. It crashed.

[atomic wigeon]

1. prompt a color scheme, color style or theme (it will adopt those colors).
2. Dreamshaper prefers incandescent lighting. Try to prompt iridescent lighting to offset this. Other ideas would be "natural light, natural shadows, bright lit room, dimly lit room, spotlight (color) or well lit.

Dreamshaper (and others) are often just prompt supports that add hidden trigger words to your prompt. At the end of the day you can prompt out of them, while still using them. It takes practice.

Also, save this google doc assembled by Emonz over at Playground. Make sure you're using the correct samplers. https://docs.google.com/spreadsheets/d/1839zMxYbzgrpeN4T7waJ0teZVRhkWhELdu082vHMEgE/edit?pli=1&gid=0#gid=0

hyper-realistic bedroom with modern wall mounted wooden headboard, cherry blossom wallpaper with built in LED lights, white ceiling, iridescent lighting, centered, symmetrical,

DPM++ 2M Karras

In the scene you have incandescent lighting for lamps (yellow trigger).
The rest of the room focused on bright because I used white ceiling and iridescent lighting.

[atomic wigeon]

hyper-realistic scene of a desert landscape at sunset, Fallout 5 style, futuristic metal structures with detailed neon accents,

Use "Fallout 5" to imply updated graphics. Never add texture unless you want to change the entire scene to that texture. You were using two triggers for sand. Desert and sand texture.

[atomic wigeon]

Very good question.

mirror image split screen of a 2D cartoon illustration of a politician

If I try the same prompt in a widescreen format the AI will want to fill in the space.

If you want to use the widescreen format intentionally, try to add "centered and symmetrical" to the end of the prompt. Less is more.

[nocturne hinge]

what did u use for negative prompt

[atomic wigeon]

I didn't.

[nocturne hinge]

aint know way XD

[atomic wigeon]

Less is more.

That one was MBBXL and DPM++ 2M Karras

[surreal rose]

always better to prompt correctly and avoid negative prompts

[quiet raft]

Damn, starting to hate AnimateDiff and why it just keep crashing.

[silver valley]

There is a setting that is needed to be turned on to make animatediff to work

I'm not at the PC to check. But its under optimisations.
"Speeds up the generation when using different lengths of prompts"

[quiet raft]

This?

[silver valley]

Nope

[severe slate]

Hello there, I made the switch from 1.5 to SDXL and I'm getting puzzled by the results. SDXL is supposed to have better quality but all I get is smudgy edges. I've tried changing the latent size, the sampler/scheduler as well as CFG but nothing gives me a crisp image like I get in 1.5? Can my (low) hardware be the cause of that? Is there specific token I need in SDXL prompts that I was not using in 1.5? Full Comfy workflows is in the images Don't know if it's the correct channel to ask for help, sorry if it's not.

[quiet raft]

Here is the full page

[silver valley]

This is it:

[silver valley]

Hey, this is the correct channel to ask. Can you show your settings you used?

[severe slate]

is this sufficient or do you need other informations?

[silver valley]

Try cfg at 7

[severe slate]

CFG at 7 really increase the contrast but does nothing regarding blurriness

Same workflow with SD 1.5 really get the photorealistic feel...

[silver valley]

Strange, I'll send you my sdxl workflow when I'm at the PC to test

[severe slate]

I'll appreciate 🙂 thanks

[silver valley]

here

i also used the sdxl fp16 vae from here: (sdxl.vae.safetensors)
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

[severe slate]

I got this image using your workflow, checkpoint and VAE which is a much better result. I'm trying different combination to see where the difference comes from but my setup being slow as hell, It takes 10minutes to generate one SDXL image especially with 40+ steps 😄
I will iterate from your workflow, thanks

[silver valley]

Nice, no problem 🙂

[rain tusk]

Yes, I did test the Loras individually.
So the prompt
"tag1, tag2, tag3, {lora:loraname:0.7 | lora:loraname2:0.7|lora:loraname3:0.7}"
will result in a fucked up result that clearly applied all 3 loras at once, while
"tag1, tag2, tag3, lora:loraname:0.7", "tag1, tag2, tag3, lora:loraname2:0.7", and "tag1, tag2, tag3, lora:loraname3:0.7" all work.

to demonstrate more clearly that it is in fact the highres fix pass prompt that fucks everything up:

[mellow kite]

Hi pepole, new here, is there a promting chat room for NSFW?

[severe slate]

it is against the TOS #✍🏼｜rules-and-tos message

[cunning sky]

idk where to ask abt models

seems like this is the best place to ask

ok not rlly gonna ask abt models but

im trying to make like a fortnite version of this image

[quiet raft]

Do anyone have any ideas where I could generate prompts from a picture of get prompts by posting my idea? I know you have Interogate clip and Deep Boru thing in the automatic1111 ui. But can I find something else and not be restricted about the prompts?

I have no idea how some people do these long and very good prompts that I can find on civtai and etc.

I am trying to do it with local ai on my PC but will not be as good and it will only be a long line of words.

[atomic wigeon]

Long prompt = bad

Short prompt = good

Try to keep it as simple as possible. There used to be a cloud based site that had image to prompt, but I forgot the name. Sorry.

[shrewd juniper]

Does anyone have an idea how I can achieve a result like the one in the picture, where the AI generator will generate all the wall layers from the perspective of the picture and write down what materials were used next to it? Below is my prompt, which is not quite working.

Create a diagram of a medieval wall showing all layers and materials in cross-section, The layers should be listed and labeled as follows: Outer layer: Stone or brick (external protective layer), Insulation layer: Stone fill or rubble and mortar, Core of the wall: Earth or clay, lime mortar, Inner load-bearing layer: Stone or brick, Inner finish layer: Plaster (internal finish). Include labels for elements like: Wall Walk (top of the wall), Side View (side view), Filling (fill), Batter (wall slope), Surface of Ground (ground surface), Bedrock (bedrock), Make sure the layers are clearly marked and the diagram shows the interior of the wall in a side view.

[quiet raft]

I don't want to have multiple lines, but would be cool to get different prompts from an image instead of just 1.. you know?

So instead of the one stable diffusion offer, it would be great to have options that are good

[sour beacon]

You could use regional prompting or ipadatper to achieve this.

Or Lumina with Area Conditions

[quiet raft]

How can I do it with ipadapter?

[sour beacon]

https://openart.ai/workflows/data_lt/nightscape-and-2girls-regional-ipadapter/IrAoW6SeLDZQhfhbVJMQ
Sorry not much time to explain it so you might simply start with this workflow and replace the loaded images with some you generate within the process.

[ocean terrace]

Hi

prompt: portrait of a fashion female model

Every single result has an open mouth.
All my attempts to get a closed mouth fail.

api url: https://api.stability.ai/v2beta/stable-image/generate/ultra

Furthermore negative prompts increase the problem e.g. negative_prompt: open mouth, visible teeth etc. lead to vulgar open mouths.

I tried keeping the seed and changing the prompt with absolutely no success whatsoever.

Am I using the completely wrong product here or fundamentally fail to understand the most simple prompts?

[silver valley]

hey, if you put closed mouth into negative it makes sense that it wont get better

better put open mouth or parted lips in the negative

[surreal rose]

turn your cfg down

[ocean terrace]

Thank you, I did that but didn't ask properly. I've fixed the typos. tldr: that doesn't change a thing. Any ideas? @silver valley

[silver valley]

try changing the seed. an fixed seed can lock the composition

[sour beacon]

Tried it myself with different models and prompts it seems the ai does not know a woman with the mouth shut 🙂
But honestly neutral face exprerssion, etc. helps at least a bit to avoid the usual sensual mouth open effect on the images.

[atomic wigeon]

Use common facial expressions. For example: smug, smirk, grin, angry, mad, happy, joyful, confident, etc. If you use words like amazed or yelling it will force her mouth open. Make sense?

[uneven mulch]

Guys i need some help, do you have any tips how to generate illustrations styled just like this moose or rhino? I tried hard to generate a Raccon with controlnet but that is best i could do

[ocean terrace]

seed is random

[ocean terrace]

what parameter in the api sets a different model? api url: https://api.stability.ai/v2beta/stable-image/generate/ultra

[sour beacon]

Prompt: sepia simple line sketch of a Moose

You could add dithered or halftone if you need more structure within the line work or screen printing

[uneven mulch]

Nice! that's just what I was looking for, which model have you used? Any Lora?

[sour beacon]

No Lora only the prompt. I tried it with sdxl sd3 and pixart/sdxl workflow. Some work better with screen printing or halftone or dithered as additional word.

[quiet raft]

I need to remove a "choker" so that the neck is free.. What can i put into img2img inpainter to remove it?

[mossy condor]

prompt:a beautiful girl under the raining sky

[teal drift]

i need to make a character to have disheveled hair

[gleaming quarry]

девушка

i need to make a character to have disheveled hair

[surreal rose]

try using the term 'messy hair'

[surreal rose]

Вы не можете генерировать изображения в этом дискорде

[floral mulch]

Anyone know why this is coming out unshaded every time?

[floral mulch]

This is even ture of extremely basic prompts

[surreal rose]

what are you generating in?

[floral mulch]

Automatic1111, but it turned out to be a resolution problem

[soft stream]

hey guys, what negative prompts do you feel are good to use all the time

[dark knoll]

am really kinda stuck. I am trying to make Ranni the witch, but I am getting some pretty nasty stuff like featured above.

I am running with 30 steps and 12 on the GFC scale
I am using Loras and embeddings too, here is my prompt:

masterpiece, best quality, IncrsRanni, Ranni, witch, woman standing in front of the moon, blue skin, wavy hair, cracked skin, extra arms, doll, joints, doll joints, white dress, hat, cloak, 8k, lora:RanniV4LoRA:0.9
Negative prompt: deformed, drawing, sketch, animated, stylistic, bad hands, bad dream, messy face, by <verybadimagenegative_v1.3:0.8>

Anything blatantly wrong?

[dark knoll]

Doll. I bet money it's because of the word Doll

[dark knoll]

Finally got a decent looking Ranni.

She'll make great inpaint practice

[surreal rose]

none. never use negative prompts unless you absolutely have no choice, and then be superspecific about the terms and start with only one

[soft stream]

The prior one it looked like the lora didnt even activate, like it ran purely on the prompts

[dark knoll]

After some in depth discussion and a major understanding of certain details, (Thanks to @surreal rose) I found out what I was doing wrong. :)

[soft stream]

Can i ask what was the reason for the bad image

[quick thicket]

Is there any difference when prompting between | and || in comfyUI and webui?

[plucky herald]

How would I get this into one of my images?

[worthy shoal]

Hello everyone, I am writing to try to understand why many images related to architecture are made with deformations? I have highlighted some of them in red, I have been trying to modify the NEGATIVE PROMPT for days but to no avail... here is what I wrote: cg, cgi, 3d,cartoon, sketch, drawing, anime, ((deform lines)), ((deformed contours)), low quality, (((low resolution))), mutation, jpeg artifacts, ((artifacts)), (((camera deformed))),bad proportions, extra limbs, flooring white.

I use architectureralmix as checkpoint and DPM++ 2M AS SAMPLING METHOD

[dark knoll]

Absolutely! So I've discovered that basically I was majorly struggling with checkpoints and Loras. I didn't realize that a Lora could only work with certain models, and so I think that was a major source of conflict. Once I cleaned things up, the Loras worked, the picture started looking nice, in general it started to just work.

[dark knoll]

I was looking for a very basic Pony model, and believe I found a good starter one. It's called "Pony Diffusion V6 XL". The desire is a very plain, standard model with which I can use to experiment. I noticed though, when going to download it, it had a VAE and a Pruned model version. I was wondering what is the difference between the two and which should I pick? Are there better basic Pony models should use to get a feel for what the model is like?

Bonus question:
I've noticed some Loras or even check points seem to have tags like "SDXL/Pony". This strikes me as odd because it's relating to two differen base models. What does it typically mean if something is being refered to as haveing multiple base model references in their name?

[silver valley]

Pony v6 was the first pony model.
Pony is based and trained on SDXL which answers your bonus question.
SDXL loras can work with pony but pony loras don't work good with sdxl models.

Pony got trained on a dataset where every image got tagged by score and source and rating.
Checkout the model description on civitai for more info on that.

It depends what direction you want to go.
Pony v6 is good for pony/furry stuff.
There are better models based on pony if you want to go for Anime (Nova anime or confetti)
Or for realism (Pony Realism)

[dark knoll]

Im going to have to write that 3rd sentence down! 😁

Definatly I prefer an anime or realistic approach with Pony. I hope it's not completely out of pocket to say I have no interest in generating ponys with Pony, LOL. I'll keep this close in mind, thank you very much!

[silver valley]

No problem ^^ than I suggest using other pony based models

They work much better for anime and realism

[dark knoll]

I will! Thank you for the tips!

[surreal rose]

there are some pony models that don't use score, if you're interested

[dark knoll]

Hmm, I am! It sounds like you have some to recommend?

[surreal rose]

https://civitai.com/models/376450/duchaiten-pony-xl-no-score this one

[silver valley]

Or this:
https://civitai.com/models/549543/chimera

Its a merge model of various sdxls and pony

[dark knoll]

@surreal rose @silver valley thank you both so much!!!

[surreal rose]

🙂 you're welcome

[dark knoll]

Please forgive a silly question, but can a merged model use Loras from places that aren't tagged as such? I.E. This is tagged as SDXL 1.0, but can it use Pony Loras?

[silver valley]

Yes it can work with pony loras as pony got merged into that model

[dark knoll]

I see, so MERGED models can use Loras from other Models. That's huge news, thank you!!

New question, I have discovered a Lora that makes a split screen of faces. It's super cool and I notice that it uses BREAK as a key word. Obviously it appears that it's used as a way of seperating parts of the prompt and saying "do one half here and another half here" but I was wondering if this is a common thing?
I have a few quetions about this Lora as it seems packed full of complexity.

Also... I hope I am not asking too many questions 😅

https://civitai.com/models/380125?modelVersionId=424387 The Lora in question. I also saw it has a window in the description I've never seen on Auto1111. Is that a comfyUI thing?

[silver valley]

Hey, it uses an extension called Regional Prompter.
With that you can define what it will generate in which section of the image.
You can generate multiple different characters with it without them merging together

[green cipher]

I'm trying to upscale some images and hopefully improve their quality with the beta upscaling endpoints. I used the conservative upscale but the output kept all the low quality artifacts. Is that something I need to fix with a prompt or am I using the wrong endpoint?

First pic is original, second pic is conservatively upscaled.

I also tried using the legacy upscaling endpoint and while the result wasn't great it did clean up the artifacts. What should I do here? I'd like a result somewhere between the two

[dark knoll]

Well, I am in it to win it, so I will give it a try!!

[plucky herald]

^

[oblique lintel]

how to achieve this style?

[true wigeon]

Hi. Please tell me how you can achieve such conceptual isometric images? I've tried - isometric view, watercolour, land art, concept illustration, digital rendering, 3 d - concept. That's not it

[past lion]

Hi, I am very new to Stable diffusion (I am just another typical coomer/degen). I am using Automatic1111's webui for generation and Anima pencil XL for model. I am having trouble to make the model output a specific prompt "floor length hair" within my prompt list. The model seems to understand the idea of floor length, but it always output fur/silk texture instead of hair texture. Any advice/help on prompting it to output the right texture would be appreciated

[silver valley]

Try (extreme long hair:1.2)

[lyric musk]

Hii, im new and i need a little help please, im trying to make some pic, and all the time the image have 2 char, how i can fix this?

[silver valley]

Hey, that happens if you try to use a to high native resolution.
1.5 models are trained on 512x512
And sdxl on 1024x1024.
Stay near these resolutions and then use upscaling to get a higher resolution

[lyric musk]

Thanks i ll try that

For a widescreen resolution, wich one should i try?

[trail compass]

how would i prompt for a pose like this? i've tried "(low angle shot/worm's eye view) of a man kneeling and looking directly at the (camera/ground)"

[winter nest]

sup do someone knows this lora or model ? i can give nitro if someone help me find

[tidal fiber]

what type of clothing is this?

[flat onyx]

what lora or model is this?

[gusty dew]

Beautiful picture Anime I like it.

[flat onyx]

yeah, the lora is sooo good

i cant find it

[lyric musk]

Im new with stable diffusion and i need a little help please... idk how to let a good quality on pictures, if i give some zoom u can see the pixels and i can use for walpapper or something, what i need to do please?

[grizzled briar]

hi, I am trying to create a character like Rocket Raccoon in "Guardians of the Galaxy" by this lovely Tibetan fox, can anyone give me some idea or some examples? thx.

[silver valley]

Set hires fix upscale to "by 2"
dont set a custom upscale resolution.
Then set the denois to 0.5 and the hires steps to 10

For FullHD set the default resolution to 960x540

You can then upscale that again in img2img for WQHD

[next niche]

#🏞｜general-with-images A touching moment of रामू and मोहन sitting together in the fields, laughing and sharing stories.

#🎥｜animation A touching moment of रामू and मोहन sitting together in the fields, laughing and sharing stories.

[autumn sandal]

Hi there! Can someone think of a cool advanced prompt for an app I've made? It's a wallpaper changer that, among others, can include various real world variables in the prompt, like:

• battery level
• dark mode enabled or not
• current weather
• the result of an arbitrary prompt to a LLM
• current season
• current time of day

Based on those parameters you can also create a custom parameter, which is a glorified if-elseif-else, basically you can test an expression and assign a different value based on the expression's value. For example, you can use a custom parameter that resolves to green if the battery level is high, yellow if it's medium and red if it's low.

The prompt I'm looking for would be premade and part of the app and I'd love it to showcase the possibilities, I'd like to have some advanced prompt that uses as much of the above (and custom) parameters to create something truly unique. I already have some prompts based on a battery level (a flower that either looks fresh, withering or dying and a city that either looks utopian, neutral or dystopian) and some prompts that include weather, time of day etc., but if someone more creative than me can think of something truly breathtaking, it would be great!

The app's open source, so I'm not making any money on that and am afraid I can't offer any payment.

[bright garden]

When i use img2img + canny controlnet should I set denoising strength to maximum value? I want to not get any color influences from reference image.

[marsh echo]

Hey guys,
I use a1111
Is there a possibility to recreate a character as portrait from a different picture?
I made a character I really like & I'd want more artwork of it. Basically use the original as a reference, similiar to Midjourney's --cref command.

[silver valley]

Hey, that can be done with Controlnet IP-Adapter

[proven hare]

I'm seeming to get a lot of similar images after a while when I use a model. I was told that it's because I have too many prompts likely. My count is 129/150 and 106/150. If someone could point me to a resource I'd appreciate it

[static vale]

Any idea on how to recreate something like this, it's insanely good and there is no metadata on CIVITAI

One more

[molten ether]

webui-user

[elfin pawn]

Is it possible to make nearly perfect, high quality image of "Tiny people flying around campfire without wings in the forest, in night and there is Saturn on the sky" in SD?
That kind of detailed images.
Is it possible?

I am new in this field an I am trying to create something like that but I can't do it. I wanna now if it is possible.
It seems impossible to me right now but I don't know. I wanna be sure about that.
If it is possible, I will continue to try to do that complex image.

I wanna learn SD in depth by creating that kind of detailed images but nearly 2 hours, I can't do it.

[surreal rose]

first question is what version of stable diffusion are you using?

[elfin pawn]

I just download A1111 and with that, SD downloaded. I don't know which version. How to learn that?

[surreal rose]

auto1111 is just the interface. there are many stable diffusion versions. there is: stable diffusion 1.5, stable diffusion 2.0, stable diffusion 2.1, SDXL, and SD3 2b medium

which are you using?

[elfin pawn]

I know. I said, with A1111, Stable Diffusion automatically installed. I don't know which version.

How can I learn the version of SD which I'm using?

[surreal rose]

you'll have to look in your install of A1111 and see what the file name of the model says

[elfin pawn]

There was this model by default. Is this mean that I'm using SD 1.5?

[surreal rose]

yes, that's 1.5. okay, here's a huge issue you're facing. 1.5 doesn't have the prompt comprehension to do what you want, no matter how good your prompt is. SD3 does, but doesn't work well in auto1111 - so to get what you want, you're going to have to do the base image, then inpaint are area in the air, prompt your people in it, inpaint in the sky, prompt saturn in it

and then probably do some upscaling

[elfin pawn]

I did that but it didn't work.

[surreal rose]

not surprised. 1.5 is a good model, but it takes incredibly complicated prompts to get anything that looks good. not the best model for a new user to learn on. and you're asking for an incredibly complicated concept. the AI knows that if something that looks like a human is flying and tiny, it's a fairy and has wings. you might have to do a lot of photoshop work on the image to remove wings and do other things

[elfin pawn]

Maybe Saturn can work but little people... SD just paint some little things or weird people.

[surreal rose]

you'd be better advised to install comfyUI and use that, so you can use SD3, you'll get farther along

[surreal rose]

the AI knows that if something that looks like a human is flying and tiny, it's a fairy and has wings. you might have to do a lot of photoshop work on the image to remove wings and do other things

and you're using SD 1.5 which doens't do good people withotu really complicated prompts

[elfin pawn]

I write wings on negative prompt.

[elfin pawn]

Thank you very much!

[surreal rose]

you're welcome 🙂

[elfin pawn]

So, SD 3 is better in everything?

What about SDXL?

[surreal rose]

SD3 has the best prompt comprehension - meaning it actually understands most of what you tell it. SD 1.5 has the least. there are a number of things that can affect how well the AI understands you, but this is the most important

SDXL does nicer images, and has better comprehension than sd.15, but not as good as SD3

[elfin pawn]

Nicer images than 1.5 or 3?

[surreal rose]

that's a question without an answer, really - you're new to this, so what i'm going to tell you is that you need to explore. start simple. start with a simple concept and prompt. just use the word apple for the prompt. generate. look at it. then try something like: an apple on a table. explore, see what it does.

[elfin pawn]

You're right.

[elfin pawn]

Is 8gb vram and 32gb ram good for sd3?

[surreal rose]

probalby

[silver valley]

Not really. Also SD 3 currently is not worth it.
Better use SDXL models

[elfin pawn]

Why my results are very low quality?

I created first photo and turned it into second photo but I want higher resolution. Did my first photo have to higher resolution? And how to do that?

[wicked sequoia]

Are you using sd3 or sd15?

Sd3 has pretty bad humans. Don’t use it if you want humans. Only use it for backgrounds, animals.

Sd15 base is also pretty bad. I would recommend you use some finetune. Those make much higher quality images.

[elfin pawn]

Sd1.5

[elfin pawn]

ComfyUI or A1111 for SDXL?

[elfin pawn]

Can I change SD version in middle of an image? For example, SDXL for human and SD3 for background. Is SDXL the best version for human images? Do I have to use finetune for all versions? Which version do you prefer mostly and why?

[elfin pawn]

I used an AI website for higher quality. It is better than previous one but it would be the best if I could get higher quality results in creation step.

[silver valley]

You can use both for it. But for SD beginners A1111 is much easier to use.

[elfin pawn]

Is 8gb vram and 32gb ram enough for SDXL?

[silver valley]

Yes

[elfin pawn]

And I need low resolution images at first stage because I need to choose one of them. I need many examples. So my base photo must be low quality. How to do that photo high quality?

[elfin pawn]

The resolution of the first photo here, is normal. But the resolution of the second photo is not what I want.

[silver valley]

You need to use upscaling for getting a higher resolution.

In txt2img you can do that by using the Hires fix

[elfin pawn]

But I choose photo from txt2img and use it in img2img. And I need many examples from txt2img to choose one. So, if all photos in txt2img are high resolution, it will take a long time.

But I can use seed to higher resolution after choosing one, right?

Yeah that makes sense.

But when I use that high resolution image in img2img, will my new photos will be high resolution?

Edits, inpaints and etc.?

[silver valley]

No in img2img it uses the resolution of the sliders

Generate low resolution images in txt2img and if you like them you can upscale them also in txt2img by enabling hires fix

[elfin pawn]

How to fix that turtle's foot?

I couldn't do it with Inpaint.

[elfin pawn]

@silver valley

[abstract valve]

Guys, I'm kind of new at this, and I wanted to make a thumbnail using this image and wanted to make it look better. Any tips on how to do that?

[silver valley]

first get some community made models from civitai.com
then try to gen a few images in txt2image to get into prompting

then try img2img and play around with the denois value

[river fern]

What's different between Hi-Res fixing in the moment VS after the fact upscaling with Tiled diffusion? Even when you use the same ESERGAN upscaler and parameters.

[sinful cobalt]

Anyone know how to prompt to get an overhead type image like the ones found at this link?

https://www.reddit.com/r/StableDiffusion/comments/125he9e/topdown_2d_map_into_isometric_map_is_it_possible/

The first one at the link. I've tried map view, aerial view, birds eye view, satellite view, orthographic view. Trying top-down view rn. Using ultrapixel.

[silver valley]

I would only use hires fix to upscale. Or upscale in img2img with SD upscale script.
Tiled Diffusion is not really different but needs less vram as you can define a tile size like with the SD upscale script.

Use hires fix + tiled VAE to get a high quality image. Then use img2img SD upscale script to get it to Wallpaper sizes

[timid kindle]

Can someone help me get a dark, night time image? I've tried terms like moonlit, dim light, night but most of the times it will look just like late evening rather than proper night which is what I'm looking for. I'm not interested in img2img, remixes etc, I just want a simple prompt

[silver valley]

it depends on the models you use. also loras can help if the model cant do nights good

[timid kindle]

Hmmmm

[thin falcon]

i'm just getting into ai... how do i make sure i'm just talking to the bot and not in gen chat or anything? i haven't used discord much

[north coyote]

Let's say you want a potato and a tomato, how would you prompt for that not to end up with a tomato that looks like a potato?
Do I have to make an image of a potato, clean the background, one of a tomato, clean the background, put them beside eachother and use that as a latent image?

[silver valley]

You can prompt for a potato and a tomato and depending on the models understand skills it will work

[north coyote]

Let's say the model doesn't. How would you proceed?

[silver valley]

Then I would use a lora.
For 1.5 models you need the EpiNoiseOffset Lora or the LowRa lora.
For sdxl you need to use the official SDXL Offset Example Lora

[north coyote]

Alright, thanks. I'll look into that

[wind kettle]

Does prompt structure worth for ai ?

[quick thicket]

Is there any difference when prompting between | and || in comfyUI and webui?

[timid kindle]

How could I stop the AI from generating uphill stairs when the photo is taken from on top of a hill, basically reverse stairs, without specifying downhill stairs because sometimes the generated picture will have uphill angles

[leaden socket]

Advice on these arms?

[static hinge]

inpaint until they are fixed?

[leaden socket]

Hoping for a more automated solution

[static hinge]

afaik there is not a good solution for fixing such details after generating. Best way is to inpaint, so you keep the part you like and re-generate the part that is bad.

If you make another gen with same prompt, model, probably you won't get bad arms, but you won't get that image either

if you use same seed, and play with different steps, or slighted different prompt, you might get that fixed but you would get a different image, and it migh ruin the thing you liked about that gen

I mean yes the only "automated" solution would be to regenerate with same seed and change something, but as I said the result is also random

[leaden socket]

i mean I'm looking more for a lora, or embedding.

[static hinge]

or to use some Adetailer, that is and automated inpaint for clothes or people. But it might not fix it.

[static hinge]

nah, that happen on that gen, and probably won't happen in the others

there is no "fix"...

everything you add to your workflow will have consequences

[leaden socket]

"On gen", no that gets applied on the k sampler.

[static hinge]

if you add a Lora for "good anatomy", it will come withstyle changes, poses changes, etc

[leaden socket]

It conerts the noise into what it things is a arm

[static hinge]

you should be "thankful" for the good generation and then inpaint the part that is bad 😆

there are not automatic fixes, a slight change will change nothing or change everything

in the generation process

[warped bridge]

ffb

[exotic mural]

Hi, I really need some help for the prompt,
i have been trying to make a picture look like this, but most of the time the characters just standing back-to-back, or stick together (like just beside each other).

please kindly advise, I am very new to SD, still learning. using Draw Things on Iphone.

what If I want both guys to look like shadow, standing just like the photo?

this is the prompt I have been trying to make the image.

(8k, RAW photo, best quality, masterpiece:1.2), (realistic, photo-realistic:1.4), ultra-detailed, (2 young asian male Japanese celebrity),perfect detail , make up,(upper body shot:1.1), a man 22 years-old, one standing on the right, another standing on the left, both facing backward, body looks like shadow

thank you so much and sorry if I post in the wrong area.

[sour beacon]

Normally you should be able to achive the result by using controlnets. Use an openpose model to get the same body pose. Or try another prompt where you include the direction.
"Silhouette of two muscular man looking towards the landscape "

[exotic mural]

Thanks, I will try that later on.

[leaden socket]

Do a image to image

[exotic mural]

Thanks, finally made something similar to what I want 🙂

[rich shell]

how to have clip skip everyone?

[silver valley]

Go into Settings - User interface - Quicksettings.
There add clip_stop_at_last_layer
Then hit apply and reload ui

[rich shell]

oh nvm i did this
thank you sir ❤️

[rich shell]

also idk if i can change image origin pose by controlnet openpose
last time seem not work
or it just work on t2t label

[rich shell]

also how can i make pointing line to the posing hand

[dark ether]

Do you guys know good prompts for waist up photos? I keep getting full body.

Esp on sdxl

Also, while I'm here, trying to balance the weights between side and front views is killing me - is there a better way to get a perfect 3/4 view for the whole body?

[rich shell]

idk how they can make detail hand like this, this is what best i can do, i hope there's someone can help me about makeing detailed hands, thank you

[dark ether]

can't go wrong with making a 3D object and running it through img2img with canny/openpose hands

the best AI art isn't 1 shot, it goes through a bunch of iterations

[rich shell]

wow i don't know the left stuffs was 3d image

[rich shell]

i hope you can tell me or have some source about it so i can learn sir

[dark ether]

Could be, could not be, all I can say is that when I do it, that's what I do to get consistent hands

I also hand paint details like that if the output is really bad

and again, run it through SD again and again until it looks good

[rich shell]

yes sir

[sour beacon]

Hi there just tried out 3 prompt with a sdxl model:
1/2 body shot of a woman, long hair (first image)
3/4 body shot of a woman, long hair (second)
3/4 body shot of a woman, black shoes (third)

As you can see the models seems to know the image "cuts" of 1/2, 3/4 quite good. But more important is to add details of the region you want to see in the image. So hairstyle etc. will help you to get the top of the image defined. You could write about the waist line or the belt to define the lower part. If you need the upper parts of the legs maybe add a description for the thighs

[dark ether]

Yup! found out about INCLUDING the upper part of the body, but EXCLUDING The lower part is hard... maybe somethign about thighs in the negative prompt?

[rich shell]

idk but how to change output image size

[versed garnet]

what is causing this in mk2 outpainting

[dark ether]

I dont know the proper name or cause for this, but sometimes changing the prompt (usually lowering the number of tags), removing a lora temporarily (remove the lora and run the generation to completion, then add back), changing model, restarting SD, or restarting your computer may help. Also, if there are any errors (eg unclosed parenthesis) they can cause this.

[mighty sage]

i give
there isn't a checkpoint or lora out there (including flux) that knows what a pickaxe is
looking for ideas without using a controlnet

thoughts?

top down view knolling layout, futuristic style sharp double pointed miners pick tool (pick mattock:1.2) used to mine ore underground, the points extend outward from each side and have a slight downward curve that would strike rock, wooden handle, glowing metal, symmetrical, laying flat on a cracked slate rock

looking for something like this

those use a controlnet

[warped bridge]

vvdv

[tired vigil]

Hello, I have a prompt from automatik1111 which uses a "style type" lora I'm unable to locate i lora-folder. No binaries, manual attempts to find via hash has failed. Because of this missing lora I'm unable to reproduce the images in comfyui. If I exclude this lora, the both automatik1111 and comfyui produces the same. Any ideas?

[silver valley]

Why not copy over the lora to Comfyui?

[tired vigil]

I was unable to figure out which lora file it was. Figured it out. Changed a setting in automatik to use filename instead of aliases, and then after creating new images with same prompt in lora, comfyui was able to resolve the missing lora correctly in a work flow.

[rich shell]

i'm trying impaniting for better hand but seem it even worse, can someone help me

[chrome shard]

I’m having trouble coming up with a prompt. The idea is Jin of bts is a mutant from xmen who has similar powers to Jean grey. Can anyone help

[chrome shard]

Can anyone help with a prompt for this hairstyle

Whenever I try to describe it it keeps giving me shoulder length or the wrong length and color and style…

[sour beacon]

@chrome shard do you try inpainting, or anything with a controlnet or source image. Or a simple prompt?

[chrome shard]

dont know what prompt to use it for though

[sour beacon]

How are you trying to achive the result.

Are you uploading an (private) image and use inpainting or other tech or are you creating the image from scratch with a prompt.
"asian man with ... wearing ... standing in. ..

[chrome shard]

im just trying to get the hair cut right is all

[sour beacon]

an asian male 20 years with pixy cut and black hair,

[chrome shard]

hold on ill try that

[sour beacon]

an asian male 20 years with side parting haircut and black hair,

[chrome shard]

ill try that though the person im using isn't 20 lol hes 31 and just got out the military lol

[sour beacon]

an asian male 30 years with short hair style haircut and black hair, wearing uniform

[chrome shard]

and hes korean lol yeah ik alot of confusion

[sour beacon]

a korean male 30 years with short hair style haircut and black hair, wearing uniform

.... and he has a dog pet friends called "walter the fish", had a bad expierence with drugs at the age of 23 and prefers rock music.....

[chrome shard]

yeah i tried that hairstyle still didnt work....

with the prompt for it

[sour beacon]

Yes because most models, loras, settings, seeds, work different. So the take away is, it is possible to get all kind of hair cuts but not from all models.

"top fade haircut"

[chrome shard]

let me try the top fade one

[verbal mauve]

how to use im not sure hahaha

please help me 😛

I bought a suvscription but where to prompt that thing here on discord ?

[sour beacon]

I guess you get help in this channel: #🗣｜artisan-support-feedback

[verbal mauve]

says prompting help here haha

HOW DO YOU PROMPT STABLE DIFFUSION HERE ?

[sour beacon]

Well obvious the automod decided that i can't help you. May the force be with you. Maybe read #artisan-faq and try to see the difference between a prompt and the process of the image generation from an input prompt.
This channel is to help people with the creation of prompts.

[tired vigil]

Should I put commas between the loras?

[rich shell]

idk why those black things keep swarming tho i;m not input it, how to fixx that

[abstract folio]

Howdy! I am new to Pony, and really XL in general. I am unsure what it is I need to do to improve my image quality. It is significantly more different from 1.5 than I expected

[prisma grail]

what's up with "core_9, score_8_up, score_7_up, score_6_up"? these prompt keywords don't make any sense!

[sour beacon]

Howdy to you,
i decided to answer you by citating the caption of the model description (pony models):
pony-diffusion-v4 - "same, but different" edition
pony-diffusion is a latent text-to-image diffusion model that has been conditioned on high-quality pony, furry and other non photorealistic images through fine-tuning.

For easier understanding: and other non photorealistic images.

So the base models of all the different pony veriants based on model training bei anime, comic, manga, hentai, ... and is optimised for these.
Maybe you look at realvis or other basic sdxl models if you can live without the hole kinky stuff...

[silver valley]

Hey Sdxl and pony are trained on 1024x1024 native resolution.
So don't generate at 512x768

Thats why the images look bad

1.5 models are trained on 512x512

[tired vigil]

Or use a properly trained version of PONY https://tensor.art/models/751621497568661037/PONY-Dream-Diffusion-By-DICE-v1

[silver valley]

Those are only for pony models

[prisma grail]

I see..

[tired vigil]

any ideas on how I can fix this?

pony

shot it dead

Im using easy negative, zpdxl

[tired vigil]

I wish I could

no weapon lora though...

everything works well other than the faces

pony model was a hoe to train and even after that i still hated it. I simply dont use it. its just out there for people that love a bit of pony

[tired vigil]

it just has a lot of loras for some reason

[silver valley]

What settings do you use?

[chrome shard]

Need help making a zombie. Like a movie dawn of the dead zombie. Can anyone help

[abstract folio]

Thanks! I’ll test out what y’all said tomorrow

[tawdry flare]

imagine

[pulsar escarp]

Anyone knows what prompt to use to get a character to cross its arms in its BACK ?

[tired vigil]

which settings do you mean?
there are a lot of them

[tired vigil]

I found the problem
I would have never thought that image size impacted the generation
I was running 512^2

[real echo]

Does anyone have any idea what this pose is called? Not necessarily on the ceiling, but the type where you hold yourself up against two ends with both your feet and hands.

[carmine iris]

Try chimney climbing technique, or chimneying. It's not specifically the suspension you're asking about, but it may set you on the right path.

[warped bridge]

fh

[plush shoal]

Hi! any ideal how i can fix this ? Is it made with sdxl with the replicate api with this prompt : An energetic young woman in workout clothes, jogging in a lush green park with a determined look on her face, sweat glistening on her forehead

What I find realy strange is that with the web version i have a way better result with the same prompt:

https://replicate.com/p/p1nd0a92s1rj40chd98a67n47m

prompt = (
f"Scene: {prompt_phrase}. "
)

input = {
"width": 1080,
"height": 1920,
"prompt": prompt,
"scheduler": "K_EULER",
"num_outputs": 1,
"guidance_scale": 0,
"negative_prompt": "worst quality, low quality",
"num_inference_steps": 4 ,
"disable_safety_checker" : True,

}


output = replicate.run(
"bytedance/sdxl-lightning-4step:5f24084160c9089501c1b3545d9be3c27883ae2239b6f412990e82d4a6210f8f",
    input=input
)
return output[0]

Here is my code for the api, maybe there is something wrong here

[faint bolt]

can someone help me - why doesnt my confiui workflow work?

i do not know what to conect to latent image

i wanted to generate simple pictures using promts - nothing to complicated

[silver valley]

Hey, sdxl is Trained on a 1024x1024 resolution. Everything to high above that causes duplicates.
You need upscaling to get larger images.

[rich shell]

why my img become so bad. yesterday it still good without changing anything, like smh it become blurry

[plush shoal]

ok it makes sense! what is the best model for vertical pictures ? ( "width": 1080, "height": 1920)

[silver valley]

You need to use sdxl + hires fix to get that

Only flux can generate at larger resolution

[crimson patio]

guys what is best resolution to flux for potrait?

i trying sdxl reslotuions but idk, its looks worse than 1024x1024

[sour beacon]

Simple click on the latent Input (pink input sampler) and drag it somewhere on the grid. It will ask you which node you want to add. Use empty latent.
(You could also simply add a empty latent node and connect it).
Then enter width and height in the empty latent widget.

[faint bolt]

i did not know this - got any other tips and tricks ?

[modern bridge]

Can we use adetailer for anime generation ?
Because when i try it it just ruin and Blur the picture
Thanks

[rich shell]

idk where to get Counterfeit vae on there, pls help me

[silver valley]

Yes

It works fine normaly

[modern bridge]

Idk what to Say it's not the case for me

[silver valley]

You would need to show your txt2img settings with an example

[modern bridge]

Okay, i'll do if idk when because SD burned my gpu wire

[livid oriole]

SD can't generate a centaur

[cobalt owl]

Has anyone been successful with simple prompts that do a few colors rather than ombré or blended colors? I’m creating stickers and I have to do a lot of photoshop editing to reduce the colors and clean up the edges. Just looking to simplify my process.

[atomic wigeon]

Put an image to image in davinci resolve and put color generators on top (or a million other options). Now export and throw back into the AI I2I.
Blackmagic Davinci Resolve is your best friend; and it's free. It blows photoshop out of the water when you know learn how to use it. It's quick and easy.

[atomic wigeon]

Otherwise try the words "muted colors", inverted colors", "chromic" and other photography lenses that dull colors.

This one came out poorly (I'm short on time). But just an example of using photography terminology.

[real echo]

anyone has any idea of whats called this pose under a blanket or any cover? been trying for so long and dont manage to make something similar

[proven jasper]

its under_covers https://danbooru.donmai.us/wiki_pages/under_covers

[real echo]

Well works, but not the way I want, it always turns them to be the whole body under sheets, not just from the mouth downwards, also most times is laying on the belly, need to put (on_back:1.4) to look lay down on back

[proven jasper]

when nothing else works u can always try with a lora like this one https://civitai.com/models/365238/comforter-sdxl

[real echo]

Going to try thanks

[warped bridge]

vbfd

[faint bolt]

so in using confiui - when i generate mulitiple pictures - how do i save them all at the same time? - im already using save picture/s node but i still have to manauly save them 1 by 1

[cobalt owl]

Chromic worked well! Thank you!

[atomic wigeon]

Outstanding. Glad to hear it!

[mystic stump]

I'm trying to generate an image with this prompt using Flux but all I'm getting is a picture of a house with lightning above it. It looks very realistic but I won't add the dome over the house:

"In the midst of a massive lightning storm, a typical residential home in Florida, stands under a protective, glowing dome of energy. The dome, a visible forcefield or bubble, envelops the entire house, clearly separating it from the chaotic storm outside. The dark skies are alive with intense lightning bolts, but each strike that comes close is deflected or absorbed by the dome, causing bright, electric ripples to spread across its surface. The dome is the central focus of the scene, shimmering with power and creating a stark contrast between the turbulent weather outside and the safety within. The house beneath the dome is a classic Florida-style home, secure and untouched, thanks to the powerful energy shield that protects it."
How could I improve that prompt to achieve the desired results?

[compact maple]

What's the proper prompt formatting to explicitly isolate prompts when i include say 3 characters and have it only give X prompts for X character?

[tired vigil]

Im using sdxl i keep only getting a looking up at person angle from the ground any prompts to make it stop doing that? or a lora that does many angles?

[sour beacon]

Well check the rest of your prompt. In general i would say the default camera angle of viewpoint is pretty much on chest high as most training images are candids or photoshoots.

[tired vigil]

i tryed putting like portrait angle and still same thing it like refuses to do diffrent angles

[sour beacon]

Usually the easiest way if adding specific angles or description like "upper body shot" don't work is to add descriptions of parts that only would be visible if the camera angle changes. Like describing parts of the pants will make sure you will get more chance of a full body shot.

[tired vigil]

ok ill try thanks

[tired vigil]

Hi Folks, I am using ComfyUI, but I also have Automatic1111 available -- I can download any models needed.
If I were to want to develop a character -- lets say a badger, who will be the central character in a story, and I want to do illustrations for the story, how do I convince the AI to use the same character over and over in different illustrations that have him in different positions, different settings, and different clothes (fantasy story) etc

[tired vigil]

hi if you use a well trained checkpoint just fix the seed then change the prompt slightly for clothes etc

hi @tired vigil can you recommend one of the checkpoints that is "well trained" -- im pretty new to this

[tired vigil]

yes bro use this one https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1

[sour beacon]

Consistent characters are tricky with ai. If you want to avoid inpainting, face swapping, ipadapter, controlnets the only option remaining would be to train a Lora. That would need you try to generate as much similar images from your character (use a prompt with much details for the character and change only position or background with 3-5 words). Keep all images in which your character looks similar enough. Then use these images to train a Lora for your character.

[tired vigil]

ok, makes sense -- thanks

[wicked wren]

Is there a way to convert a photo taken with a camera (such as a person or an object) into an image similar to the style of an illustration, watercolor, etc.? I'm trying to play with the img2img settings, but it's not working well. Sorry for my bad English.

[tired vigil]

yes img2img and use a anime style checkpoint and raise the denoise to aroun 5

[faint bolt]

can my 4060ti 16gb run this?

[tired vigil]

yes you can my man

[faint bolt]

confiui i expect?

[tired vigil]

It runs on Forge ui, Compfy and swarm

thats it on forge

[faint bolt]

do you maybe know how to use confiui?