📝｜prompting-help

[static hinge]

At so point I become frustrated with this new models as I couldn't run them with my RX 6700 12 GB VRAM, even in Ubuntu. Maybe in the future I would try to run them again.

I wonder if they would perform on Windows and I can ditch Ubuntu for good xD

but I started with RX 480 and Ubuntu was like heaven

Also Flux and SD3 for image generation alone took ages

SD XL was some perfect niche for me (!?), generations time and quality was really confortable

and Comfy gave me a hard time as I was used to Automatic1111 that everything that was suposed to work worked most of the time

[unkempt mason]

are there LoRAs that help with generating better Eyewear (Glasses, Sunglasses)?

[silver valley]

If you use Auto1111 or Forge Webui you can use the Adetailer extension and then you can load the sunglass model from Civitai.com into it

[proper quest]

Anyone here who understands [sep] [skip] for adetailers?

[willow shell]

Auto111 better or forge for beginner and learn to pro from there ?

[silver valley]

What's your GPU?

[willow shell]

Rtx 3070

[silver valley]

Mostly Forge is better as its more updated and gives a bit better performance

Then I would go with Forge

[willow shell]

Ots prompt base or node

I just one to learn from beginning so ....asking recomendation and learn from zero

[silver valley]

Prompt based with the same ui as auto1111

[willow shell]

Setup is same as auto

[silver valley]

Mostly, but it has different launch args

Checkout the first link of the pinned messages in the #🤝｜tech-support channel for all the install Guides

You can also setup both Auto1111 and forge and compare with what you like to start with

[willow shell]

Hello i cannot install reactor on extension why......it is removed ? Whats the best alternative

[viscid thistle]

hello everyone, can someone help me. i generated good picture but it have bad eyes, someone have a guide, promt or can someone help me to generate good anime eyes USING INPRINT on stable diff? thank you for answer !

change 1st to second or just for better eyes

[silver valley]

Hey, which webui do you use?

[viscid thistle]

easy diffusion

[silver valley]

Ah okay, other webuis like Auto1111 or Forge have an extension that automatically fixes eyes.
If easy diffusion supports inpaint then try mask the eyes. Set inpaint area to mask only if available. Then set the denois to 0.5 and generate.

[viscid thistle]

i tryed but idk eyes are shitty anyway, i tryed a lot of promts and loras idk what i am doin wrong

[viscid thistle]

and there are no denois and inpaint area in easy

[silver valley]

Then you should checkout Auto1111 or Forge Webui

[viscid thistle]

as i know easy works on auto111

[silver valley]

Easy is a webui itself like Auto1111

You have to install Auto1111 separately

[viscid thistle]

ok got it

[silver valley]

A guide for that is in the first pinned message of #🤝｜tech-support

[astral trench]

Hey guys, does anyone have a guide or extension (for forge webui) with all quality tag and his effect ? Thank

[nimble shuttle]

does anyone know why my images dont come out as good as the example images for the lora lol. im new to stable diffusion and ai image generation so anything helps

[proper quest]

First, you have a fk ton of useless negatives (unless specifically stated to use some of them on the model page).

Then if u r on SDXL u should use 1024x1024 as base resolution (that means any res containing the same amount of pixels when multiplied, example 1152x895),
(ignore this if ur model is SD 1.5, which it is)

Then you should set the sampler to karras

[proper quest]

just use masterpiece, best quality, absurdres, highres

@hearty kayak

@pearl pilot

[nimble shuttle]

when you said i should use karras, this is the option you were talking about right?

[proper quest]

ye

also, try using less CFG, like 5 for starters

[nimble shuttle]

okay, i put the cfg down to 3, removed the negative tags (i only used those because it was in the video i was watching) i'll play around with the prompt more and see if the results are any better

[proper quest]

you dont need to remove all of them

[nimble shuttle]

with the negative prompts

without it 😭

[proper quest]

^

also dont use cfg 3 try 5 for starters

[nimble shuttle]

what are some negative promts i should keep then? i don't know what is and isnt good

[proper quest]

remove the bodly parts in negs, and instead prompt portrait and head only

[brittle stream]

hey there....when using flux (i use forgeui) does the lora location in a prompt make any differnce, especially when using more than 1 lora? i ususally put the all at end, but when i moved them to start i think it makes them more obvious, but that may just be me.

[silver valley]

Hey, in Forge / ReForge and Auto1111 which forge/reforge is based on the prompting is so that the earlier (more to the left) a word is the more attention it gets

That works also for loras

[brittle stream]

thank you! So its more down to the UI than the model you are using by the sounds of it. Very good to know 😉

[silver valley]

NP, also you can adjust a loras strength by changing the number at the end of it

If a lora is to weak just go for 1.2 and if its to strong then try 0.8

[brittle stream]

i will try this

[inland mantle]

Some loras have very aggressive styles, so if they are changing the overall base style of your model too much you can go as low as 0.4 and still get some effect.

I've even got a couple that you WANT to set at 0.1 lol

[dapper vault]

can someone help me, im trying to create an image where a katana is embedded into a wall of marble, the problem is just nothing really works.
this kind of marble as in the picture i sent,but maybe more like unevenly uncut, i tried some different stuff with loras new prompts etc but i never really works.

[torn swift]

anyone know how to fix this?:

RuntimeError: mixed dtype (CPU): expect parameter to have scalar type of Float

[inland mantle]

thats a question for #🤝｜tech-support i think

[torn swift]

I sent it in there too but some rando who probably wants to steal my discord account answered so im just waiting

[proper quest]

are u on integrated gpu

[torn swift]

no my integrated gpu is disabled, im on my actual GPU

AMD Radeon RX 7900 XT

[proper quest]

does this show up in the cmd when u try to launch it>?

[torn swift]

lemme check

This the entire startup

Already up to date. venv "C:\Users\ramon\Desktop\SD-Zluda\stable-diffusion-webui-directml\venv\Scripts\Python.exe" WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next. Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr 5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)] Version: v1.10.1-amd-37-g721f6391 Commit hash: 721f6391993ac63fd246603735e2eb2e719ffac0 ROCm: agents=['gfx1100'] ROCm: version=6.2, using agent gfx1100 ZLUDA support: experimental ZLUDA load: path='C:\Users\ramon\Desktop\SD-Zluda\stable-diffusion-webui-directml\.zluda' nightly=False Skipping onnxruntime installation. Installing sd-webui-controlnet requirement: changing opencv-python version from 4.11.0.86 to 4.8.0 You are up to date with the most recent release. C:\Users\ramon\Desktop\SD-Zluda\stable-diffusion-webui-directml\venv\lib\site-packages\torch\cuda\__init__.py:936: UserWarning: CUDA initialization: CUDA unknown error - this may be due to an incorrectly set up environment, e.g. changing env variable CUDA_VISIBLE_DEVICES after program start. Setting the available devices to be zero. (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\pytorch\c10\cuda\CUDAFunctions.cpp:109.) r = torch._C._cuda_getDeviceCount() if nvml_count < 0 else nvml_count no module 'xformers'. Processing without... no module 'xformers'. Processing without... No module 'xformers'. Proceeding without it. C:\Users\ramon\Desktop\SD-Zluda\stable-diffusion-webui-directml\venv\lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: pytorch_lightning.utilities.distributed.rank_zero_onlyhas been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it frompytorch_lightning.utilities` instead.
rank_zero_deprecation(
Launching Web UI with arguments: --use-zluda --update-check --skip-ort
Warning: caught exception 'CUDA unknown error - this may be due to an incorrectly set up environment, e.g. changing env variable CUDA_VISIBLE_DEVICES after program start. Setting the available devices to be zero.', memory monitor disabled
ControlNet preprocessor location: C:\Users\ramon\Desktop\SD-Zluda\stable-diffusion-webui-directml\extensions\sd-webui-controlnet\annotator\downloads
2025-06-23 14:21:19,667 - ControlNet - INFO - ControlNet v1.1.449
Loading weights [a810e710a2] from C:\Users\ramon\Desktop\SD-Zluda\stable-diffusion-webui-directml\models\Stable-diffusion\waiNSFWIllustrious_v130.safetensors
[LyCORIS]-WARNING: LyCORIS legacy extension is now loaded, if you don't expext to see this message, please disable this extension.
2025-06-23 14:21:20,032 - ControlNet - INFO - ControlNet UI callback registered.
Creating model from config: C:\Users\ramon\Desktop\SD-Zluda\stable-diffusion-webui-directml\repositories\generative-models\configs\inference\sd_xl_base.yaml
Running on local URL: http://127.0.0.1:7860

To create a public link, set share=True in launch().
Startup time: 31.5s (prepare environment: 44.0s, initialize shared: 0.3s, other imports: 0.7s, list SD models: 1.1s, load scripts: 1.6s, create ui: 0.6s, gradio launch: 0.6s).
Loading VAE weights specified in settings: C:\Users\ramon\Desktop\SD-Zluda\stable-diffusion-webui-directml\models\VAE\fixFP16ErrorsSDXLLowerMemoryUse_v10.safetensors
Applying attention optimization: InvokeAI... done.
Model loaded in 10.7s (load weights from disk: 0.5s, create model: 1.0s, apply weights to model: 7.3s, load VAE: 0.3s, load textual inversion embeddings: 0.1s, calculate empty prompt: 1.2s).`

[proper quest]

i dont see anything wrong in this

[torn swift]

yea

[torn swift]

tho if i try to start any generaration i get that error

[proper quest]

sooo what the issue then

forgeui?

[torn swift]

i have no idea, i deleted the venv folder and reinstalled it that didnt help

[proper quest]

well i mean what ui r u using

[torn swift]

uh, do you mean stable diffusion?

[proper quest]

well yes but what ui

comfyui, a1111, forge, swarm

[torn swift]

a1111

[proper quest]

and lemme see ur generation settings

[proper quest]

remove the vae

[torn swift]

still the same

[proper quest]

is that an sdxl model

cuz from what i know it is

[torn swift]

it doesnt work with any checkpoint, i cant use anything

[proper quest]

im sure a1111 is a bit outdated by now

since most models are trained on sdxl

you can try using --no-half-vae command arg

[torn swift]

might need to look into comfy ai

[proper quest]

or --no-half

(as in one or the other)

[torn swift]

I'll try

[proper quest]

anyway, forge is basically a1111 but newer/better

so its same simplicity, comfyui..well its not comfy rly, its imo a bit higher on the skill usage

[silver valley]

Pls don't use --no-half

I already told you the fix to the issue in #🤝｜tech-support

[proper quest]

or that ^

[willow shell]

Hi there,is there any sd prompt generator , free or paid one , not prompthero and online based, i heard there are many like andrewongai etc......

Any experience with those any

[proper quest]

im not sure what ur question is

[willow shell]

Is there any best prompt generator for stable diffusion besides chat got ,deepkseek,claude.

[inland mantle]

He wants to know if there's a prompt generator that spits out random prompts other than asking an LLM that is free. The answer is no, not yet.

[river fern]

is there any method or trick the make the generation focus on a certain set of colors?
I imagine with two its easy, but im intending 3. so yea

[sharp night]

The wildcard extensions might be of interest - some are linked within webui in the extensions tab

[molten spindle]

what chatgpt prompt to make a realistic portait picture that seems real and not doesn't have so perfect skin that it will not be so obvious

[proper quest]

as bro said wildcards or use {color1|color2|color3} and it will randomly pick one of them

[balmy vale]

I'm trying to use regional prompter on auto1111 but it doesn't seem to work, it's like the thing isn't even activated, i have the same result with and without it

i tried with BREAK, i tried with the ADD thing

with base prompt and comon prompt

none of the thing in the tutorial works, it all blends in instead of creating that image

the regional prompter is activated, the box is checked

i tried simply having 4 characters with simple columns too next to each other, doesn't work either

[hollow tapir]

show the whole webpage please

[balmy vale]

what exactly do you need to see cause i can't take a pic of the whole screen

i tried using mask too but i'm not sure how those works, couldn't find a tutorial on them

[silver valley]

We need resolution, steps and other settings above regional prompter

Also an example output or the prompt too if its sfw

[balmy vale]

i did give an exemple of the prompt above

also here :

like when separating 4 character, it just blend them like i didn't use regional prompter at all, i see no difference when activated or not

[silver valley]

Ah okay

[balmy vale]

is that enough?

also i made sure it's up to date

[silver valley]

Its enough, can you try just to split the image into half with 2 different characters?

[hollow tapir]

What the prompt? EDIT ok there it is #📝｜prompting-help message I havent played with regional prompter in a while I ll try to remember and fiddle with it after work (6/7 hours)

(It s easier if you screen the whole webpage)

[balmy vale]

no rush, thanks a lot

[hollow tapir]

what do outputs look like ?

[balmy vale]

but it really looks like the thing is just not activated

[balmy vale]

i'll post one in a bit, i'm cooking right now

[balmy vale]

here's an exemple using 4 columns :

BREAK doesn't work either

same result when i deactivate it, so it's really like the extension is just not applied

[inland mantle]

What model are you using? I know some struggle with complex prompts or compositions with multiple characters etc. Also, if it's not a NoobAI or Illustrious model, it likely won't have strong enough training on just character names. Another thing to consider is the aspect ratio of the image. With 4 evenly spaced columns you'll sometimes need a slightly wider aspect ratio, otherwise the model tries to cram two characters into one column at times. Regional prompting is finicky at best and most tutorials sell it as a guaranteed thing, when in reality most of the time it won't work (personal experience), though some people have reported better success using character loras in each column as well

[balmy vale]

it's illustrious, and you need larger than usual aspect ratio? Cause with normal prompt you can get a lot of character in

also do you know how to use mask? I read multiple people said it worked better with it than matrix (column and rows)

[elder spoke]

tangent: since you're using illustrious you definitely don't want to be on 768x512
these are the usual sizes you should choose

[balmy vale]

oh, been using that size for a while and worked pretty fine, lol, but thanks for the info!

[hollow tapir]

illustrous is sdxl based if I remember so yeah aim for something close to 1024x1024 (+- 30% in any dimension)

sd1.5 => 512x512
sd2.1 => 768x768
sdxl and above => 1024x1024

[balmy vale]

you say +30% for regional prompting you mean?

[balmy vale]

i managed to make it work somehow, but even with bigger aspect ratio, i still get some character blending together... I don't get how other people are doing it so perfectly (and a lot of images in a row, so it's not like they're only getting a good result once in a while)

[hollow tapir]

I mean you can deviate 30% of the training resolution of the model in any direction without getting any degrations / duplicate artifacts.

so for sd1.5 models you can aim for 512x768 pretty safely for instance.

[balmy vale]

i see, thanks

i wish i could use this damn regional prompter well, i'm so jealous at what i see everywhere else, lol

[balmy vale]

even with only 3 characters, and while the separation works, some of the settings, like (smirk) get on all characters instead of just the one where i put it... is there additional setting to weight the blend between the region or something?

[silver valley]

quick test, works like that, no spill over, each character has their clothes, colors and expressions

[balmy vale]

can you copy past the prompt so i can try it myself? see how similar we get

[silver valley]

im on AMD so dont expect the exact same image

masterpiece, best quality, sitting on stairs, in park, intricate iridescent, <lora:gnd-sdxl:0.5>, summer, evening, sundown, autumn leaves, windy, BREAK 3girls, rias gremory, blue torned shorts, black croptop, medium breasts, puffy lips, blushed, crossed arms, evil smug face, BREAK 3girls, esdeath, slim fit, straight hair, long hair, blue hair, blue eyes, pleated skirt, white shirt, medium breasts, looking down, covering privates, surprised, BREAK 3girls, saber \(fate\), slim fit, straight hair, long hair, blonde hair, green eyes, wearing armor, holding sword, angry, attack,

and neg:
blurry, deformed, worst quality, ugly, 3d, realistic

[balmy vale]

yeah i know it wont be the exact same, but i just want to see if the expressions blend in or not

i don't get it... it works here, and the settings compared to mine are mostly the same, i dunno why the smug expression kept going onto the other characters

[silver valley]

Did you had Adetailer enabled?

[balmy vale]

i don't have that

[silver valley]

Ah okay

[balmy vale]

i'm losing so much time on this and i'm never getting the results i want, it's depressing, lol

[inland mantle]

Welcome to the wonderful world of AI image generation lol unless your using Flux or hidream, then you get exactly what you ask for, for better or worse.

[balmy vale]

i'm sure there's a way to get something more consistent, i just can't figure out how

[balmy vale]

trying to follow a guide on mask for regional prompt, but i get a "IndexError: list index out of range " error

the guide says to activate both base prompt and common prompt for the mask, but it cause this error

turning off base prompt make it work again, but then the mask doesn't seem to work at all

[nimble yacht]

Hi, I’m looking to install a prompt-enhancing extension for ComfyUI. There seem to be several — Flux, Prompt Expansion, Prompt Enhancer, etc.
Which one are people here actually using and liking right now?
I want something that helps improve phrasing without bloating my system.
Any solid recommendation?

[runic hare]

id:browse

[balmy vale]

Is there a way to randomise the groups in regional prompter? Like if i have a bunch of characters in certain positions, but i want sometime character 1 to be in column 1 and sometime column 3

is there an option for that or do i just have to use dynamic prompt and hoping all characters will be different?

[tired vigil]

Does anyone have an LLM prompt that can create random image prompts continuously? I like being surprised with different unplanned concepts

[quiet zodiac]

It's not LLM, PromptCrea node can kind of do that. You could feed that result into Searge LLM.
https://github.com/tritant/ComfyUI_CreaPrompt

[subtle cove]

game？

[subtle cove]

66

[subtle cove]

https://tenor.com/view/mochi-mochi-cat-mochi-mochi-peach-cat-yes-nod-gif-9552934121675609737

[humble lynx]

Hi, @everybody

I have an AI image effect issue now and looking for your hellp.

There are some edited phots of original photos and the effect is similar to sketch.
I want to build workflow to upload original photos and get same effected photos.

How to build this kind of workflow?

These are edited photos

And this is original photo example.

Help me to find solution: which is the best model and how to fine-tuning it.
If you can send edited photo of this photo, it would be great

[spice ore]

Is flux kontext discussion allowed here?

[proper quest]

why wouldnt it be

[glad patio]

Guys how can I achieve this level of style? If anyone knows pls suggest me Model and LoRA’s .

[sweet nexus]

guys

what u prefer to use?

civtai or apob ai?

been using savro and i'd like to test what is better

[proper quest]

most of the influencer models operate on flux models underneath anyway, so its irrelevant

[primal pulsar]

How specific can I get with colors in Stable Diffusion? I've only been very basic with it so far, like "Green Eyes" and "Brown Hair". Is something like "Army Green" recognized? Is there perhaps a way to use RGB color codes?

[magic grail]

Anyone know a good way of removing text from an image such as this one

[proper quest]

yeah thats not gonna happen

too complex

[proper quest]

no rgb/hex codes

[balmy vale]

what's a good prompt for a canvas/textured paper effect? i can get watercolor easily on illustrious, but not the former

[tired vigil]

anyone willing to give me some feedback on a lora i made? im trying to replicate yusuke muratas art style such as opm. image isnt really nsfw but this is an example of what it can make

[light obsidian]

I wanna know what the prompt is to remove human ears

[quaint agate]

How would I prompt to get a literal blade for a hand, like a prosthetic or whatever that would be covering the stump from the wrist/forearm down

I’m trying to prompt for making blade fist orcs from WoW but it’s uh not going great

[quaint agate]

i guess theres not much help for that kinda thing.

[sour beacon]

Well there are not much of these images in the training sets of most models. So either some of the anime models or specific character Lora’s would maybe generate this.
Otherwise I would recommend using a canny controlnet. So generate the orc, generate a nice blade merge it in a photo editor of your choice and use it as input image for the canny, depth or lineart controlnet.

[quaint agate]

ah.. all stuff beyond my skills. could i , start with a image with a blade on its arm already

and try to work it into becoming an orc with a blade for hand?

im still a bit new to img2img ..SD forge just does resizing and cropping. how do i take that and make it something new if all it does is that. or does it allow me to prompt it from said image to a new one?

[light obsidian]

I need help, my cow keeps getting human ears even when I put in (human ears:1.2)

[stone path]

how do I change the light source in a picture? I'm having trouble with illumination hitting a character directly from the front and illuminating the face of said character

[proper quest]

hardly actually

there isnt a tag for character being well lit from the front

[proper quest]

try puting extra ears in negative

[sour beacon]

You might take a look at the sdxl relighting models. For example

https://huggingface.co/spaces/lllyasviel/iclight-v2-vary

[proper quest]

still doesnt have front lit

[reef torrent]

Hey, I'm new to SD, how do I make a prompt for 2 people appearing on the screen (both fully visible) each of them with different outfits/pose?
Trying to generate a scene when 1 guy is giving the other a handgun but no matter how and what im inputting it either mixes their features/poses or one doesn't appear on the image at all (maybe his hand). Any example of that?

I wanted one of them to have a grin on the face and dark hair while the other was supposed to be blonde and have unsatified/mad look but no matter what I do I can't get them to have different features (that is when I manage to get both of them on the screen, usually I don't).

[hollow tapir]

Search for "regional prompting".

[naive salmon]

hey, so I'm trying to use dynamic prompting to make specific colors constant through dynamic prompting (like making two body parts the same color)

${skxn=!{green|purple|orange}} these are the wildcards and I'm using it like (${__lxrd__} scales),${__skxn__} skin, but then I'm ending up with yellow-skinned characters```

what am I doing wrong?

[proper quest]

Im sorry what?

[proper quest]

What do. Uwant to do

In plain english

[naive salmon]

I want to make a variable so I can make a prompt that reads (Var1) skin, (Var1) arms using dynamic prompting so it's randomized btween generations but always equals the same in the same generation. If I have {blue|green|red} at those spots instead of Var1 it rolls the dice for both variables instead of just once.

[proper quest]

'prompt1|prompt2' is gonna select either 1 or 2 randomly

[naive salmon]

not if I set a wildcard variable according to what it says in the image I sent

[proper quest]

Ffs discord on phone cant do __

If u want it to be chosen twice as often, just add it again

Thats how i play with odds lmao

[unborn prism]

bird, (Engine:2.2), (hybrid:1.5), (bevel-edge:1.3), (translucent:1.3), (refractive:1.2), (metallic:1.2), soft-shadow, (antistrophic-filtering:1.2), (jade:1.3), (crystal:1.4), (diamond:1.4), (jewel:1.4), (lens: medium-format), (intrinsic:1.3), hyper-realism, post-processing, (texture:1.2), (depth:1.1), (terrarium:1.2), forest, (blend: multiply), isometric, (steampunk:1.4), (bio-organic:1.2), (gypscope:1.6), diffused-lighting, (fractal:1.7), (inlay:1.5), iridescent, silk, anatomically-correct, engineering, mechanical Negative prompt: little-girl, missing-limb, tessellation, collage, extra-limb, human, Steps: 15, Sampler: DDIM, CFG scale: 4.5, Seed: 713084654, Size: 640x1024, Model hash: f47ad22837, Model: stokedRealityXL_v10, Version: v1.6.0```

Am i doing something wrong? Why cant i get my old sd quality back

[silver valley]

The widht and height are not the same here

Xl models are trained on 1024x1024 so stay near that resolution

In the top image you used 640x1024 but in the bottom one 616x376 which is way to low

[unborn prism]

Oh wow does resolution actually play a part?

[silver valley]

A big one yes. SD1.5 models got trained on a 512x512 resolution. When using these with a to high resolution they will output duplicates.
As SDXL got trained on double the size it has problems using really small resolutions making the output really bad

[unborn prism]

Amazing insight, thanks man

[silver valley]

No problem 🙂

[buoyant quartz]

Hi

[silver valley]

Hey

[buoyant quartz]

I want to prompt image on this channel

#🆕｜sd3

[reef torrent]

is regional prompting the best/only way to achieve that?

[hollow tapir]

Not the only way, you can try to constrain the output using controlnet. Use another model with a somewhat better "lexical analyzer" such as sd3 / flux / etc. You can generate one character and then the second with outpaint / img2img functionality (assuming they re not overlapping). Start from a rough sketch and use img2img to repaint it. Etc

[hollow tapir]

#artisan-faq

[silver valley]

Forge couple is good too

Inpainting can work too but is much more work

[reef torrent]

https://github.com/Haoming02/sd-forge-couple is that it?

[silver valley]

Yes

[reef torrent]

i've been trying for some time and i still dont understand how to make something more complex than just standing next to each other
can't make a proper background, can't make them interact with each other, can't make a longer promp describing what they are doing, it just mixes everything together even in advanced mode with tiles

[glossy dock]

is there a trick to make the same face without lora or checkpoint until you make like 20 images so you can make a checkpoint?

[sour beacon]

Not a trick but flux kontext is a model that is great for character consisty.

[glossy dock]

i dont seem to have flux kontext

is it like a checkpoint?

[sour beacon]

Yes and no. It is a concept for changing images with text prompts based on the flux image model.

[glossy dock]

how can i implement that in my stable diffusion?

[sour beacon]

Which user interface are you using?

[glossy dock]

AUTOMATIC1111's WebUI. i believe

[sour beacon]

I am not sure if kontext is already implemented with auto.

[glossy dock]

what do u use?

[sour beacon]

ComfyUI but it is node based approach. Very flexible and very fast in implementing new models thru the modular node architecture. But most people who come from a more artistic direction need a while to adapt to the way of using it.

[hollow tapir]

Sure it s not

[wanton swan]

I wana generate this kind of image some one help me

what model is this ? settings ?

[kindred blaze]

Looking for some help to try to achieve this high gloss shiny look on both the body of a figure and her face, either a prompt or if anyone knows of a Lora that can accomplish it please ❤️

[wet saffron]

Anyone need help with Loras training?

[still turret]

What's the catch?

[fiery pecan]

i find it annoying to inpaint in comfy, I don't think it's all there like the ease of Automatic

[sour beacon]

If you like inpainting you might take a look at krita and the ai plugin. It connects to comfyUI and works like a charm for outpainting and inpainting tasks.

[wet saffron]

Anyone need help with flux or sdxl lora trainings?

[reef torrent]

@unborn gyro

[silent moon]

hi

[pulsar moat]

what are the different prompting words for artsyles? Things like watercolor, splash color, etc. I'm trying to look for them on danbooru but they don't seem to specify

[hidden mauve]

Anyone needs someone to build workflows for them?

[languid veldt]

hello there i need some help no matter what i try i dont get a darker/dim lightning like candlelights, does anyone knows a way?

[sour beacon]

it total depends on the model you use. There are some models which won't create dim images. You could always use some thing like flux kontext which the prompt change the lights to dim, cozy atmosphere, ....

[hollow coral]

is it possible to make stable diffusion "forget" its memory?

several projects ago, i prompted for tanlines. i have had multiple computer resets since then, but it still insists on giving my gens tanlines

i am using the forge through stability matrix

[hollow tapir]

don t follow that link, it s a scam
EDIT : cleaned, bye bye scammer.

[proper quest]

First time hesring such thing

[empty chasm]

I need to create images of jewlry with prompts like "modern signet ring, oval face, engraved gothic initial ‘M’, high-polish sterling silver, subtle reflection". I'm new to SD, any help is appreciated on prompting strategy.

[acoustic trail]

How do I generate an image where 2 characters are facing each other? e.g I tried to generate a 1940s bar image where the bartender is serving a customer but it kep fusing them together?

[solemn flame]

Can I also ask for Lora training help here?

[proper quest]

probably region prompting

[proper quest]

ye

[elder spoke]

fusing them together
have you tried a wider resolution?

[acoustic trail]

Yeah, I even typed in the prompt separate characters, character one and character two you putting things in like white shot, extra white shot but I could never get them to face each other separately

[proper quest]

@pearl pilot

[quiet zodiac]

@acoustic trail Try supplying a reference image for your scene to a controlnet. Then the prompt doen't have to work as hard.

[sacred fulcrum]

If a lora has a trigger word, do I just put the trigger word at the beginning of the prompt? I see that some people use the trigger word, and some poeple put lora:triggerword:1
The lora is already in the beginning of my comfyui workflow, so I am not sure what I am supposed to do

[quiet zodiac]

Just put the trigger word at the end, so it is not as important as the rest of the prompt. The <> syntax, I think comes from online sites, not necessarily Comfy compatible.

[languid veldt]

hello there im looking for a good/easy to unterstand tutorial for regional prompting in ComfyUi, im planing to create some images with multiple different characters and as far as i know cant ComfyUi or a Image generation "AI" handle this topic well. THX in advance

[long tartan]

Hello everyone!
I’d really appreciate some advice from experienced AI artists.

Prompt structure — how did you come up with your “ideal recipe”?

I’m trying to figure out the best sequence of tags and descriptions (score, source, style, etc.) that actually works well with SDXL models.

I drafted a structure for my own character + custom LoRA, but I’m not sure if it’s correct or if I’m overloading it with useless stuff. Here’s an example of how I currently organize my prompt:

---

General quality block:

score_9, score_8_up, score_7_up, perfect proportions, uncensored, digital painting, best quality, 4k, very aesthetic, masterpiece, vibrant, professional, source_western BREAK

Environment (examples):

solid background, cityscape, landscape, indoors, outdoors, scenery

Character:

1boy, hichigo_shirosaki, solo, long hair, hair between eyes, hair volume, white hair, white skin, black sclera, detailed yellow eyes, pale lips, adult male, toned body, masculine, looking at viewer

Clothing: (to fill in)
Pose/Action/Expression: (to fill in)
Lighting/Atmosphere/Vibe: (to fill in)

Key tags / LoRA:

shirohichigo, hichigo_shirosaki
lora:shirohichigoIL2:0.7

---

What I’ve noticed: on Civitai, most prompts under images look pretty chaotic — no strict structure, everything mixed together. But I’m curious:

Do you personally follow a certain structure (blocks like quality → face → hair → clothing → pose, etc.)? Or do you just write freely?

Are there any good articles or guides that helped you shape your own prompt structure?

Do you use “style” tags to improve the generation, or do you prefer dedicated style LoRAs instead?

Thanks in advance for sharing your experience! 🙏

[smoky sigil]

/create A magical birthday celebration, balloons flying, cake glowing, kids smiling, cinematic animation

[arctic heron]

@silver valley

[silver valley]

flux is not compatible with auto1111

[arctic heron]

oh

[silver valley]

so it falls back to an other model

[arctic heron]

would ihave to use comfyui for that?

[silver valley]

you would need Forge or comfyui

Forge looks the same as auto1111 but more updated

[arctic heron]

rips

[gilded turtle]

hm. need some help

my genned pics keep having really long necks, or like warped clothing and faces in bodies

[silver valley]

Thats because of the resolution + the high denois.
A few tips:
Try a resolution like 832x1216
Don't use _ underscores in the prompt. They are not needed.
Reduced the steps to 25-30 as more doesnt make the image better.
For hires fix always set the hires steps to 10-15 or 20. don't leave it on 0.
And then set the denois slider to 0.4-0.5

[vast pendant]

Hi

[gilded turtle]

that.. didnt really help at all

[silver valley]

Can you show me how it looks like with these settings?

[gilded turtle]

its like, getting betetr but its still distorted a lot

[silver valley]

did you used the same seed? try with -1 (random)

[gilded turtle]

seed is random

[silver valley]

oh okay

[gilded turtle]

for some reason it keeps making the neck super long?

even though i have long neck as negative promt

[silver valley]

can you try euler a as sampler?

[gilded turtle]

its already euler a

https://tenor.com/view/kid-awkward-worried-nervous-embarrassed-gif-4782036

[silver valley]

oh

[undone crow]

im a beginner in SD. currently using counterfeitv30(mostly)/control v11p openpose sd15 and waiNSFWillustriousSDXL_v140/openposeXL2 as check point/controlnet.

can anyone give me some advice to get a better result.
i am trying to create a character using this (skeleton) poses. i keep getting results like this.

either there are faces or there are some furniture adding. what am i doing wrong?

PS. when im using SD1.5(counterfeitv30): i can generate multiple poses from multiple skeleton in a single image it has the same problem.

[sour beacon]

What is the resolution you are trying to generate. SD1.5 is max 768x768 as the base resolution it was trained on is 512x512. If you use 1024x1024 he adds normally more objects in the image.

[undone crow]

im only using the minimum 512 512 for both

[sour beacon]

And your prompt only describe the character?

[sour beacon]

No Problem with the pose in my opinion

[sour beacon]

Character sheet is kind of messy but still "ok-ish" if you try to squeeze 14 positions of a character into 512x512 pixels....

[undone crow]

okay ill try that. can you give me sample prompt you use for it? maybe my prompt is the problem to make it cleaner

[sour beacon]

Sure. Prompt was: character sheet, no background,

[undone crow]

i see maybe using white background generates something white that has mixed setw

"white background, character sheet, 1 boy, man, masculine, BREAKmessy black hair with silver streak bangs, BREAKschool uniform, blue blazer, loosened tie, high quality, high detail"

[sour beacon]

Even i would think you are talking about different characters 1boy vs masculine man. ... The break stuff is also not very precise

[sour beacon]

the "school uniform" tag leads to chairs and tables...

[sour beacon]

And some of the poses don't work out realy good...

[upbeat forge]

Hi everyone.
I'm stuck in my progression. There is something I'm missing. I hope you'll be able to unstuck me.
The picture from @sour beacon just above this message is a perfect example. I can produce this kind of character sheet from a prompt. I use a 'squeleton' sheet (from DWprocess I believe) and I obtain a random character.
Now when I want to use a personnal image on this squeleton sheet I face a problem: I don't know how to deal with two images as input. I didn't find the right controlnet node/workflow that let me do that.
I vaguely see the flow but can't wrap my mind around it.
Any lead?

[sour beacon]

Controlnet-Pose and Image2Image is pretty much a no go.
Either the input image determines the result or the pose. Both simple don't work. You can set the strength of the controlnet to low and let the image denoised by a large potion but you won't see either the pose nor the source image in this mix up.

To different approaches would be try qwen-edit or flux kontext and ask for another pose of the same character (not realy usable for NSFW Content....) Or create the new Person and change the face...

[upbeat forge]

If I understand you correctly, there is no obvious possibility to use a skeleton sheet with another image.
I then must try the checkpoint Qwen or Flux kontext for different approaches.
Does it mean that, in a context of gathering images for Lora training, my best shot is to create different skeleton sheet with different poses, edit different character with these sheet (txt2img), and swap the faces on these character to the one I want ?

[sour beacon]

Well there are other options nowadays. For example some used 3d meshes (and blender) to create characters and use the bone rig system to pose them. But the texture on the mesh can be off. WAN got some nice Loras to create movements like a rotation, so you can start with one image and then rotate the character 360° to get more sides.
But again to get different clothes, background, poses etc. a image to image and controlnet is not the way.
Try flux kontext or qwen edit and see if it helps. Work step by step otherwise it might miss your attention.

[upbeat forge]

I will play with Kontext and come back. Qwen is way above my hardware league...
thx

[thorny pelican]

Hey. Can someone help me with getting a a Anime style character together. Im currently using Forge Webui. If there is any suggestions im open to em thank you in advance 😄

[silver valley]

hey, can you describe a bit more what you want to do? xD

[tender crest]

Hey,
I'd like given an image like the first clean one, to add mud splatters onto the camera lens of the image and get something like the dirty image provided

Additionally if it can be done with inpainting but so far I feel like SD is very bad at giving me a dirty image.

The idea behind this is to take a dataset of clean images to get a dataset of dirty images and their segmentations to train a downstream model that can identify where a camera is blocked due to dirt mud etc.

Any advice?

[thorny pelican]

Basically im saying im interested in making my very own Anime Character. But i want it to be consistent

[silver valley]

For that you can try controlnet with IP-Adapter for noob-ai(Illustrious)

[unkempt swan]

how can i add more details? like when i do a big picture, i want it like HD or more.
I got a Lora (more Details Lora) but it isnt thaat helpful..

Or how can i add more details while upsclaing?

[young topaz]

okay, so i want to learn how to post process with img. Like upscaling it and impainting it with img2img. I Don't need to impaint anything at the moment. but does anyone have any good img2img upscale prompts and screenshots i could borroiw

[silver valley]

For Inpainting always set the Inpaint area to "only mask" and then set the resolution sliders to a square resolution based on the used model. (1024x1024 for sdxl, 512x512 for 1.5).
Then set the Denois slider to 0.5 to start and see a difference. The Denois defines how much the image will change at the maked area and how good it will belnd over into the image. Then play around with the Denois slider to your liking.
The prompt should be very defined like only describe what will be seen. Nothing like "nice details or etc", more like "sunglasses, green eyes, black hair etc," when inpainting face or changing the hair color.

[young topaz]

nods and writes this down

thank you so much, i am not great with this

[silver valley]

For img2img there are no specific prompts, just make sure you remove all words that refer to characters to avoid morphing or diformations. like remove 1girl, eyes, lips, hands, hair, etc.
For upscaling in img2img make sure you also set the resolution to a sqaure (no matter the size of the input image).
At the bottom under scripts select SD upscale script.
Then again play around with the Denois Slider. Usually start with 0.2-0.3 for the best results. Going higher will mostly break the Image.
There are techniques for example using controlnet model tile to prevent that. But its a bit more complex.

[young topaz]

Okay, thank you so much!

[pulsar rock]

Is anyone able to help me undertand why I get a black background on this image when I type White background?

[sour beacon]

Maybe you might elobarate which tool / model you are using. In general the Pipe-Symbol does have another meaning

[pulsar rock]

Is this where I submit the URL I'm at and get in trouble

or is that allowed

It's at "sdxlturbo" site
I replaced | with a comma and that helped
vs
"nvm fixed it"

[glossy grail]

Hi, i am a architecture student learning how to use stable diffusion. You have any material of prompt engineering to learn how to write appropiately prompts?

[sour beacon]

Hi Lucas,
most architecture use cases that are discussed here are creating more realistic images from model (image 2 image / controlnets) or image to sketch and sketch to image scenarios. Some even use the ai to generate different scenes for inspiration.
But there is not one prompt to fit all that and different models require different prompting.

[glossy grail]

very thank you, there is a specific group or you know web community for the architectural use?

[undone crow]

Needing some help.
im trying to use image to image to generate different views of specific part of this image(2nd image attached). like wide front view of the 2nd image attached. is it impossible to do it?

[unborn ivy]

Manga style, black-and-white ink, dramatic contrast, cinematic angles. Sequential panels, consistent characters, tense horror-thriller mood. Silent library, frustrated writer, masked killers, surreal ending. Each page shows panels with continuous story flow.

---

Page 1 – Library
2 panels: vast empty modern library, tall shelves, rows of tables; closer view of books and dust in silence.

Page 2 – Writer
6 panels: close-up of man (30s) writing furiously; pen in hand; wide shot alone at table with books; messy scribbled handwriting; crumpled paper; shadowed angry face.

Page 3 – Intruders
4 panels: library doors open, masked men enter; close-up of cold eyes; killers moving between tables; man tapping desk, killer behind.

Page 4 – First Kill
3 panels: disruptive man tapping; killer grabs his hair; throat slashed, blood on table.

Page 5 – Girl
4 panels: young woman gasps; killer covers her mouth; “shhh” gesture at Keep Silence poster; silenced pistol shot, she collapses.

Page 6 – Writer’s Rage
3 panels: writer slams fist; killer behind with knife; suspenseful knife over him.

Page 7 – Break
4 panels: writer rips page; killers vanish, library empty; writer breathing heavy; fist smashing wall.

Page 8 – End
3 panels: glass door shatters; writer crushed under shards; close-up of shard with “Do Not Disturb, Keep Silence.”

---

i have this prompt how can i generate images by this prompt, pls help me , i am new to this

[hollow tapir]

cf #artisan-faq, or build locally, or use some online services

[unborn ivy]

like do i need to enter any special command

?

[silver valley]

Use a local webui to generate images for free.
Or pay for the bot in #artisan-faq

[unborn ivy]

u know any free webui , pls let me know

[silver valley]

If you have a good GPU, the local ones a free.
Like Automatic1111 or Forge, ComfyUI, SwarmUI, InvokeAi

[unborn ivy]

anytging free available on discord??

[silver valley]

Only Dreamdiffusion

[unborn ivy]

link for it?

[silver valley]

Nope

[unborn ivy]

then how can i find it

[silver valley]

Using a search engine probably

[marsh echo]

How do I get rid of the typical stable diffusion model face?
All females look the same, even when using makeup prompts, it's always the same structure. I even tried with LORAs but had now luck.
Any model recommendations?

[hollow tapir]

Not sure what you mean by that. Could use different models indeed or adetailer with better / more specific description for faces.

[quiet zodiac]

@marsh echo Start prompting or gathering a set of faces to use for face transfer. You can use the IPAdapter or other technologies to swap in a reference face. What I've discovered, is that you can blatantly crossfade two different face images, convert that to latents and use it as a unique face.

[unkempt swan]

someone here known in Anime prompting?

[unkempt swan]

need some help to get better 😅

[silver valley]

hey, show an example maybe ^^

[unkempt swan]

maybe this one, is there a solution for eyes? because the more far a person gets, the more the details are getting lost.
or can i somehow work on them seperate after image creation? i mean inpaint is a thing, but i dont get it.. too much for my brain xD

i already use an eye-Lora and a dramatic lighting lora.
about the lighting, the face never gets light on it, always shadows, but the lora increases the details a lot!..

[silver valley]

The overall solution for face and eyes is the extension called Adetailer.
It automatically fixes the eyes and face depending on the Adetailer model used. It can also be used to fix other stuff.

[unkempt swan]

how to install and use? ^^

[silver valley]

Click on extensions tab, click on Available, Load From.
Then in the list you search for Adetailer and click install

Also install the booru tag autocomplete extension for better anime prompting

Then restart the webui

For Adetailer just enable the checkbox when you want to use it.
Additional tweaks can be made if needed

[unkempt swan]

massive thanks, but there's so much to select xD
wow, dont know where to start.
i might test around a bit

[silver valley]

for the start just enabling it does the most part

then using the denois slider can effect how much it will change the face/eyes etc

[gusty pagoda]

any tips on how to make it generate graphic design style generations? let’s say i want a physics chemistry wallpaper it needs to be artsy etc, currently looks generic.

[unkempt swan]

omg, just did some tests.. this is amazing xD
do you got any ideas how i can improve the grade of detail?

[elder spoke]

sampler / steps can have a big effect too
try e.g. deis 40 steps

[tribal kayak]

Hello guys, someone can help me with a specific pov I can't get it? I got it 1 time between 1000 and I didn't even know how, if someone can help me dm me please

[robust gate]

Why does my stable diffusion always divide the picture into 3 parts when I didn't ask for that

[sour beacon]

I would guess because you use a model with a base resolution of 512x512 pixel and create an image with 512x1500 resolution….

[robust gate]

Well to add onto it it’s only on some generations as well as the resolution is 1024

[hollow tapir]

that would be the same problem tho.

If you ask for a native resolution output too high compared to the training resolution, it will result in duplicate artifacts

you want to ask for something within +-25/30% of the training resolution and then upscale the result (be it with hires.fix, img2img, etc)

[ocean timber]

So, I'm trying to make a character image I'm trying to generate for a DnD campaign, trying to get the image of him to be laying with his back on the bed. As if he's relaxing or in thought. However, constantly I'm seeing the generation put him in a position where it's face down toward the bed. y prompt says this. relaxing in bed on back, lost in thought as he looks at ceiling, bedroom with a window looking outside, side profile shot

Any ideas of what I'm doing wrong? Either the image tries to be sitting in the bed or face down towards the bed?

[hollow tapir]

It depends of the model you're using, settings, etc

[prime rivet]

#bot-commands

[fair hearth]

How can I include two "checkpoint" models in the generation?

[unkempt swan]

so you basically want a dakimakura(or atleast the pose)? did someone helped you?

[ocean timber]

if you mean that by how it sits flat? Sure. And no.

This is the basic result I'm trying to get.

pose wise

[unkempt swan]

try something like : sideview, from the side.. etc

[ocean timber]

Yeah, I get the side view easy enough. However, they never are laying flat on the bed

usually get them sitting on the bed, or they are looking face down into the mattress.

which, is both parts funny and frustrating

[tender vale]

need help making consistent characters

[ocean timber]

use more descriptive prompts, this is the first and easiest way to make character consistency go up a lot.

if you use something like, 1 girl, big breasts, koala ears on top of head, stone grey hair, bowl cut, slightly tan skin, faintly glowing amber eyes, dark yellow chiffron dress, as starting point and add in more details, you'll get a lot more consistent looking characters

[tender vale]

That's great but I have a pixel art style that i want to make consistent but DALL-E is the worst tool I have ever used

[thin aurora]

You can use image2image approahc

[tender vale]

Thank you. Any info will help

[misty monolith]

Really hoping somebody can help me. I just need to conclude that im not too good at prompting. Have en Project going where im creating logos for fictional nfl teams, mostly in flux.. I had success with some but there are a few where i just cant replicate what i made in sora originally. How would you prompt this?

[proper quest]

maybe maybe as hrfix refiner

[hollow tapir]

tldr you don't unless you do two passes.

you don t drive faster with two cars.

[languid veldt]

hello there, i need some help im currently making some images for character creation and i have every time the same problem, as you can see in the IMG there is hair that lies over the shoulder of the character and thats the problem i dont want this i tried "sidelocks" as negative prompt, "short sidelocks", "hair behind ears" some natural language and much more but nothing works. im using the WaiNSFWIllustrious model and thats my prompt (i let the NSFW prompt aside):

1girl, solo, cute, (makeup:0.5), long hair, black hair, hime_cut, hime bangs, blunt_ends, purple eyes, (tsurime:0.75), (shiny_skin:0.5), medium breasts, looking at viewer, happy, light smile, looking straight

resolution: 1024x1024
Steps: 30
CFG: 8
Eula_Ancestral
Clip skip: 2

if some one knows how to fix this please let me know as fast as possible.

Thanks in advance

[remote sage]

Hey fellas, i'm prompting some PVC-like anime figures and having good results overall, but the problem is that i'm having a hard time replicating an specific hair style (2nd pic) Tried controlnet with different models, modes and so on and no matter what i try i just can't get something even close. Checked my prompts many times and nothing seems to work. Any help is appreciated! Model https://bit.ly/48iocY7

[languid veldt]

i have a similar problem and for me using the IPAdapter works... sometimes... maybe give it a try (just look on YT for IPAdapter FaceID and you probably need some manual tweeking and trial and error)... but as for my Experimente it dont work very good with Loras and im not sure if tis works with your PVC-figurine project since i make only 2D images.
the two other methods i would use is a character LORA or prompting (from the IMG you show i would use "purple_Hair, short_hair or Medium_hair, Parted_Bangs, yellow hairband,")

[burnt yacht]

hello, i am trying to recreate a... for lack of a better word, sepia filter? i have tried many "anime style" or "golden hour" loras promps related etc but nothing seems to give a similar effect, i know it probably depends strongly on what checkpoint i am using but some ideas would be appreciated

[drowsy minnow]

Where do you go to talk to not to create images

Is it that hard

[acoustic wharf]

Hello - forgive me if this is the wrong channel - but is there any appropriate place to look for hiring someone to help me build an AI pipeline for my game assets?

[fickle bramble]

Hello, I can help you with that, I'm a python developer with more than 5 years of experience and I'm also very skilled in AI and LLMs.

Reach out so we can work on it.

[languid veldt]

hello there, i need some help, im using ComfyUI and want to use one of those Slider LORAs but they dont work for me, i´m either to dumb or i missing something if someone knows how to run those properly please tell me how (im using the normal LORA loader of ComfyUI)

[hazy lagoon]

Hello, I don't suppose anyone could help me with what feels like a weird, abstract problem when it comes to trying to generate an image of a known object? I'm trying to generate imagery of a distinct type of naval mine but it seems SD 1.5 and SDXL as models as much as I try to prod the prompts around just... can't? Could this be a model training set problem?

First image is said type of general shape of landmine and the other two are just of my attempts and the results that generally come out.

[hollow tapir]

you re generating 512x512 images with SDXL (e6bb9ea85b) which is made for 1024x1024. That s why.

[hazy lagoon]

I was getting the same problem in SD1.5, in that it tends to produce anything but that distinct spheroid with the protruding triggers. Using SDXL to generate things like this (soviet pioneer, balloon, etc) seems to hold up fine at 768x512 resolution for me.

I'll try running at 1024x1024 in SDXL then and report back.

[hollow tapir]

you can usually deviate +-30% of the training resolution without getting any visual artifacts

[hazy lagoon]

I see

[hollow tapir]

More than that would be gambling.

[hazy lagoon]

So ~1.330+ and ~720-

[hollow tapir]

about that yes.

If you want higher res you ll have to use upscaling / hires.fix / img2img

[hazy lagoon]

I was trying to nail settings at lower resolutions to quickly re-iterate through my prompting to get the general "look and feel" down for the object.

[hollow tapir]

good idea, but yeah don t lower it too much otherwise you ll get garbage outputs.

[hazy lagoon]

I'm quite unfamiliar with SD after the interface* changes (played with it briefly around 2022-2023) and I'm effectively trying to create... Well, this (rough draft) image out of it with with the naval mine in effect so I've been trying to iterate prompts quickly enough to get the visual style right.

...Not good enough to touch img2img yet to "nail it" and/or I don't think I have the prompt right to make it switch the balloon with the naval mine.

[hollow tapir]

You could also use this as a reference with controlnet or use it with img2img

[hazy lagoon]

Yeah, I was going to try to learn img2img/inpaint from video tutorials where I've seen people paint out the replacement section but I needed to nail the prompt right for the mine.

It seems the distinct Hertz horn (the little spike/studs) fuse isn't "spikes" for the prompt.

I assume I'm not wrong in treating the prompt more like a tag system instead of using natural language, right?

Ex:

(naval mine:1.5), sphere, masterpiece, (photorealistic:0.5), (photo:0.1), (spikes:0.05), studded, atomic bomb, [chain]

Instead of something like "chained (naval mine:1.5) with external protruding fuse studs" or whatnot?

[hollow tapir]

For starter, you could just give it that image with an empty prompt and tweak the denoiser strength to make it redraw the whole thing without changing it "too much"

[hazy lagoon]

Ah, scatter it with the denoiser from a single seed I like?

[hollow tapir]

that way it will redraw some details, unify the seams from any quick and dirty photomontage.

it s very basic but it will get you started with img2img

[hazy lagoon]

Gotcha

Do I want to give it any negative prompts or just leave those blank as well?

[hollow tapir]

leave it both blanks.

One thing to keep in mind, is that in img2img the gui is kinda lying to you regarding the number of steps it does.

The "real ammount of steps" it does is denoiser strength*steps+1

[hazy lagoon]

I see, so dropping sampling steps down as well should be helpful

[hollow tapir]

if you just want to do some basic upscale with minimal modifications to the image. Something in between 0.1 and 0.35 should be fine. (haven't done it in a while, my numbers might be a big wrong. I m at work so I can t check right now)

[hazy lagoon]

Yeah I'm just trying to get that distinct naval mine look without it being a stock image (...the massochistic option for me is to photoshop the text out since this will be used as a low resolution thing)

These very, very early naval mines have those fuses sticking out and are visually distinct compared to more modern ones

[hollow tapir]

contact naval mine might fair better.

or floating contact mines

[hazy lagoon]

Getting close, thank you for the guidance earlier

Negative: (Alamy:5) did a lot of work...

And a careful bit of balance on denoise strength (~0.45 makes the text turn into light reflection)

...I'd take this even in nuking the text out

Thank you again, this got me in the right direction.

[merry linden]

any tips on proper bow nocking on illustrious? sometimes its close but havent had any perfect ones

[harsh kayak]

How can I generate fine tuned ui for mobile apps? I have extensively searched loras too but couldnt find anything with a wow factor. How troublesome would it be to train a lora on 3-4 mobile apps ui?

[sonic grove]

Hi everyone,
I want to generate images with a style similar to a reference image I have — i find it difficult to describe the style.
For those of you experienced with prompt engineering:

What kind of prompt keywords or structure would you recommend to capture a specific visual style?

Are there any LoRA models that work well for achieving such looks?

Any tips on keyword placement or style modifiers that can make the result closer to the reference image?
Thanks in advance for any suggestions!

[sour beacon]

This Prompt:
"A highly stylized digital painting of a futuristic luxury car parked in front of a surreal, impressionistic city skyline at sunset. The car is sleek and aerodynamic, with smooth metallic surfaces that reflect its surroundings in a chrome-like finish. Its oversized wheels have a modern, turbine-inspired design, and subtle blue lighting glows from beneath the chassis. The scene is bathed in warm golden-hour light, with intense contrast between warm oranges and cool turquoise hues. The background features a painterly, semi-abstract cityscape resembling New York, with iconic skyscrapers like the Empire State Building rendered in bold vertical brushstrokes. A large, monolithic, sculptural tower rises prominently in the midground, splitting the composition and adding surreal tension. The sky is heavily stylized with a gradient from soft peach to deep cyan, dripping downward like oil paint on canvas, enhancing the dreamlike atmosphere. The ground reflects the city and car slightly, adding depth. The image blends photorealistic rendering of the car with abstract expressionist elements in the background, creating a futuristic, cinematic, and otherworldly mood."

Leads to the composition but the special "drawing" and lineart of the image is not captured.... 🙁

This prompt:
"A highly stylized digital artwork blending clean, futuristic lineart with painterly, impressionist backgrounds. The central subject is a concept luxury car drawn with precise, aerodynamic contours and smooth, flowing lines that emphasize speed and modernity. The lineart on the car is crisp and technically accurate, with subtle, controlled curvature, tight perspective, and detailed rendering on features like the headlights, wheels, and paneling. The line weight varies dynamically to create depth, from bold outlines on the body to finer, more delicate strokes on intricate surfaces.
The background contrasts this precision with loose, vertical brushstrokes reminiscent of oil or acrylic painting techniques, creating a semi-abstract city skyline. Architecture is implied more than defined, with jagged, impressionistic forms and minimal line detailing, letting texture and color suggest volume. The skyline bleeds upward with elongated vertical lines, creating an almost liquid sky — a technique that evokes dripping paint or soft rain. Colors are applied in layered gradients and washes, rather than rigid fills, producing a soft, flowing effect, especially in the sky where blues and oranges blend like watercolor bleeding.
Reflections on the car’s body are rendered with painterly digital smudging and airbrushing techniques, blending hard-edged realism with stylized lighting. The composition relies on contrast between structured, technical draftsmanship in the foreground and expressive, abstract mark-making in the background. The image overall demonstrates a fusion of sharp digital illustration and loose, tactile digital painting — merging mechanical precision with artistic fluidity"

Leads to a more conceptual image but still got problems with the line art in the background 😭

[torn thicket]

Regarding the Slider LoRA issue - they often require specific custom nodes and aren't compatible with the standard LoRA loader.

Quick troubleshooting steps:

1. Make sure you have ComfyUI-Manager installed, as most Slider LoRAs depend on it.
2. Try loading node groups like "Impact Pack" or "WAS Node Suite" - they contain specialized nodes for advanced LoRA handling.
3. Check the download page for that specific Slider LoRA - authors usually provide example workflows (.json files) that you can import directly.

If you're still stuck, let me know which specific Slider LoRA you're trying to use and I can help pinpoint the exact nodes needed.

I've packaged several optimized workflows that handle these compatibility issues out of the box - might save you some setup time.

[torn thicket]

Training a LoRA for 3-4 mobile app UIs is definitely feasible. Here's a practical approach:

Key steps for your case:

• Data: Gather 20-30 clean screenshots per UI style. Consistency is more important than quantity.
• Tool: Use Kohya_ss GUI - it's most reliable for style transfer like this.
• The Secret Sauce: Your captions need to describe the common design patterns across all your references (e.g., "minimalist navigation", "rounded input fields", "consistent spacing").

The real challenge most people hit isn't training, but efficiently testing and iterating on the trained LoRA. That's where having a structured generation workflow becomes critical.

I've documented a complete process for this specific use case - happy to share the workflow details if you get stuck after the training phase.

[torn thicket]

What's the most effective way to handle multiple LoRAs in one ComfyUI workflow without crashing? Specifically dealing with memory issues on 8GB cards.

Tried different loaders and execution orders but still hit limits. Any proven node setups or custom scripts?

[chrome moth]

hi guys good morning, afternoon, or evening any of yall know how to fix this?

[silver valley]

Hey, you need remove the lora stuff and only try to replace the lora value.
In the prompt make it to lora:add_detail:X
Then in Prompt S/R set it to X, -2, -1, 0, 1, 2

[chrome moth]

oh ok ok i only add lora cuz i was following a vid alr thx

[sonic grove]

thank you!!! you are a genius!

[azure flame]

https://generativeai.pub/the-math-art-of-artist-0thernes-not-the-typical-96e009060bc1

So I got published. Take a read. Tell me what you think?

[chrome moth]

hey quick question. i tried what you said the other day and the output was the same. so i decided to use and XL detail tweaker and it kinda work
im using illustrious base model
and my question is: is there any illustrious detail tweaker? or XL is compatible with illustrious

[silver valley]

The xl is compatible with illustrious and pony. Because boath are based on sdxl.
But it doesnt work the other way round

[chrome moth]

Ohhh ok ok. Cuz the output I generate is not really detailed only a little. But is ok I'll just keep trying it till I get it now

Thx

[hollow tapir]

btw judging the quality of those images, it looks like you re generating 512x512 outputs with models designed for 1024x1024.

[chrome moth]

Ohh. So my model or Lora designed for 1024x1024?

Thx for the tip I'll try generate but in 1024x1024

[formal thicket]

Good morning. I would like to know if there is a specific prompt when you want to represent several characters on an image

[hollow tapir]

depends of the type of model you're using

[formal thicket]

I'm on illustrious. but if you have other models to recommend to me I am interested

[hollow tapir]

ok so something SDXL based.

Then I'd recommend using some kind of regional prompt extension. https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/wiki/Regional-Prompt-Control (just in case you're at work, the examples used in there are very ... well endowed)

the tldr idea behind it is "splitting your image into multiple regions. Each region holding one character being dictated by one prompt"

[formal thicket]

Okay. And what if the characters have the same outfit? Do I have to re-specify it every time?

[hollow tapir]

Probably yes. You could try to specify it in the "common prompt" that will be shared across all your other prompts but I'm not sure about how good the results will be.

[formal thicket]

Okay I will try this. Thank you

[modern lodge]

If you want multiple people, prompt is enough, if you want more people with unique faces you need to inpaint them, use IPadapter, controlnet and LoRA for clothing to get consistency

Train character outfit LoRA I guess

[formal thicket]

What is IPadapter qnd controlnet ?

@hollow tapir I tried to make a prompt based on the link you sent. the character descriptions kept getting mixed up

[hollow tapir]

Can you take a screenshot of your UI ? Showing us what you tried.

[modern lodge]

Things that guide the image generation

[formal thicket]

This ?

[formal thicket]

I have to write that in my prompt, where are these LORAs to add?

[modern lodge]

I don't use Automatic111 Web GUI but ComfyUI.

[formal thicket]

What is it ? Are there any differences?

[modern lodge]

Just check it. It serves the same purpose but it's more flexible. You have nodes that you connect together to create a workflow for your needs, workflows are also available on the internet or create your own.

[formal thicket]

I looked at what it looked like. This reminds me of how textures work in Blender. I'm installing it and will test to see

[proper quest]

this isnt going to work

[modern lodge]

it will work, but it'll be painfully slow with an unpredictable result

[formal thicket]

And how can I repair this ?

[modern lodge]

choose xl, karras is probably too much (slower with dpm sampelr) though it's accurate or better for upscaling I found out

I prefer euler_ancestral and 20 steps is enough for that model (yes, I'm trying that model uut right now)

also idk how that highres works, but first create an image and then upscale

I checked Highres but don't understand it at all from the description on this website https://www.runcomfy.com/comfyui-nodes/efficiency-nodes-comfyui/HighRes-Fix-Script

[formal thicket]

Okay. I have also this problem, can you help please ?

[modern lodge]

go into ComfyUI Manager and download missing nodes

that's what ComfyUI Manager is for

also rgthree nodes are a must

at least for me

[formal thicket]

@modern lodge What is the problem here please ?

[modern lodge]

not sure what you are trying to do here, but maybe msimatch between models? Show me the workflow

[formal thicket]

I hope french is not a problem

[modern lodge]

uggh, why are you trying to use Flux nodes with SDXL model?

I haven't tried FLux but would like to try that one too, the models got just too big though

[formal thicket]

Because I'm a big noob ? XD

And I don't know where I can see this informations

[modern lodge]

check the tutorial series: https://comfyui-wiki.com/en/tutorial

[formal thicket]

https://youtu.be/RXuTNuyM6GI?si=Etjs9IhNOfH1h28O I use this tuto

[formal thicket]

I deleted the Flux nodes, and the error message changed

[modern lodge]

It's not about deleting the flux nodes. Create a basic workflow and try again or use templates and load a basic workflow for text to image generation.

[modern lodge]

It tells you you don't have the checkpoint in a folder where a guy, who made it, saved the checlpoint.

So whenever you download a workflow, you have to switch the paths if you don't have them the same as the workflow author.

[chrome moth]

hi guys. can one of yall help me find a lora that can turn my images into a lineart. like an lineart lora i want to use it on img2img and txt2img. everytime i tried to generate a prompt with a lineart lora it always come out not what i wanted heres some 2 prompt i made

lora:ayase_seiko:2 seiko_wz, white hair, 4k illustration, anime coloring, masterpiece, ultra detailed, anime, perfect image, good anatomy, detailed illustration, beautiful, anime style, 1girl, (girl in crop top:1.3), yawning, (stretching:1.1), overhead stretching, in a Japanese house, cinematic lighting, volumetric lighting, diffused morning light through window, cinematic shadows, highly detailed face, ultra detailed face lora:LineAniRedmondV2-Lineart-LineAniAF:2 lineart, LineAniAF

lora:ayase_seiko:2 seiko_wz, white hair, 4k illustration, anime coloring, masterpiece, ultra detailed, anime, perfect image, good anatomy, detailed illustration, beautiful, anime style, 1girl, (girl in crop top:1.3), yawning, (stretching:1.1), overhead stretching, in a Japanese house, cinematic lighting, volumetric lighting, diffused morning light through window, cinematic shadows, highly detailed face, ultra detailed face, lora:lineart_style_illustriousXL:2.0 lineart, solo

i use 2 different lineart lora and they both are not good

can someone help me or know any lora or prompt method that turn my images into lineart

[hollow tapir]

looks like you re generating 512x512 with an SDXL based model.

should be 1024x1024 (then upscale if needed)

[hollow tapir]

wasn't at home sorry (still am not) but that s not how to use regional prompting

[proper quest]

Ur lora weights shouldnt be over 1 , unless its a slider/detailer lora.

Also, lineart can probably just be prompted instead of using a lora

[chrome moth]

alr i'll try lowering the weight thx for telling

[silver valley]

Normal loras range is from -2 up to 2.
Values like 0.8 or 1 work good.
Only slider loras can go higher/lower without destroying the image

[zinc bridge]

is there one beast way for structuring prompts like?
embeddings, style lora BREAK character lora and description BREAK scenario lora and description, angle, setting, other details

or is it kind of random and each structure can produce good results? I use illustrious based models and tried a lot of different orders of sections with or without BREAK and can't really tell whats best

[modern lodge]

Hi. I don't understand what's that BREAK but it's simple... Make the structure clear e.g.: appearance, position, action, composition (angles and so on)

That's just an example

You could use join "string multi node" or something like that

Just search for "concate", "string" and other keywords and you'll get your result

[compact citrus]

Hi guys
I'm getting miniature, bad quality doodle people in the background of my images.
Like, it's like a second charater with the same characteristics as the main character of the image, but it's chibi-ish and/or doodle-ish. What can I do to get rid of it?

[silver valley]

Hey, try add solo, to the prompt and if hires fix is used, lower the denois value

[compact citrus]

Thanks! Will try that.

[jade terrace]

Hey guys i recently downloaded a llm modle
I want to use it and for that i need the searge node but after downloading it i get the error llm-cpp
Dose anyone have an idea on how to fix it?
(I use setup version with an Nvidia gpu)
HELP ty...

[main crane]

This prompt created a video with several errors and distortions. Does anyone know how to fix it?

Anime 2D 6s (24fps), Naruto 2002 style.

Setting: floating platform with neutral background and ethereal grid lines.
Camera synchronized with the character's rotations and the surrounding space.
Smooth interpolate. 1920x1080.

[0.0s] Character in the center, fighting stance, right fist retracted, neon blue energy pulsing in the arm.
Camera: medium frontal shot, static.

[1.5s] Character rotates torso 45° to the right, right shoulder advances, legs flex in a dash.
Camera rotates laterally (low-angle ascending), following the torso rotation.
Opponent (gray silhouette) to the left of the frame, torso turned 30° backward.

[3.0s] Character completes 90° shoulder rotation, fist explodes in a frontal punch.
Camera rotates 360° around the axis of the punch, keeping the fist centered. Impact to the torso: white flash 1 frame, blue sparks, strong camera shake.

[4.5s] Opponent thrown backward; camera pulls back in wide shot, following the movement.
Trail of dust and neon blue energy; background grid distorts slightly with the rotation of space.

[desert needle]

I'm kind of stumped on what i'm doing wrong here. I'm attempting to join 2 videos together in a seamless flow by adding the last frame from the first video to the first frame of the second video. Right now it's only a single frame. I could join the 2 together by other means but the second video, 1/2 of it is garbage and it doesn't continue where the first one left off. If someone already has a working .json file that does something similar I don't mind starting over.

[modern lodge]

check this workflow for reference, it has what you need: civitai.com/models/2039923/not-so-simple-or-is-it-wan-2122-lightspeed-i2v-workflow?modelVersionId=2312835

[unborn saffron]

good

[trail forge]

Hello

[limber jetty]

I've launched a prompt sharing program 🔥

https://www.getsnippets.ai/earn

[agile remnant]

video_wan2_2_5B_ti2v ive been stumped for days, noi matter what settings, scheduler, sampler, wan2.2_VAE.pth or wan2.2_vae.safetensors, model shift 1.0-8.0, denoise 0.2-1.0, steps 10-40, cfg 2.0-8.0,

i seem to get a blocky mess of a subject and color distortions

[balmy vale]

Any recommendation for a prompt to get full black eyes for a cartoon image? Something like scott pilgrim? Mine are always like gray with some gradient, never full and pitch black

[modern lodge]

but Wan 2.2 has a high and low noise model

you are missing the 2nd one, don't you

[tight ferry]

I have an annoying problem i am trying to generate a character with black hair, but the lighting makes it look grey almost white, any idead how i might me able to prevent this? it is in every image

[formal thicket]

Hi. Sorry to bother you again with this, but I'm still stuck on my problem with Comfy UI. I understood that the model and the nodes were not compatible, but I don't know how to change the model so that it works. If anyone can help me a little more precisely, I'm interested (at the risk of sounding like an idiot)

(as a reminder)

[sour beacon]

Well whenever you download workflows from any source they have values in the different drop down of each node. In your case the checkpoint loader tries to load a model that you don’t have within your folder (models/checkpoints). the node show have a red box around it and with the dropdown you could chose another model as long it is the same architecture (sd1.5/sdxl/flux/qwen/…)

[formal thicket]

And for that ?

[sour beacon]

Would say the file you chose is corrupted. Download not finished or broken

[formal thicket]

that didn't solve anything

Just the template files, is it in "checkpoint" that you have to put them right?

[daring compass]

I was directed here. Can someone help???

[hollow tapir]

it would probably help if we saw some examples of what you're trying to achieve.

What do you mean it's not following your prompt ?

Screenshot the whole webui with some results so we can see

(also I'm sorry to say but automatic1111 is quite outdated by now. You might want to install forge or forge neo instead later on. Check the pinned comment in #🤝｜tech-support for some guide about that)

[daring compass]

thank you! How do you upload pics (I have them).

[hollow tapir]

drag and drop or copy paste here

[daring compass]

my wood sign and so I did generate with promts Image2image a great pic of a woman holding this court officer wood sign. But now it wont work . Unrelated results.

[hollow tapir]

Yes but what are those unrelated results ? What do they look like ? Show them to us.
Also show us the settings you've used to generate them so we can see what settings to tweak.

[daring compass]

I will upload my SD Interface next

+Want the code that geberated to great picture of the girl with the sign?

generated

[hollow tapir]

Yes because the screenshot you just posted doesn't mean much. The prompt is empty, it's using sd1.5 model which is 3+ years old by now. Not gonna help us solve your "output unrelated to prompt" issue if there's indeed no prompt.

[daring compass]

Here is the prompt that worked...:
PROMPT:
Remove the background from the sign pictured below (uploaded) Create an Ultra-realistic photo of a woman in her 30s holding this carved Wood sign chest level, smiling proudly. Create a background of a court room with Judge atop a podium and witness chairs. Cinematic side lighting, soft rim light, 85mm lens, f/1.4, hyper-detailed skin textures, photorealistic reflections on acrylic, shot in a modern studio, 8K resolution, extremely crisp detail. Image of sign:
"C:\Users\Owner\Dropbox\AI\AiPicsIn\IMG_0003.JPG"

Negative PROMPT:
blurry, distorted hands, extra fingers, low resolution, plastic skin, cartoon, grain, artifacts

+++Also do I need to upload a newer sd model? Can I update??? if so how do I do that and what is the best sd model?

[hollow tapir]

Image of sign: "C:\Users\Owner\Dropbox\AI\AiPicsIn\IMG_0003.JPG" doesn't mean anything to the model if you used this as is in your prompt. It's not gonna pull the file from your disk for you (Personally I would be worried if it could access files without my consents like so)

[daring compass]

Understood.. I did upload the picture of the sign.

And this was GREAt but I cannot duplicate it again or with other signs

[hollow tapir]

Hard to say which is the best among the hundreds of community trained models, but you can give Juggernaut XL a go https://civitai.com/models/133005/juggernaut-xl it's fast, relatively easy to use and the outputs should be a massive step up from that sd1.5 model I saw on your screenshot. Download the model and put it in stable-diffusion-webui\models\Stable-diffusion then rerun the webui-user.bat.

[daring compass]

can I use A1111 to add this? Do I need to use dos cmd.

Is it Free??? Everyone wants my $$$$$$$$.

[hollow tapir]

You can keep A111 for this as it's still an SDXL based model. But I would really not recommend using A1111 especially with a RTX 5xxx gpu. I don't think A1111 has even been fixed to correctly support those out of the box.

yes it's free.

[daring compass]

We fount the upgrade and installed it. Again my system is a Ryzen7 + Geforce 5070ti + 64 gig ram and a 2 ssd hard drive.

[hollow tapir]

I'm gonna head to bed for now. In the meantime, I would really encourage you to switch from A1111 to Forge Neo. Using this guy should make everything easy https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#nvidia-forge-neo. You can keep your old A1111 as a backup if you want.
Forge Neo's UI is very similar to A1111 so you should not have to relearn everything.

[hexed charm]

What prompts or models do i need if i want realistic image of anime characters

that makes it look like they are cosplaying

[formal thicket]

👆

[mortal gate]

Am I allowed ask a question about an nsfw image here?

[proper quest]

you need a checkpoint for it to start with

[proper quest]

ask in dm if u havent gotten an answer yet

[formal thicket]

Could someone please answer me ? 😭

[proper quest]

dont use comfy simple

[formal thicket]

But where is it marked that these are Comfy samples?

And is there a way to change the samples? (other than manually)

[proper quest]

What samples ru talking about

[formal thicket]

You told me not to use confy samples

[proper quest]

i told u not to use comfyui

in general

[formal thicket]

Oh ok sorry. So I come back to Stable Diffusion basic?

[proper quest]

just get forge vanilla/classic/neo if u want to gen images

[formal thicket]

So I come back to my first question: how to represent several characters without the characteristics being mixed?

[proper quest]

forge couple

[formal thicket]

I must download forge ?

[proper quest]

https://civitai.com/models/339604/how-to-generate-multiple-different-characters-mix-characters-andor-minimize-color-contamination-or-regional-prompt-adetailer-and-inpaint-or-my-workflow

[formal thicket]

Does it work with OCs?

[proper quest]

why wouldnt it

[formal thicket]

And so that my characters all have the same outfit, are there just LORAs?

[proper quest]

read the guide first

[formal thicket]

This one ?

[formal thicket]

I haven't changed anything in Stable Diffusion, and yet it gives me this error message. Is Comfy Ui doing this?

[proper quest]

what gpu do u have

nvm thought u replying to me

[formal thicket]

AMD Radeon RX 6800

[proper quest]

should prolly remove the --cuda argument you got there

[formal thicket]

And how can I fix this ?

[proper quest]

try by removing the arg

[formal thicket]

We'll have to be more precise (I'm bad at this)

[proper quest]

in your webui-user.bat

[formal thicket]

How can I remove this ?

[proper quest]

by editing the webui-user.bat

[proper quest]

remove the venv folder and run webui-ser bat again

[proper quest]

try adding this arg to the list --device-id=0

if this dont work then im clueless

perks of being an amd user i guess

[formal thicket]

Didn't work T-T

Well thank you anyway. I'll see if others can help me. And if not... I guess I'll have to redownload everything

[proper quest]

@silver valley

[silver valley]

Follow the Forge with Zluda setup guide from here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

Let me know if anything isnt clear, but I'm off now.

[chrome moth]

hi guys can any of yall help me with my prompt? i was trying to make an lineart prompt using a img2img and controlnet but the outcome is messy some section of the images is filled like a giant black circle or other and everytime it has shading. does any of yall knows any negative prompt to stop it.
if so much appreciate

[proper quest]

why would u use controlnet to achieve lineart?

[chrome moth]

because i found a video that teaches how to transform img to lineart using img2img and i was interested and this is my first time doing it and the video said use controlnet canny.
and once i did the outcome was good and for the 2nd time im having a trouble with the prompts cuz the img has filled in color in some sections of the img and has shadows and shading and i tried so many ways changing prompts to fix it
and im asking that does anyone know any prompt or ways to stop the filled in colors and shading/shadows?

[formal thicket]

I have this error message at the end. Can you help me please ?

[silver valley]

wrong channel, can help in #🤝｜tech-support

[crisp ridge]

hey guys im trying a cyberpunk lora, and it works well, but it keeps putting a cyborg body on the character's chest, i want the cyberware only on his face, how can i do something like that?

it keeps doing this

i want something more like this

https://civitai.com/models/1517871?modelVersionId=1717297

im using this lora, the illustriousXL version

[digital herald]

So I'm fiddling around with ((Front side back three views)) for a character concept image. Can I then specify (Front view: Description), (Side view: Description), and (Back view: Description) or is it difficult for Stable Diffusion to interpret that?

[digital herald]

Even just using ((Front side back three views)) is giving me mixed results

[digital herald]

Ah, i think it's because of img2img and denoising strength.

How might I develop a preferred look/appearance for a character that I can reproduce with different camera angles, clothing, and back drops? Any online guides are appreciated!

[crisp ridge]

how do i indicate wich hand should be mechanical on my character, i put the prompt (mechanical hand:1.7), and it makes the character with 2 mech hands sometimes just one, but i want it to be just the right hand

i want her to have this setup

[proper quest]

first use the correct tag, mechanical arm

and you cant specify left or right

[river fern]

I pleasantly noticed yesterday that Adetailing faces to high values (eg 4Kpx) actually adds tremendous detail than if you scaled to 1K or 2K. Way beyond anythive i've seen on general skin even upscaling to high values. (with USDU, Tiled Diff, HRFix)

1. Is there anyway to just make adetailer run at a high res rather than making the whole image run in a high res.
2. Is getting Adetailer level quality for the entirety of an image possible? thanks for reading

[mortal gate]

Is it possible to create an audio file from a text prompt using a certain model or UI?

[velvet charm]

Hi guys quick question do you know if these images were made on stable diffusion? If so does anyone know what the prompting would be, it’s 1980s style

[signal juniper]

i wanna get this exact artstyle

[proper quest]

do two passes of adetailer

[native meadow]

how do i make dymanic promt to use 4 of words from file ? {4$$file} ? {4$$file} ? erm , (nvm figured out)

[low roost]

https://www.twitch.tv/Nekomipii
playing my favorite game join us on the live game

[formal thicket]

HI. In terms of the prompt, how should I formulate if I want to make a piece of clothing in two different colors, or two eyes in different colors?

[leaden socket]

Hey lads, I'm trying to make a creature image to video on wan 2.2 but NOTHING I prompt seems to allow me to do an animation with the mouth closed, the creature always seems to be moving the mouth like it''s talking, suggestions?

[leaden socket]

I would use a image to image workflow.

[river fern]

hi I just noticed your comment, what do you mean by that? Do you simply mean use adetailer at multiple points of the workflow?

[proper quest]

use two face detailers or whatever one after the other

[pulsar moat]

So I want to do some genning of characters without the usage of loras through danboru tags. However, I noticed that there's difficulty genning certain characters even if they have 2k posts on danboru. While others with only a couple hundred can accurately be genned. How does models work with the danboru tags?

Do like more recent models encapsulate more recent of the danboru archives?

[sour beacon]

Different model architectures and models where trained (and /or) finetuned with different base image sets. Some anime based got a great understanding of characters even without loras and others are great on realistic objects (chair, table, influencer) but do not know any game, anime character at all.
Therefore people created the loras to add specific character, style or objects to their favorite model.
Some models are trained with danboru tags (image description of base images where done with tags) other on natural language. So you need to search for a model generic enough to creative the relevant images (Background, Styles etc.) and specific enough to know your characters....

[pulsar moat]

awesome ty

[unborn smelt]

I see some ppl use \( \) in prompt, what does this means

[twin yoke]

Do you mean something (like this) or sometimes (like this:1.2)?

Or even (((like this))).

[hollow tapir]

The backslash is used to "escape" a special character

Parenthesises are special characters used to emphasize a certain word in a prompt

Therefore if you want to actually use proper parenthesises in your prompt you need to "escape" them

[unborn smelt]

\(like this\)

[twin yoke]

what they said ☝️

[sharp osprey]

no matter the model be it flux, qwen, sdxl, ill, pony. and no matter how i prompt i can't seem to get the focus to not be on the character described or make the character further from the viewer. anyone got any suggestions or loras?

[sour beacon]

NeonLighting as shown in the other discord the images created by nano banana pro or qwen image are able to capture a person far away if prompted accordingly.
I guess that you have precise descriptions part of the character so it becomes a dominant part of the image. Some people want the ai to create a character with a scar under the left eye, wearing a black ring, nike air sneakers, .... but at the same time this detailed describted character should be 30x60 pixels of the 1024x1024 image. This does not work well. So focusing on the surrounding or whatever needs to be in focus and add a "figure" to it.
Later you can still change the figure with the character if needed (flux context, qwen image edit, flux2 or what ever you like or licence wise can use)

[quiet zodiac]

@leaden socket Try instructing the character to do something or specify a turntable animation. You also need to spell words correctly inside your prompt. There is no such word as "curiose". When AI runs into words that don't exist, it often hallucinates a place holder.

[leaden socket]

Encoding and transformers handle misspellings, it uses sub tokens not the full words.

[lyric musk]

Anyone knows how to use forge couple? I need help please.

[tulip relic]

illustrated vector art for apply on a access card of a important way to enter in dubai emirate this art shoube about dubai's heritage and future of dubai and this art should create with only 4 colours.

[proper quest]

there are guides on civitai

[lyric musk]

can u send me the link please?

[proper quest]

https://arcenciel.io/articles/2

https://civitai.com/models/339604/how-to-generate-multiple-different-characters-mix-characters-andor-minimize-color-contamination-or-regional-prompt-adetailer-and-inpaint-or-my-workflow

[lyric musk]

Thanks

[thorn hearth]

is it better to do (illustrious checkpoint)
prompt
BREAK
style loras
BREAK
character
BREAK
place/pose
or
prompt
BREAK
Character
BREAK
Style loras
BREAK
pose/place
or even keep the style loras at the bottom

[proper quest]

i keep style loras at the bottom

everything else, in order of importance

[tired vigil]

I was wondering does it have any impact if I name any famous/known artist names within the prompt like "style by alberto seveso, Range Murata, Alphonse Mucha" - for example. Is this doing anything? I am not sure to be honest.

[sour beacon]

Depends on the model you chose to work with. SD1.5 for example is really good with the artists name. Later in the process models were often trained without the artist in the image descriptions. So at least the less known artist have disappeared. Van Gogh is like Rembrandt still good to use.

[tired vigil]

What about Pony and Illustrious?

[sour beacon]

Pony uses Danbooru Tags. Not quite sure if there are ones for artist. There are lists with tags online. I guess if there are ones it would be anime artist focused.

[tired vigil]

What is the best prompt/Tag against 3 fingers? (Thumb and 3 finger). Generally is it better to tell the AI Five finger or Four Finger. I mean I would count the thumb as a finger, but how is the AI working in this case?

[proper quest]

i often use four or five fingers for inpainting

idk what model r u running on, but on a lot of my prompts where the hand is visible i use five fingers, cuz its easier to reduce from 6 rather to add one to make it a 5 from 4

[tired vigil]

What is the best ADetailer? Or Workflow.

[rapid basin]

Trying to fix hands with just prompting is a nightmare lol. 😅 I switched to a Python workflow using Llama 4 Scout, it’s way better at reasoning and keeping the character consistent across poses.
​I made a dedicated generator tool for this. If anyone needs the source code, feel free to check my profile.

[cedar brook]

I've been practicing prompting for anime style but kept messing up the tags. So I hacked together a simple site to help me stack tags visually: https://aniprompt.xyz/ I hope this will be helpful.

[vital mulch]

Hello can someone help me with img2img? Everytime i try to generate a picture it only shows a black picture and im not sure what model is good for a anime style one

[silver valley]

Use a negative prompt like blurry,

For anime go with Nova Anime XL

[vital mulch]

Not sure why its such a different image lol

[vital mulch]

Oh okay im using that one now

[vital mulch]

Coz tbh i just wanted this to look like an anime cant figure it out on how tho haha

[thorn hearth]

can anybody help me develop a good stable prompt?
the prompt style ive been using is self taught, and I want to refine it so i cna have more control over it

[thorn hearth]

please some SD guru help me...

[balmy vale]

is there a list of colors that you can use in prompting in general? like brown works, but dark brown doesn't seem to work, is there another way to write?

weight doesn't seem to affect it either

[proper quest]

what ui

[balmy vale]

isn't the checkpoint more important? i'm using illustrious on auto1111 and forge neo

[proper quest]

how many steps in ur image

[balmy vale]

20-25

[proper quest]

right, starting from now u will generate with 30 steps.

Any you will type [brown hair:black hair:25] (adjust if necessary)

[balmy vale]

what does [ do? i always used (, and with small number between 1 and 2 for weight

[proper quest]

it will generate brown hair, and at 25th step it will start generating black

[balmy vale]

ooooh i see

interesting, thanks

[tiny grotto]

​🚀 LOOKING FOR A CREATIVE PARTNER (AI ANIME SERIES) 🚀
​The Vision: I am a 15-year-old creator building a high-quality AI Anime Web Series. The goal is to build a solid portfolio and eventually pitch to major platforms like Netflix/Crunchyroll.
​Who I am looking for:
​🎨 AI Artist: Someone who can generate consistent Anime characters (Stable Diffusion/Niji).
​🎬 Video Editor: Someone who can animate images into cinematic scenes.
​🌍 Age: Preferably 15-18 years old (Beginners are welcome!).
​The Deal: * This is a Long-term Partnership (50/50 Profit Share).
​No upfront payment, we grow and earn together as a team.
​Let’s build a masterpiece from scratch!
​Interested? Send me a DM

[shrewd prawn]

Even when using I2I (Image-to-Image), you still need to properly fill out the image prompts and then use ControlNet to lock in your imagery.
This is a simple example. You can see all the parameters and models I used; feel free to modify them according to your own needs

[agile quail]

what app/program is that?

[vital mulch]

I think its forge neo.

[balmy vale]

so i'm using dynamic prompter to get various results

but is there a way to randomize the CFG scale?

like i do 15 images, and randomly some will have a CFG scale of 2, other a scale of 5, same way i use dynamic prompter for hair color and such

[proper quest]

and why would u do this

[balmy vale]

variety of results

[proper quest]

cfg is barely gonna help u with that

[balmy vale]

and yet with my settings it does

i guess there's no way of doing it then?

[proper quest]

i mean maybe on comfy, not so sure about forge

[balmy vale]

thanks anyway

[proper quest]

Wait, if u r on forge, i think you can try xyz plot to perhapse change cfg

Idk about the randomness part

[wide edge]

Hiii Im using regional prompt and it doesn't work I need a little help haha-

[silver valley]

Feel free to share an example and whats not working

[wide edge]

Thanks, I'll try tomorrow with a friend. If the error keeps happening I'll document it here

[frigid plinth]

I want to digitalize a printed photo but AI change and removes people. Write me a prompt.

Can anyone enhance this for me

[naive bolt]

That picture tells some history. Very cool

[frigid plinth]

I need the prompt, if you can enhance it; do it.

[lyric musk]

So, this is happeing alot with me latellyu, how to fix this?

[hollow tapir]

looks like you're not using adetailer properly

it's repeating your whole prompt on the faces

[proper quest]

which is fine lol..The problem is his denoise is too high

[proper quest]

^

[lyric musk]

Here are the config that im useing

[proper quest]

thats not related to adetailer

[hollow tapir]

Not really fine, they should use a different prompt for adetailer. Even something generic as "face" should help.
lowering CFG to the point where the reused global prompt does not show in the output is kinda counter productive.

[proper quest]

2nd, reduce the denoising strenght to 0.35 for hrfix, and change the upscaler model to a standalone model (ultrasharp, remacri)

[proper quest]

i made over 25k images with the prompt field being blank

[lyric musk]

i just use this

[proper quest]

so it is fine

[proper quest]

and its settings

[hollow tapir]

two options
1/ like I said you put "face" as ADetailer's prompt and it's gonna redraw the face using previous output as a base
2/ you lower adetailer's CFG to the point where it does not care whatsoever about the prompt because it won't have enough weight to do any significant work, turning it into some basic upscaler.

[proper quest]

3rd option, dont listen to the guy above me

[hollow tapir]

your choice...
Both options are valid but one is miss using the tool. It will have some effect on the output for sure. It might even be better. But I'd say it's best to know how to use the tool the way they were designed for before hacking them.

[proper quest]

my fault for not breathing on discord

[lyric musk]

cfg sacale or highres CFG Scale?

[hollow tapir]

again, method 1

method 2, lower adetailer's cfg

[wide edge]

Update! It worked. It seems like one of my loras was responsible. Thank you for offering help anyways!

[keen beacon]

Genuine artist desperately still looking for help creating music videos, happy to pay $100 for successful help, I had to delete everything & start again as I ended up with 100gb+ of pointless files due to misdirection & failed guides. If anyone would just be kind enough to help guide me as a complete noob, I have partial brain damage & suffer from too many chronic migraines to figure it all out by myself... 🙁 Just please no more scammers!

[midnight needle]

Hi guys, I would like to ask. I want to create several high school girls or uniformed wearing girls characters. How do I make the uniform consistent?

[native pier]

I'm losing my mind. I'm trying to generate a photo of an EMPTY swimming pool next to a motel. Everything I've tried fills the goddamn pool with water. HELP PLEASE! I feel like I'm going insane.

[inland token]

lol try putting water in your negatives and use dry concrete in the prompt

[native pier]

I got it! - prompt that worked was "an empty pool, it's empty, full of dry cracked concrete, the setting is next to a rundown motel in the american midwest desert" with "water" in the neg

[inland token]

nice one glad it finally worked out

[native pier]

Now I have to add a diving board and metal ladder

[inland token]

inpainting will save you a lot of headache with the placement honestly

[native pier]

Does Z-Image have an inpainting model yet?

[inland token]

no official one yet, but z-image-edit or the fun-controlnet union should work

[native pier]

Without inpainting, not too bad! I couldn't get a graphic of a Wave on the sign, cause it kept filling the pool with water 😄 so a sun will have to do.

[inland token]

nice one, semantic bleeding is a pain, regional prompter usually fixes that

[native pier]

Final prompt was "an empty pool, it's empty, full of dry cracked concrete and bits of dry tumbleweeds. There is a metal ladder on the ledge, and an old broken diving board. The setting is evening, next to a rundown two-storey motel in the american midwest desert under a starry sky and a crescent moon. There is a large motel sign saying "TURQUOISE WAVE MOTEL" with a neon graphic of a smiling cartoon sun."

[inland token]

sick prompt, maybe try adding logo or decal to the sign description

[native pier]

besides the sun?

[inland token]

yeah, 'wave logo' or 'decal' makes the ai treat it as a graphic, not water

[native pier]

Great idea... I said neon logo of a wave

[true siren]

https://deut.li — FREE local metadata extractor (no cloud, no install). Works with SD/Flux.

[mighty wedge]

Hi, I’m trying to replicate a specific art style from NovelAI images and I’m looking for advice on the best approach (prompting, LoRA, or workflow).

My goal is to reproduce the style and shading so I can generate many different characters with the same overall look. I don’t need to recreate the exact same characters — consistency of the art style is the main priority.

I’m currently using Stable Diffusion (SDXL) with:

• StabilityMatrix (Windows)
• waiIllustriousSDXL_v160
• animagineXL40_v4Opt

What I’ve tried so far:

• Training a LoRA using PixAI (auto training from uploaded images), but the results weren’t very close
• IP-Adapter (ip-adapter-plus_sd15)
• Prompt adjustments
• Trying to reproduce directly in NovelAI (free credits), but I couldn’t get close

I’m wondering what the most effective approach would be for this type of style.
Would training a proper LoRA locally be the recommended solution?
If so, how many images are usually needed for good consistency?

I’m still relatively new to LoRA training, so any advice would be really appreciated.

(Images attached below — NovelAI samples)

[inland token]

illustrious is your best bet, try training a local LoRA with about 25 images

[mighty wedge]

Thanks! That’s really helpful.

Do you think OneTrainer would be a good option for training, or would you recommend Kohya instead?

[inland token]

kohya is the standard for a reason, way more tutorials if you get stuck

[mighty wedge]

Got it, that makes sense. Thanks!

I’ll try setting up Kohya and preparing a dataset.

[vital junco]

Anyone up for a quick chat I have a question?

[proper quest]

Sup

[proper quest]

First install derrian lora trainer

Then let me know when u have it working

[rotund sandal]

Hello guys please help me out with this images because I seriously nobody is helping I think nobody knows but I do not know why this is hard for anyone to generate guys I do not have high and pc so I can't run high and a model even do not know about the stable diffusion properly I can't use it I have a very medium not medium I have low Vram PC but if anybody knows how to create these kinds of ultra photo realistic military images absolutely these are mind blowing images please help me out with the prompts and also please and also please tell me which AI online tool can generate these types of images please

[inland token]

those look like tactical bodycam shots, try flux on tensor.art for that level of detail

[rotund sandal]

Thank you for your respond but I do not have high and PC I have very low Vram6gb PC I used to stable if diffusion of for one day but I am unable to get this kind of quality I seriously even do not know about the prompting so please help me out please help me out

[inland token]

use 'tactical bodycam, gopro style' on tensor.art, 6gb is a bit low for local flux

[rotund sandal]

I know that bro I have very low PC but is there any online AI image generation which can generate these kinds of images if yes please tell me which one or also please provide me prompt how to write these types of image from please brother

[vital junco]

. I’m doing a very specific style clothing brand. Some of the additions on my pictures that I need are, backgrounds, lighting, gloves, helmets, but all in a specific style. I’ve gathered a bunch of images as well. i just need some help getting started training a lora. or maybe not even. this is my first time using a genertor like this

[wary narwhal]

How did you create this style?

[proper quest]

@hollow tapir @silver valley

[wintry rapids]

Right... how do i start this... so im using wailillustrious on Forge Neo, and I wantedto try out Forge Couple, originally wanted to have this two character hug each other but they dont really hug that well, so i tried bleeding their mapping section into each other, now they hug properly, but then, because of the overlap the girl on the right started to gain the features of the girl on the left,

So i tried to inpaint the glasses away atleast, but it doesn't remove anything, well more like it doesnt do anything even after adjusting the noise strength, and if i switch to latent noise setting, it just error the faces that i selected

So then i tried inpaint sketch, still doesnt work, i have to what to do with this, or does a simple restarting the UI fixes this???

I suppose this is a question about in painting more than prompt?

Edit: Got it, just need to disable forge couple

[uncut granite]

Hello everyone !!!!! ❤️
I’m not sure if I’m in the right place, so apologies in advance if this topic has already been discussed.
I’m currently working on a project: I’ve already created my AI’s face using several photos, and now I’d like to generate a very specific body type (very slim waist with wide hips and prominent glutes, BBL-type proportions).
To be clear, I’m aiming for a realistic human look, not an anime, cartoon, or stylized render.
The issue I’m facing is that on Stable Diffusion or Higgsfield, I quickly run into limitations due to NSFW filters. I’ve been stuck on this for several days now, so any help would really mean a lot.
I was wondering if anyone here has tips, workflows, or best practices to handle these constraints while still achieving a realistic, human result.
Thanks in advance for your help, and sorry again if this question has already been covered.

[fossil pasture]

re Stable Audio: am trying to understand how I can input a particular sample, and then make its output sound more like sound fx, than music - if you know what I mean. Sort of musique concrete-ish. Are there prompts that can make my input sound like "tumbling pieces of cardboard" or just "metal", as in the object/material "metal"?

[proper quest]

@hollow tapir

[woven birch]

How do u find the name 4 an artist? I want 2 make things in the style of dr stone but idk what artist or style 2 prompt 4?

[true siren]

Hey everyone, I need a quick favor for some stress-testing.

We are finalizing a 100% local, zero-server metadata extractor for DEUTLI, and I need to push our browser-based parser to its limits.

Could someone drop a few massive PNG files (15MB minimum) that still contain their original Stable Diffusion or ComfyUI metadata?

I absolutely do not care what is actually on the image — it can be pure noise, a failed upscale, or a literal black square. I only care about the large file size combined with the embedded tEXt chunks (workflows/prompts) to see how the client-side parser handles the load.

Feel free to drop them here or in DM. Thanks in advance!

[mild night]

🙂

[lilac zephyr]

https://www.linkedin.com/posts/akshita-goel-52744b282_assistivetechnology-accessibility-userresearch-activity-7437148557192658944-5v2e?utm_source=social_share_send&utm_medium=android_app&rcm=ACoAAETD4S8BR5vsA7E9vSxEQe-MZ17_YkIfgAU&utm_campaign=copy_link

[lyric musk]

There is anyway to put speech balloon and put what i want?

[proper quest]

If sdxl it will be gibberish

Otherwise u can use blank speech bubs and write what u want ur self.

If u are on any bigger models (flux, qwen, zi) they should understand text

[lyric musk]

IM sorry i dont know the meaning of gibberish, is a extension or what?

[proper quest]

Mean incomprehensible text

[lyric musk]

Oh i see

[wintry rapids]

No need anymore, figure it out

[hollow tapir]

good cause that was a scammer

[velvet charm]

Hi guys do you know if these images have been created on stable diffusion?

[elfin star]

been a while...how would you prompt something like this to AI?
(ideally with 1.5 based models \ finetunes, low end PC, that is faster for iteration)

Idea is similar to what is search-able by "sword beam" (not gonna post long link), but 1.5 seems to suffer with it.
In my usecase I need it specifically as projectile, not related to sword in any way, without sword on the image

[upbeat current]

Are there any tools for prompt optimization? For example if i give it a list of prompts it reorders them?

[grave abyss]

do you want prompt?

[upbeat current]

what?

[grave abyss]

i mean do you want prompt?

[restive oar]

I need a cinematic photo of a clean, modern office bathroom, without a bathtub, just a spacious area where people can move around easily, with a modern sink and a mirror that reflects the stainless steel details. The tiles are so clean they reflect the light, with some plaster details on the ceiling and a stainless steel trash can.

[somber abyss]

Anyone with the tool wanna help me out?

10 sec video

[little sleet]

Hey @somber abyss, I can help you with that!

[somber abyss]

Do I have to pay you?

[hollow tapir]

if they only talk in dm, be careful.

[somber abyss]

Yeah, thats the issue

If I had a pc would have done it myself.

[pine junco]

Struggling with prompt weights and syntax? 🤯 I found it way faster to just sketch the composition roughly instead of describing it.
I built a quick tool that turns simple drawings into detailed prompts/images without needing to type a novel. Might save you some time if you're stuck on the wording: draw.freefroai.com

[modern oak]

helloo

I need some assistance in correcting the prompt I am using for a vton project?

Whould any with the expertise in this field assist me?

It would be a great help

[little sleet]

Hello @modern oak, I can definitely help you with that.

[modern oak]

Can you please dm me?

[little sleet]

Yes

[proper quest]

Hes prolly a scammer

[proper quest]

@thin carbon

[thin carbon]

whats up, discord was down

[proper quest]

There was a mrbeast scammer

Someone may have nuked him

[thin carbon]

ah i couldnt see it, thanks for pinging but i think the other guides cleaned him up

[lunar rune]

cloud flare burp actually

[thin carbon]

As usual from cloudfare

[proper quest]

@thin carbon that allowed? 👆

[thin carbon]

ah didnt see this, @low roost please follow rule 5 thank you.

[proper quest]

out of all places to advertise this place is the weirdest one xD

[thin carbon]

eh probably a bot honestly

[proper quest]

i mean shes been in the server for 3 years

[thin carbon]

and all her messages were twitch related lmao

[proper quest]

@thin carbon

[thin carbon]

got it thanks!

[proper quest]

@thin carbon

Sad to see a 3 yr old member go

She got cooked 😭

IM SO tempted to click on this image to see if they would get me too 🤣

Hurry before i commit sepukku 😭

@hollow tapir

[thin carbon]

hoops i was busy sorry

thanks for the ping!

[raw bough]

Good morning guys^^

Im using JuggernautXL first time, i wanted to ask for some prompting help with it:( is it the same Score_9 etc. Like in PonyXl, or its just normal text?

[thin carbon]

Hi hi, the scoreing is only for pony models (i recommend you use illustrious anyways for anime so theres no scoring and same prompting)

For the prompts, i remember using natural language

But on civitAI you can see how others prompted with the model

[proper quest]

@hollow tapir

[thin carbon]

Sorry, rule 5.

[warm knoll]

For those who frequently use external API providers, my universal API node now supports language models as well (previously, it only supported image models). This allows you to distinguish between user and system roles in the prompt. The node is controlled by an external JSON file, so you don’t need to modify the backend code if something changes—such as adding a new model or service. The example is a simple prompt enhancer, but it can be used for anything else as well. Manual: https://github.com/CosmicLaca/ComfyUI_Primere_Nodes/blob/master/Workflow/Manual/nodes/uniapi.md

[prime urchin]

dunno if this is the right place to ask but can someone help me out? im tryinig this all for the first time and was hoping someone will dm me so i can get a proper crash course on all this. i dont learn well over just like messages so a vc would probably best if thats alright. thank you!! also im not looking to pay someone or deal with a bot no thanks.

[proper quest]

You are saying a whole lotta nothing

[midnight pagoda]

😂

[warm quiver]

.