✨｜sdxl

[topaz spade]

/generated i use invoke, novaFurryV8b first, then then animeArtflow_v14 for the upscale

/generate Ultra-realistic artistic digital portrait of Radha Rani, close-up front-facing smiling pose, glowing soft white skin texture, sharp facial details, natural pink lips and golden brown expressive eyes, long styled hair adorned with vibrant peacock feather and pastel-colored fresh flowers (pink, orange, yellow), wearing heavy traditional jewelry and beautifully designed ornate clothes in soft pastel pink and golden tones, intricate ornaments with gold and gemstone details, warm pastel color palette, cinematic natural lighting, detailed textures, subtle artistic pastel background painting, professional DSLR depth of field, ultra-HD 8K --ar 2:3 --v 6 --q 5

[hard plinth]

/create: big red mouse

[graceful solstice]

Painting realm

[ancient owl]

what's the size of the latent image do you use? using low resolution images may cause the image to deform a bit. have you tried latent upscaling on A1111?

you can try upscaling the image too. or you can start with decent img dimension like 896 x 1152 768 x 1344 such that. something lower like 448 x 800 may distort the image.
then try adding "deformed" in the negative prompt.

[tawdry badger]

a cat sitting on the blanket

[normal monolith]

trying out new stuff... I really like vibe of those

[glacial star]

One message removed from a suspended account.

[frail lodge]

Anyone know some solid anime inpainting models for SDXL or SD 1.5? Thanks!

[bold sky]

many many beautiful flowers and a cute white cat

[devout flame]

Anyone please help, I'm new how I generate text to image using stable diffusion in discord?

[native knot]

You don't unless you want to pay.

[spare patio]

How do I get SDXL on macos

[boreal oriole]

https://diffusionbee.com/

Try this one.

[spare patio]

Is it reliable?

[boreal oriole]

Of course.

[spare patio]

Thanks g

[boreal oriole]

Welcome anytime.

[turbid urchin]

is clip skip still neccessary for sdxl and other models built on it?

[native knot]

Depends on the model. Some do fine without, some need it to perform.

[rare rampart]

try Draw Things - https://drawthings.ai/. Very fast, very actively developed (eg has Qwen-Image, Flux, WAN 2.2 etc)

[rare rampart]

12 secs - SDXL

[spare patio]

I need sumn that’s gonna remember specific things Y’know? Like how ChatGPT does it

[next folio]

Need help trying to make an angle where the character looks to others a the distance while the character is close to the camera, like looking in profile but in diagonal

[graceful solstice]

Bedroom in a seaside cottage, in the style of Kim Jung Gi

[quaint harbor]

In automatic1111 where do i put the upscaler in safetensor format? The GUI doesnt seem to find it.🤔

[slim wren]

what kind of upscaler ?

if it s gan one, models/ESRGAN

[quaint harbor]

4xultrasharp. Dont think its possible in 1111. In forge its no problem.

[slim wren]

pretty sure it s a gan one, it should work.

[frank garden]

https://i.imgur.com/unHXAPL.png

[slim wren]

read #artisan-faq and stop spamming, thanks.

[south horizon]

Doug?

[faint shore]

Chuck Tanner

[pine glade]

?help

[frigid thicket]

把左边人物衣服颜色换成红色

[rapid charm]

/generate landscape image with aliens

?help

[native knot]

lol, no.

[junior vigil]

https://x.com/5anjoker/status/1968553879148347711?t=TDU859UzqZ_CoVMqhWl0vQ&s=19

5aNjOkEr_XL (@5anjoker)

🌅💠 seductive but untouchable…✨

#5an_AI | #AIart | #AIイラスト︎ |#AIgenerated | #digitalart

X · Sep 18, 2025, 5:53 AM

[junior vigil]

https://x.com/5anjoker/status/1969094022909010371?t=A1puYAgNm1TB7YCN4R97Mg&s=19

5aNjOkEr_XL (@5anjoker)

Catch the tide — and my hat if you can 🌊🎡😝

#5an_AI #AIイラスト #Anime #DigitalArt #AIArt #SDXL

X · Sep 19, 2025, 5:39 PM

[heady pivot]

is this ai? 😮

[earnest viper]

Yes

[heady pivot]

wow, congrats 🫶

[heady pivot]

what model did u use?

[earnest viper]

SDXL + WAN

[heady pivot]

kk ty

[junior vigil]

https://x.com/5anjoker/status/1970730215157702746?t=aU4FncD-PKOSAixZQNPSgQ&s=19

5aNjOkEr_XL (@5anjoker)

Wrapped in gold, veiled in mystery, and crowned by power 🔥👑✨

#5an_AI | #AIart | #AIイラスト︎ | #AIgenerated

X · Sep 24, 2025, 6:01 AM

[opal stream]

Hello

[opal stream]

Caution : Stay away !
Guys i got Scammed by @earnest viper , he showed me these images, I paid $450,
I Have the full chat and full video recording of what happened.
Now he deletes his chat from discord, blocks me and runs away.
So finally i find him here.

[jolly flame]

HOLA

hello

the image in anime digital semirealism

[slim wren]

Pre emptive reminder to keep things classy in here. No nsfw stuff here, thanks. Just in case...

[jolly flame]

So you can create it? I want a digitally drawn profile picture.{

[shy kelp]

@mortal rain

[turbid orbit]

What upscale models do people swear by these days for sdxl?

[faint shore]

When I upscale, I just use the same model from the first sample pass, but send it a higher-resolution latent set with a little noise injection before resampling.

When upscaling pixels, I mainly stick to 2x models like Real-ESGAN. None of the pixel upscalers stand out as being amazing. Because pixel upscale can be applied at anytime, I often skip that step to save disk space. When I find a real gem of an image, it becomes a candidate for upscaling.

[shy kelp]

I use Upscayl

[vernal shale]

A lion sitting at a school desk, wearing glasses, holding a pencil, cartoon style, colorful, funny” ,4k,drawing,wide lense ,perfect ........ why dose it lock so bad?

its real vis 5 and some random lora

cartoon

[wide tapir]

a bald dwarf, muscular body, realistic face, detailed expression of rageA photorealistic masterpiece, shot on Arri Alexa, cinematic it, shaking his head furiously. The camera is handheld, with a slight, almost imperceptible shake, enhancing the raw emotion color grading. A bald dwarf who is an exact likeness of Cristiano Ronaldo is sitting on the floor of a bedroom of the scene. The visual style must match the provided reference images, focusing on gritty realism, deep shadows, and des. His face is a mask of visceral anguish and pure rage, with hyper-detailed, wet skin and realistic tears streaming down hisaturated colors.

[stark sedge]

"its real vis 5 and some random lora" there is your answer

[hasty shell]

A clean futuristic digital illustration showing the evolution from traditional search engines (search bar, magnifying glass icon) on the left to AI assistants (ChatGPT, Gemini icons) on the right, minimal flat vector style, professional tech design, blue and yellow corporate accents, white background, concept of AI visibility 2025, modern business look

[frank marten]

Create a modern scientific laboratory scene with clean white counters, chemical storage cabinets, safety signage (like PPE reminders and hazard symbols), and realistic lab equipment such as microscopes, beakers, and fume hoods. Include subtle lighting and a slightly dramatic tone to suggest a challenge or escape room atmosphere. The layout should be modular and clear, suitable for overlaying interactive hotspots or puzzle elements.

[humble sandal]

What is the best XDSL check points you guys used for illustration nsfw

[late canopy]

hey everyone. I'm trying to do inpaint with an sdxl model + a lora of a character into a specific background image. Now I cant seem to achieve that. I use swarmui. Do I have to get better at my control of specs such as denoise and mask blur etc or is there a better way to do it ? I usually do a remove background of the character and then paste it on said new background that I want but that has problems when I want to animate, as the I2V video gen AI will see the subjects body is not blending well (in the scale of small pixels), for example on a chair its sitting. It will see it as not sitting and the subject my start to fly away as the AI will see it floating, even if by a few pixels, etc. I have discovered that it matters to do a good mask too, not just but a rectangle box where you want the person, and actually try to give it think the mask legs, arms, a head. But I still cant get a good result and am a bit lost. Should I up my prompt game ? Should I mention the background as well ? What to do ? Any helps and tips will be gladly appreciated ! Thanks everyone !

[shy kelp]

ramthrustsNSFWPINK_alchemyMix169

[humble sandal]

Looks good any other one ?

[shy kelp]

I'm not google

[shy kelp]

what model?

[normal monolith]

not a single model. small anime illustration made with some illustrious model evolved with im2img into bigger one with dreamArtFusion

[shy kelp]

ah.... thanks anyway

[ornate rain]

generate island taro

#✨｜sdxl hero

[clear quarry]

🎉 LIVE DEMO: 8-Step > 20-Step Diffusion Breakthrough

It's an implementation of the paper: Hyperparameters are all you need. I've just launched my HuggingFace Space where you can test this yourself!
the link of the paper is below: https://arxiv.org/abs/2510.02390
🔗 Try it:
counterfeit v3.0 version: https://huggingface.co/spaces/coralLight/Hyperparameters_Are_All_You_Need
xl version: https://huggingface.co/spaces/coralLight/Hyperparameters-are-all-you-need-xl-version
original sd version: https://huggingface.co/spaces/coralLight/hyperparameters-are-all-you-need-sd-version

some examples is below

✨ What you'll see:

• Side-by-side with DPM++2m
• 2.5x faster, BETTER quality
• Works with any model

📊 The impossible made possible:

• 8 steps generate images with FID performance comparable to the 20 steps
• No training/distillation needed
• 60% compute reduction

Would love your feedback! What models should I test next?

[dapper wasp]

are you the paper author or did you just implement what's in it?

[clear quarry]

I am the author

[dapper wasp]

very cool
my suggestion after skimming is, DPM-Solver++ while popular is not SOTA for ODE by any means; res, unipc, gradient estimation, etc. all boast superior FID scores in their respective works

also you included turbo, lightning, but not hyper for some reason

[clear quarry]

We report DPM-Solver++ because it's a popular baseline in 20-step inference. You are right that UniPC is also very popular in few-step inference, but we want to prove that our training-free solver in few-step inference can beat the most popular ODE solver DPM++ 2m even when the latter one uses 20-step inference. We also compare our solver with the latest training-based ODE solver AMED, and our performance is also better than that in the few-step inference. Meanwhile, our algorithm provides a plug-in asset that is suitable for most of the training-free solvers, like unipc. And by using our algorithm in dpm++2m, the 8-step inference result beat the 20-step inference result.

The comparison with DPM++2m follows with the latest jobs in training-based ODE-solver, like AMED and EPD.

Meanwhile, we also include Flash Diffusion, the latest distillation algorithm proposed in AAAI 2025, and DMD2, not just the lightning and turbo. I think the main FID bottleneck of diffusion distillation is caused by the diversity, and the algorithm using DMD loss is a natural choice for me, since it aims at distilling the distribution rather than a single sample. Hence, I chose these two algorithms, rather than hyper. We also provide some insight into why diversity will be reduced by the distillation algorithm, and do some experiments to prove diversity really matters by using PRD metrics to measure the latest distillation algorithm.

Oh, I see, the Hyper-SD also uses the DMD loss. But we also compare with DMD2, I think it's acceptable.

[mental tundra]

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?

[leaden sigil]

Hi,
I’m trying to place a glass bottle in a new background, but the original reflections from the surrounding lights stay the same.
Is there any way to adjust or regenerate these reflections without distorting the bottle itself?

[clear quarry]

Actually, DPM V3 also utilizes the technique proposed by UniPC. Our latest code in GitHub combines the DPM V3 with our algorithm. You can also check it. Our 6-step inference result is better than its original 8-step inference result.
Check it there:https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/txt2imgV3.py
Some of the results:
results from the normal dpmv3 8-step inference:
https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/outputs/txt2img-samples/normal-v3-result-8-steps/00000.png
https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/outputs/txt2img-samples/normal-v3-result-8-steps/00001.png
https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/outputs/txt2img-samples/normal-v3-result-8-steps/00002.png
https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/outputs/txt2img-samples/normal-v3-result-8-steps/00003.png

results from our 6-step inference combine with dpm v3:
https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/outputs/txt2img-samples/samples-6-1o-f2/00000.png
https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/outputs/txt2img-samples/samples-6-1o-f2/00001.png
https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/outputs/txt2img-samples/samples-6-1o-f2/00002.png
https://github.com/TheLovesOfLadyPurple/Hyperparameters-are-all-you-need/blob/main/outputs/txt2img-samples/samples-6-1o-f2/00003.png

[tough stream]

Guys, is there a faceswap workflow?

[high skiff]

Hey guys, has anybody else noticed SDXL being slower than ever lately? I remember I used to get like 4.1it/s on my 3090, now I can't even hit 3it/s anymore. I thought comfy recently got an update that was supposed to speed up image gen

[faint shore]

I've been using the 4 step model. It's quite fast.

[proud sierra]

美化一下

[edgy smelt]

[amber escarp]

Where can I find advice/instructions for installing SD xl to a new system?

[gray lagoon]

#🤝｜tech-support in the pinned messages by Cs1o

[amber escarp]

thank you

[blazing cipher]

very dreamy

[heavy lintel]

A lot of help

[heavy lintel]

Art

[vale vapor]

What model and prompt did u use for these. They are gorgeous

[normal monolith]

I used Dream Art Fusion. older model, but still my fav for realistic images. just mind that it doesntvwork well with initially big images. no loras were used, just a bit inpainting and postprocessing

[vale vapor]

What prompt did u use

[normal monolith]

I don't share my prompts

but since this model recognizes quite a lot of celebrities, try adding 3-4 of names... this model quite nicely blends features of named people

[inland maple]

hello

[heavy lintel]

very surprised

[glass forge]

sdxl still going strong

XD

[bleak badge]

or do you have know any model that the architecture was the config.json is in the root?

[sterile carbon]

Have tried downloading the sdxl model from civit ai and put it in model/checkpoints

[winged spear]

Trying to learn controlnets in sdnext... I'm getting black outputs despite following the doc pretty much exactly.

[faint shore]

I got black outputs on qwen when I launched with sage attention enabled. If you're using that, try launching without it.

[fathom musk]

guys where can i access sdxl 1.0 deployed version?

this is for research purposes

[slim wren]

deployed version ? deployed where ?

[glass forge]

It's been fun! But Z image is probably going to take over SDXLs role fast.

its just as fast and so much better

🤯

[gray lagoon]

Unless z image anime ver comes out (waiting) illustrious is still relevant

[shy kelp]

I use CyberRealistic Pony and im goonin to every images i generated.

[molten gull]

what's with flux2 ? any good?

[random swan]

does anyone know how do i get a hold of marginal vectorfield parametrization from Flux ?

2. can i run it on 8G vram ?

[dense oasis]

Dream

[shy kelp]

yall better try z-image it's much smarter than sdxl.

[high canopy]

baby

[clear quarry]

I want to do some propagation. I develop a new fast ODE solver that suitable for the diffusion model. And the huggingface repo is there: https://huggingface.co/spaces/coralLight/Hyperparameters-are-all-you-need-4k
You can sample a 4K image in 8 steps, and its' result seem as better as 30 steps dpm++2m SDE with karras. And the 6-step inference result is also acceptible. The paper is there: https://arxiv.org/abs/2510.02390

The dreamshaper XL version is there: https://huggingface.co/spaces/coralLight/Hyperparameters-are-all-you-need-UniPC-Dreamshaper

And there is the nova anime version: https://huggingface.co/spaces/coralLight/Hyperparameters-are-all-you-need-UniPC-XL

This is a training free solver that compatible with all of the U-Net latent diffusion model.

[peak atlas]

what is z-image? a model?

[peak atlas]

Oh ok, thankyou

[coral violet]

where can i read more on this diffusion
i want to do a project for a client and i have to use the stable diffusion

is there like a blog official one that i can look at?

[slim wren]

cf https://stability.ai/

[coral violet]

thanks

[forest egret]

a cute dog sleeping in a blanket

[jagged depot]

#✨｜sdxl of

#✍🏼｜rules-and-tos one

[toxic wasp]

Here's another one.

[royal tapir]

#✨｜sdxl dog

[gray lagoon]

🐕

[glass forge]

It's been fun.

[glass forge]

It's the future.

[peak atlas]

interesting, is this kind of model available in civit.ai?

[glass forge]

oh boy is it

it already has a ton of loras

[peak atlas]

aight, will wait for NAI and my fav model to come out with ZIT model

thanks for the info

[next folio]

Anyone knows of any way or online tool to train loras faster? don't have the hardware for it and all the options i found take almost an hour and most times aren't very accuarate. Also, no longer using Civitai, too few credits i receive to train.

[gray lagoon]

A hour seems reasonable though?

[lilac tulip]

This is too much for me.

[glass forge]

yep

sdxl is now abandoned

since Z dropped no more posts

[gray lagoon]

more like this server tbh, granted sdxl was already dying due illustrious, flux and others

[pure stream]

What the heck happened to this server?

[gray lagoon]

lack of moderation, lack of community interaction etc

[glass forge]

z image

[normal monolith]

there. something made with sdXL... Actually I use only SDXL models, but illustrious mostly so it lands in anime...

[glass forge]

Emad had great mascot potential.

[ionic cradle]

hi all

[wide turtle]

/dream一个小溪旁边坐着冥想的女人，小溪清澈透明，周围绿草丛生，树木，落叶，

[glass forge]

Yes this was Emad's dream.

[wide turtle]

/dream一个小溪旁边坐着冥想的女人，小溪清澈透明，周围绿草丛生，树木，落叶，

[dry viper]

6666

[normal monolith]

she's a bit 6666 herself

[waxen crag]

any advice on better servers? looking for a nice community, maybe with daily contests etc...

[gray lagoon]

Ill dm you one

[novel heron]

Hello!
I'm having an issue with something on (I think) version 1.5 of stable diffusion, is this the right place to ask?

[rare socket]

Anyone recommend a good illu checkpoint model that has good base quality and anatomic consistency? Pref one that doesnt deform character loras. Or a noob checkpoint that works well with illu loras

[normal monolith]

you want more anime style or 2,5D, semi real?

then again, I don't use character loras, so I'm not sure if I can recommend anything specific

[rare socket]

Hmm i dont really know how to describe it well to be honest, its kind of in between anime and 2.5D. Kinda like anime with detail and quality to a point it looks a bit 2.5D. Something like morimee style from civitai

[normal monolith]

would check out morimee, but civitai is having tech problems again

something in style of one of these?

i prompted very generic prompt in a few of my favorite models

[rare socket]

Kinda like this ye

[normal monolith]

one of the left is diving anime, the one on the right is electrum cinnamon, both of them on civitai and both very solid

for more flat look I could recommend animij

[rare socket]

how's the situation look with anatomy? hands, feet, extra limbs etc

[normal monolith]

both are solid, sometimes 1 finger more, but usually good enough

actually when it comes to illustrious, I couldn't find better ones

ymmv

[bitter frigate]

hello! could i dm you for some advice please?

[normal monolith]

you can ask here

[bitter frigate]

thanks! im really new to all this and ive gone on civitai and grabbed some loras and models, however my images generated are really bad even when copying prompts, am i missing something?

ive noticed some models are under "trained checkpoints" and some are "checkpoint merge" is there any difference to it?

[normal monolith]

I never worried about that

[bitter frigate]

for example, im using this model rn https://civitai.com/models/827184?modelVersionId=2167369

[normal monolith]

can you post an example of what you made?

btw I tried to like that model, so far no luck

if you could post image and a prompt you used?

[bitter frigate]

it comes out like this

not sure what im doing wrong as im copying the prompts based on the image posted on civitai

[normal monolith]

not entirely sure, but it looks like: 1. the resolution is way to small for that model, 1024x1024 is probably a better starting point. 2. you seem to have some sort of artifacts on the image, and that is strange, but first try larger resolution

[bitter frigate]

ok thanks

[normal monolith]

what gpu do you have, how much vram?

[bitter frigate]

i have a 4070 with 12gb vram

[normal monolith]

let me know how it came out with bigger resolution

[bitter frigate]

it came out like this with 1024

[normal monolith]

better, but you still got artifacts, face should be readable in that resolution

you're using loras? maybe they are the problem...

btw for euler ancestral and illu models i rarely go above 30-35 steps

try generating something that doesn't require loras, and see if results are better

[bitter frigate]

im getting this now using 1 loras

[normal monolith]

now it looks ok

loras are a bit tricky, you need to see how many you can load, and how much of each you can add before the image gets messy

[bitter frigate]

thanks, how do i improve on these

as its still quite far from the image on civitai

this was the image with the prompts on civitai that was posted

[normal monolith]

at least your's has a proper number of fingers xD

[bitter frigate]

true

do i have to get other stuff as well?

[normal monolith]

what other stuff?

[bitter frigate]

like loras for styles etc

[normal monolith]

with illustrious you've got A LOT of styles embedded in the model at start

unless you've got very specific need, you can go with one model for months xD

for example, try adding at the end of your prompt "image by bacun and rariatto" and see what you get

[bitter frigate]

thanks, do i just need to improve on my prompts?

ok

[normal monolith]

depends on what you're after

[spark hull]

yeah hate to say it but, theres alot more going on in the civitai discord. its always got people chatting, and theres always media being uploaded there. not sure why

[jade raptor]

https://civitai.com/posts/26065901
https://civitai.com/posts/26066307

[glass forge]

I'll never forget you my friend. you were my first. Well technically it was sd1.5 that was my first.

[wintry steeple]

Hey everyone, I need some expert advice on my ComfyUI workflow for a realistic ai character.

Current Setup:

• SDXL LoRA (self-trained)
• Checkpoint: Juggernaut XL
• Face Detailer with a separate prompt

I'm struggling with facial consistency across different prompts (the face looks slightly different depending on the setting). I have a few specific questions:

1. Trigger Word in Face Detailer: Do you recommend using the LoRA's trigger word inside the Face Detailer's separate prompt as well? Or is it better to keep the Detailer prompt generic (e.g., "high quality skin, sharp eyes")?

2. PuLID for Consistency: Does it make sense to add PuLID into the mix to "lock in" the identity? If so, does PuLID usually play nice with an active LoRA, or do they tend to fight each other for control?

3. Face Detailer "Guide Size": I’ve seen conflicting info online. Some suggest 512, others say 1024 for SDXL. What is the sweet spot for Juggernaut XL?

4. Denoise Settings: What's your go-to denoise strength for the Face Detailer when trying to maintain a very specific trained identity?

Any insights or workflow tips would be greatly appreciated! Thanks in advance! 🙏

[glass forge]

Just use z image

[kind bronze]

too true. Z-Image being soooo good had made me almost stop using Flux AND Ideogram )

• SDXL is very much but a legacy historical artefact at this point.. but you can use it for very affordable refiner / post processor .. and yeah, some checkpoints are still competitive for specific styles

[faint shore]

I still come back to SDXL because there is no replacement in any of the new models for the IPAdapter's style transfer. No other model looks as good as SDXL in this category.

[deft ginkgo]

I want to share my story for anyone who might be in a tough season right now.
Not long ago, I was working as a dispatch rider when a serious accident left me unable to walk for six months. Overnight, everything changed. Providing for my family became a daily struggle, and mentally, it was one of the hardest periods of my life.
Through it all, my wife never gave up on me or our family. She stayed strong, supportive, and kept us moving forward when things felt uncertain.
During my recovery, I started searching for a way to earn online. That’s when I discovered Shopify and automation. I didn’t rush or look for shortcuts I followed the training step by step, stayed consistent, and trusted the process even when results were slow at first.
By combining Shopify with simple automation systems, I was able to build something that worked even when I wasn’t physically able to do everything myself. In the last 30 days alone, this system generated $37,000 in revenue.
Today, I’m grateful to be able to provide for my wife and children again. More than the money, I’m thankful we didn’t quit during the hardest moments. Shopify and automation gave me a real second chance, and I share this to remind someone else that a breakthrough is still possible if you stay patient and committed.

[glass forge]

so no one moderates here anymore

@mortal rain spam alert

[azure seal]

Hi. If you need help with training Flux/SDXL Face Lora you are welcome.

https://www.behance.net/gallery/243708697/Stable-AI-Influencer-Private-Flux-Face-LoRA

[hardy trench]

16 emoticons of chibi orange tabby cat in 4x4 grid, Chinese New Year themed, each panel showing different festive action:
(1) cat clasping paws greeting with golden "Fu" above head,
(2) hugging big red envelope with "Bao Fu",
(3) jumping with sparkler, night fireworks,
(4) eating sticky rice cake with crumbs,
(5) spinning chasing fluffy tail,
(6) covering mouth giggling with lantern light,
(7) sticking crooked Spring Festival couplet on door,
(8) kneading on red cloth with peanuts and dates,
(9) hugging big fathead fish proudly,
(10) curled in fur ball exhaling warm breath, snowy background,
(11) shaking bronze bell with tail swaying,
(12) lying on back paws up asking for petting,
(13) holding small lantern exploring night alley,
(14) rubbing owner's neck in cozy home,
(15) carrying willow branch knocking door,
(16) three mini cats stacked with tiny cup, family silhouette,
uniform panel size, clear borders, round edges, vibrant red and gold color palette, cute cartoon style, high detail, clean lines --ar 1:1 --style cute --no text, watermark, signature

[fading scaffold]

Hey guys, anyone has !!colab!! notebook for sdxl lora training, which is current and works? Also needs BLIP captioning? Can someone share link on git?

[high meteor]

Hey not sure if it’s the right place and I’m sorry if it’s not allowed I can remove this message. For those who want to learn how to use SDXL in A1111, use NSFW checkpoints, best settings, how to create a LoRa, I create a full training about it : https://www.skool.com/ai-influencer-model-training-4492/about?ref=8a94f10b02214a008aac821766685904

[ashen hare]

Aesthetician Logo for Website.

doh

[azure seal]

Hi. Who knows why on Comfy there is such a mistake on KSampler node??

What's wrong?..

im trying to generate img2video on flux checkpoint

[faint shore]

You're right it should work.. Because you're posting in the SDXL channel, I assume that's what you're trying to generate, but I notice you're using a Flux model. Try supplying an SDXL model in your current workflow before resorting to updating. Otherwise, exit Comfy and update.

[random panther]

I like the skin tone that came out of this it is very realistic tones, could I poke you for the prompt please? thanks!

[faint shore]

@random panther Try downloading the attached image and dropping it into Comfy. The prompt may still be there. You can also use an auto-prompter to help with prompt development. The fastest Comfy auto-prompter is Florence2, but it's a little limiting in what it kicks out. Another to try is Janus. Also give Qwen a try, more robust, but slower. I often add an age sentence to any prompt that kicks out people. Also consider giving people an occupation or title instead of just "man" "woman" or the worst "1girl."
A 59 year old house wife.

[tame junco]

loras ?

[faint shore]

The only lora you really need with SDXL is not a lora at all. Use the IPAdapter to transfer styles from one image to another.

[tame junco]

sd_xl_base_1.0.safetensors ?

Is it possible to create LoRAs using Colab?

[lethal vapor]

Need help

Flux lora generate

Hello guys am new to this stable diffusion world. Am a graphics designer, i want some high quality images for my works. So i want to use flux. Is anyone free to tech me how to generate a lora model for flux. I allready have automatic 1111 and kohya ss installed please help me a little guys.🫠🫠🫠🫠

[faint shore]

@lethal vapor Why are you posting for flux help in the SDXL channel?

[lethal vapor]

Where can i post ?

[faint shore]

Stable Diffusion doesn't even make Flux. Try joining Pixaroma discord if you haven't. You'll find a lot of help there.

[gray lagoon]

Or the OneTrainer discord, works great and a1111&flux wont work iirc

[random panther]

I tried it wouldn't read the webp. wasa simple enough promt. thanks.

[faint shore]

@random panther A Discord Tip. The first click on the image is a .webp. If you click again, however, you'll get to see the original posted image at full quality. This is often a .jpeg or .png. Right-Click Save Image as. Only the full-quality image contains any workflow. Try supplying the full-quality image to Comfy.

[random panther]

I never actually knew that about this dscord thing. i hate using it tbh (old man - this is way to much clutter). but thank you i know how to get the source pics now!

[chrome oyster]

Is there any models for controlnet for sdxl/illustrious?

[faint shore]

Try using the all in one. Search for AIO.

[next folio]

any recomendation for inpaint models? tried with wai illustrious but the characters merge and distort, hope someone can help

[fluid moon]

Yes, can help you that!

[lethal vapor]

help guys why i am getting so bad results

[gray lagoon]

No negative prompt is a start, could try messing with a different sampler

Go on civitAI under juggernaut and copy someones settings

[lethal vapor]

i am still not getting good results please help

[lethal vapor]

I have a 2060ti 12gb vram and 16gb systm ram what model should i use ?

[lethal vapor]

So will it work on my pc

No man there is a 12gb version

I have it

Thanks

But sorry to disturb you again

I am new to this

Can i get download links of those please 🥺

Ok 🙃

I use forge neo

Its update of a1111

For lora training what you use ?

It need to have continuous internet connection right ?

Internet connection

Whenever i disconnects its like turns of

Man am dont know what happens it just turns off

Ok

Ok it just download let me see how it goes

🥲

🥵

Oh bye the what the settings will be ? 🙃

Haha its heavy still stuck here for like 2 minutes

[lethal vapor]

My system ram used like 98%

But vram at 7-10%

Lol 🙃

Bro forge and a1111

Is same

Any tips ? 🥲

got it working its fast now

just used the same text encoder but the fp8 version

but now the problem issssss

she is not very detailed

no i mean not the image qulity

i am talking about the detailes in her face

ok

i have a very noob qustion lol

is my juggarnaut xl lora not gonna work with this ?

ok thanks

thanks a lot for the help man

no that images was not good anyway

claude meaning ?

oh

by the way what u do man ?

😒😒😒

i mean what u do for living ?

your a fun guy man

where u from ?

🤣🤣

ger meaning germeny ?

can i ask for little bit more help from you ?

lamo

need help with aitoolkit lora traning for z image

😒😒😒 i see you got a high spcs pc

ok

me to

but my father was brock

his father too

and now i am earnig

me too borck as fuck 😊😊

lmao

friend no one is replying there 😢😢

[lethal vapor]

Its really hardbro they are not helping much

I dont know what is happening with my aitool kit

Why its asking for internet connection

Probably

[lethal vapor]

Is the downloading from hugguing face necessary?

?

[gray lagoon]

Yes useually if its missing essential stuff to run a new model

[lethal vapor]

Oh so first time i have to download it ?

[lethal vapor]

can i use one trainer for z image lora traning ?

[gray lagoon]

i recommend asking in the one trainer discord

[ebon niche]

Hi everyone,

I’m currently working on creating an AI influencer using ComfyUI (SDXL).

I already managed to get very good and consistent results for the face using PuLID (the identity is almost perfectly preserved across generations).

My current issue is with generating the full body.
When I try to add ControlNet (pose) and generate a full-body image, I start losing consistency between the face and the body, or the proportions / realism don’t match properly.

What I’m trying to achieve is:

• keep a 100% consistent face identity
• generate different full-body poses
• control proportions (slim waist, wider hips, etc.)
• maintain realistic skin / lighting consistency between face and body

My current setup:

• PuLID
• IPAdapter
• ControlNet (OpenPose / DWPose)
• SDXL checkpoint (RealVisXL / Juggernaut)

If anyone has a simple and stable workflow (not too many custom nodes) that works for consistent characters / AI influencers, I’d really appreciate it 🙏

Thanks!

[silk cedar]

https://huggingface.co/Bl4ckSpaces/AllInOne-XL-V12.5

[hot depot]

ghbdtn

[silk cedar]

Hi everyone, I just released my latest custom merge, Animagen XL v3.
It’s a specialized 2D/anime SDXL model built from 9 different models (including Illustrious, NoobAI, and Pony). Instead of a standard flat merge, I used custom block-specific U-Net routing and targeted Text Encoder adjustments to fix common multi-model merging issues.
⚙️ Technical Highlights:
Pure Danbooru Prompting: The Pony Text Encoder was deliberately set to 0% weight to eliminate syntax conflicts. You don't need to use score_9 or source_anime. Standard Danbooru tags (masterpiece, best quality, etc.) work natively.
Deep Block Routing for Anatomy: To prevent style models from degrading human anatomy, the deepest U-Net blocks are heavily weighted toward NoobAI, Illustrious Janku, and Pony V6 for better hands and dynamic pose stability.
Mid-Block Routing for Faces: Illustrious Wai and Hassaku are prioritized in the middle blocks to maintain clean, highly detailed facial features and eyes.
💡 Quick Settings:
Sampler: DPM++ 2M Karras, Euler, Euler a
CFG Scale: 7.0 - 8.5 (Keep it at 7.0+ to prevent color desaturation)
Steps: 25 - 40
I've attached some raw unedited preview generations below. Feel free to test it out and let me know your thoughts or feedback!
🔗 Link & Full Documentation: https://huggingface.co/Bl4ckSpaces/Animagen-XL-V3

[lethal vapor]

How can a lora file be corrupt or invalid

[gray lagoon]

download going wrong

[faint shore]

Or your hard drive is going bad.

[lethal vapor]

Yea thanks man i found the dowaladed file was not completed

[wheat hamlet]

Guys help I get an error on ComfyUI setup I can't even troubleshoot

[glass forge]

https://www.youtube.com/watch?v=LPZ-pErAM8Q

[twin shell]

🥂 oh.. turns out... hes a deck supposedly. 😭

[next folio]

Any news in lora training for XL models? i havent trained one in almost a year. Mostly looking to train illustrious artstyle loras, i've seen a bunch of tools on youtube but seems quite old and not sure how good are the results. Any recommendation for tools or checkpoints for good results in artstyle training?

[worldly pawn]

You can use flux or sdxl

They are good for Lora training

[next folio]

yes, i said illustrious for a reason

[worldly pawn]

For Illustrious style Lora, the current best approach is still Kohya ss with SDXL, but using newer tricks like LoRA+ and lower network dims (around 64). Illustrious already has strong style priors, so shorter training (6–10 epochs), minimal captions, and a well curated dataset matter more than heavy tuning.

[worldly pawn]

Base model choice is key Illustrious or Animagine XL works best for illustration styles. The workflow isn’t radically different, just more refined for better consistency and less overfitting

[slim wren]

rule 2 please.

[wheat hamlet]

Base model illustrious sucks

go get a fine tune

[pine oyster]

Guys I am a rookie, and I have a question. I am using SD Forge, and for some reason the images generated every time have the exact same face. I am using juggernautXL_ragnarokBy.safetensors and the seed is set to -1, maybe because I was trying to use ControlNet? But I turn it off later, can any expert tell me what could be the problem and how could I fix it? Because I used to be able to generate random faces.

[stiff saddle]

Hi, quick question, if I upload some AI images somewhere on this server, is there any chance someone could recognize the art style/model/checkpoint and be able to tell me what it is?
I’ve been trying to find images and use their metadata to see, but they come up blank with SwarmUI and Forge Neo.

[gray lagoon]

why dont post em anyways and hope for the best,
getting precise things are probably impossible

[shy kernel]

I just joined the AI own generation scene again and i have no idea with what model to start

[red vale]

See what you like, choose what you like

[tepid ginkgo]

Can anyone give me any tips on how to get multiple characters in one image? I consider myself to be pretty good at using Automatic1111/ReForge but I can never for the life of me figure out how to use multiple character LoRas in one image. I've tried things like Regional Prompter and Inpainting (I'm terrible at inpainting) and nothing has worked so far. If anyone here regularly creates multi character images I'd love to know your process (I'm also experienced in ComfyUI so if you use that I'd still love to know how you do it!)

[ashen shell]

would you like me to teach you or do it for you?

[slim wren]

be careful of dm scammers

[glass forge]

Maybe use Klein flux to put more than one characters in one scene. I didn't try yet but I think it might work.

Place them roughly where you want them in the scene tho in photoshop.

Early days. but it's possible.

[glass forge]

https://www.youtube.com/watch?v=-CNqDIKTC2Q

Play it loud.

So emad can hear it. This is all his fault.

[tepid ginkgo]

Well to be honest I already know how to do it. I can create pretty much any image I like with 2 character LoRas as long as it's not like insanely complex. My issue starts at upscaling. i usually generate at 1024 x 1024 or any aspect ration stemming from that. My base images are usually perfect both characters show up just fine without any mixing. However when i try to take my image to 2048 x 2048 through a latent upscale pass that's when it starts to mix the LoRas and I have yet to find a fix to this. I've tried controlnets, low denoise (usually doesn't work during an upscale you typically need a higher denoise to do a latent upscale or it just comes out blurry), and inpaint how ever inpaint usually requires me to do an img2img pass after to fix any small issues inapint causes but that just takes me back to square one which is img2img causing mixing.

The TLDR is I need to find an upscaling method that wont mix my characters.

[ashen shell]

Yeah that’s a pretty common issue with multi LoRA images When you do a latent upscale the model isn’t just enlarging the image it’s partially reinterpreting it With higher denoise, both loras start influencing the whole frame again which is why the characters begin to mix

the cleanest workaround is to avoid doing a heavy latent upscale to the final resolution Instead do a very light latent upscale just to add structure then switch to a pure pixel upscale (like ESRGAN/UltraSharp) so the identities stay locked If you still need refinement do a very low denoise masked pass per character with only that character’s lora active

Also lowering your LoRA weights during the upscale stage helps a lot strong weights tend to cause cross contamination at higher resolutions
That approach usually keeps both characters stable without falling back into the img2img mixing loop you are describing

[tepid ginkgo]

I actually ended up figuring out a really good work around and am now able to create complex images with multiple LoRas with ease. I wrote a guide on civitai to help anyone who's having the same problems. https://civitai.com/articles/28071

[ashen shell]

okay that is greate

[vapid creek]

what model are you using?

[rain wagon]

Flux and lora trained in dataset

[vapid creek]

so your own trained lora?

[rain wagon]

💯💯

[vapid creek]

what gpu do you have and how did you make it?

[rain wagon]

Send me a req, let's discuss better

[gray lagoon]

Be careful of any dm scams, they are rampant here

[vapid creek]

I will

[blazing belfry]

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.

[blazing belfry]

its my first time here , idk what i need to do, my english is trash, i drop here just searching a img generate. i barely undertand a lot of words but i got all that usay ❤️ and im so gratful and glad for u anime.

idk u but u had a bit of my heart

[glass forge]

SDXL still kicking ass.

[twin shell]

I need to finally in depth embark down this rabbit hole 😭

And properly route so my gpu vram handles it & not overflows to ram 😭😭

3080 dont fail me now

bricks lol

[wet plover]

Anyone know what batch size i can safely use with 5070 TI (16gb) when generating images in 960x1280 using hires?

[red vale]

https://tenor.com/view/head-scratcher-confused-perplexed-huh-clash-with-cam-gif-10198268

[wheat hamlet]

IS IT POSSIBLE to get Illustrious models running on an S25 EDGE

[drifting cove]

设计一个手表

[wheat hamlet]

Which model should I use for pixel art sprite generation? Illustrious fine tunes don't work

[wheat hamlet]

which one? flux.1 kontext dev, dev, dev quantized, schnell, schnell quantized

[fading ibex]

where can i download a working comfyui fully loaded

[sick arrow]

You can get ComfyUI from its official GitHub repo and set it up fresh. Most “fully loaded” packs online are modified and often break or come with issues so a clean install + adding only the custom nodes you need is the most stable way. @fading ibex

[vapid creek]

Hey can anyone help me with eyes? everytime it generates an image the eye are always f'd up i tried other models, alot of other loras, also im using comfy ui with zluda so the face detailer is not working or im doing something wrong. please help me fix this issue, im using an sdxl checkpoint

[hollow nimbus]

Yeah i can help you with it

[gray lagoon]

Be aware not to download stuff or follow any links people send you in dm's! Anyone who really wants to help, would do it here

Hmm you could try disabling the loras when doing the adetailer

But eyes are always a tricky thing with illustrious

[twin shell]

Virustotal scan any links to be safe. 👍 & even then its not 100% certain

[vapid creek]

so what should i do?

[ashen shell]

I have work with this many time and I know what cause it I can help you with that that the eye will be good and perfect

[vapid creek]

dont tell me you're also going to ask money for it

[ashen shell]

Okay no problem

[vapid creek]

wdym 😭

[gray lagoon]

also a bot account, they linger around for a bit

if adetailer doesnt work, at most you can try to do a manual inpaint after upscaling the image (to increase the amount of pixels it can work with)

or adetail after upscaling

[vapid creek]

the problem is i need to make atleast 30 images in a day and i dont have enough time to manualy inpaint everyone

[vapid creek]

adetailer and face detailer is giving me straight errors

[gray lagoon]

only 30?

try to see if you can get swarmUI working with the comfyUI zluda backend then impainting is super easy

and eh the segmentation tool (adetailer) is better imo

[vapid creek]

i mean each images take about 58-160 minutes depending on the prompt so...

i want to do more but i need to fix the eye issue first

[gray lagoon]

and your using illustrious? thats.... impressively long..

please tell me that your using a autocomplete feature 😭

[vapid creek]

im new to this so im using a very basic txt to img workflow and upscaling it with remacri at the end

[gray lagoon]

explains why your taking so long.
i really recommend using swarmUI with the comfyUI backend using zluda.
theres just a lot of features that will make your life a lot easier

[ashen shell]

But you said I should not tell you

[gray lagoon]

the first one is less comfyUI noodles and just a useable UI

the second one is autocompletion, better adetailer (that works on more stuff then just faces)

[gray lagoon]

meaning you were gonna ask money for it. vile ngl

[vapid creek]

okay so what should i search for this or do you have any guide that i can use

and is it going to take another 10-15gb? 😭

[gray lagoon]

im not familliar with AMD cards but i recommend the swarmUI discord. it should be able to just hook up to your existing comfyUI backend so it shouldnt take up that much space.

https://github.com/mcmonkeyprojects/SwarmUI but again im not that familliar with zluda but it should work

[vapid creek]

can this make the eyes fix?

[gray lagoon]

^^

currently im at work but if you manage to get your existing comfyUI linked to swarmUI. once im home i can show you how to work the thing

[vapid creek]

can the swarm ui run on python 3.13?

[gray lagoon]

SwarmUI uses the comfyUI backend. if comfy handles it. so does swarm

swarmUI basically is a layer over comfy

[vapid creek]

okay

its not python 3.13's problem right? (the eyes)

[gray lagoon]

yes its not a pyhton problem.

[ashen shell]

Yep

[gray lagoon]

well this is not a place for predatory "support" smh

[vapid creek]

there is one other guy too from this server

he was trying to convince me to borrow money from my friend

https://tenor.com/view/higuruma-hiromi-higuruma-jujutsu-kaisen-jujutsu-kaisen-meme-jjk-memes-gif-10370496866020481076

[gray lagoon]

i see a lot of these. they use LLM's to first scope out people in need and once someone bites they swipe in personally

[vapid creek]

okay so this is what i got

[gray lagoon]

you dont have dotnet installed

[vapid creek]

wat does that even mean

[gray lagoon]

simpler terms : a program using a simple version of chatgpt to give surface level advice and convince you to pay for "help"

[vapid creek]

ong it felt like ii was talking to chat gpt

like saying that would magically increase the amout of money i have in my bank

[vapid creek]

there is also one error at the end can you check that one too

[gray lagoon]

whered you place it?

in what folder?

[vapid creek]

umm

[gray lagoon]

Should be fine

Weird, try relaunching it, otherwise i recommend joining the discord listed on the github page

The open help channel / help thread there will get you a long way

[vapid creek]

here is the new log but still getting the error at the end

[gray lagoon]

^^

[vapid creek]

oh you cant see the error

[Error] Self-Start ComfyUI-0 on port 7821 failed. AutoRestart ignored as this was an initial launch failure.

here ☝️

[gray lagoon]

im busy at work so i cant really deep dive into the logs

[vapid creek]

ahh sorry sorry i forgor can you tell me when you'll be available

[gray lagoon]

seeing how you gotta link your own comfyui instance to swarmUI first, after that i can really help with the automatic inpainting etc

5-6 hrs

[forest imp]

Rip. He got hacked.

[vapid creek]

who?

[forest imp]

Looks like the earlier message got deleted.

[vapid creek]

Oh

[bronze ridge]

Hello, ya all

has it now become possible to train quality Illustrious LORAs locally on 6 GB vram and 16 GB ram? Cannot currently buy expensive shiny equipment

[gray lagoon]

its technically is possible but eh its not gonna be super

your probably gonna be training on 768x768 and take a long long time

[bronze ridge]

🙁 That is what I read about as well, I was hoping some people here might have experimented with it a bit already

I wanna be able to train character loras 😭 with around 30 - 40 high quality images

[gray lagoon]

hmm weird question but have you tried cloud stuff instead?

hmmm i think 16gb normal ram wont work

[gray lagoon]

anyways i think you might be on the edge of it being possible or not but if you want to be sure i recommend visiting the Onetrainer discord to ask just in case

[bronze ridge]

thank you !

[bronze ridge]

would I have to pay using credit card etc?

[gray lagoon]

hmm depending on the service but iirc 10 bucks should be able to do 2 loras on civitai for illustrious depending on the settings you use

or atleast when i last made one

[bronze ridge]

civitai changed their payment model hence I am trying to flee to local gen 😅

[gray lagoon]

hmm only for NSFW gens, on civitai.com uses " green buzz" and is sfw only

and uses creditcard

otherwise something im less familiar with would be renting a cloud GPU by the hour and train loras on there. if you do a lot of loras it might be cheaper to do that

[bronze ridge]

Yea, I just wanna train every now and then, hopefully something cool happens—

[gray lagoon]

Hmm the newer models are getting heavier and heavier I'm afraid

[loud moth]

yea i just realized i may have gone buck wild downloading models and need to be sure i don't start filling up my data store lol

[faint shore]

For SDXL, you really only need one or two. I favor the 4-step models because they render fast. Try RealVision or Juggernaut.

[gray lagoon]

haha yeah ****

[loud moth]

stop spamming with that bs.

[faint shore]

Insta block.

[elder fog]

Is it normal if I am using sdxl but its super slow on a rtx 2070 and isnt even touching the vram?

20 steps in about 2 minutes 16 seconds

[slim wren]

you're outputing at 1152x2016 (without upscaling I m guessing) so :
1/ yes it s gonna be slower
2/ also it s gonna be buggy, you ll have some duplicate artifacts because sdxl is trained for 1024x1024
3/ you can still output at higher res but you ll need to do it in two pass, one at lower res and the second one being an upscale pass. (there s probably an "hires. fix" option for that in your UI)

[woeful kernel]

You are on CPU.

[gray lagoon]

Please no advertising, rule 5

[loud moth]

is there a great breakdown of all the sampler types?

this is my list

[narrow tendon]

Copied your list via screen grab, pasted into ai chat, with this question: can you provide a summary for each of these: RESULTS ARE: (too long to list here - see attached file)

[loud moth]

nice, didnt realize how many of these may actually be appropriate outside the stock default

[faint shore]

Many of the SDXL models are tuned to a specific sampler. It's worth reading their model card to determine which one to choose.

[grim bison]

Hi..i want to generate Generate a realistic and functional furnished interior layout based strictly on the provided floor plan image. which way should i proceed. i tried https://api.stability.ai/v2beta/stable-image/control/structure but didn't got the desired image output.

[loud moth]

happy accident trying to render a unicorn coming out of a 'crystal like lake' somehow intepreted the unicorn body and horn as crystal. so i may roll with this lol

[terse ice]

Disha