🏞｜general-with-images

[viral frost]

1792x1008 without (left) and with (right) DyPE. Generating at 4K now.

[viral frost]

Hmm. Well, it's sort of an improvement in composition, but the artifacts make it unusable. (without/with)

[soft hedge]

i've been having fun with this cyberpunk lora, im using it to make npcs for my campaign

[noble sequoia]

so beautiful

[viral frost]

Why do you keep reposting other people’s images?

[still rose]

This lowk tuff 🙂‍↕️

[trail cipher]

hey guys, does anyone know the best model and settings to use for creating city skylines

[viral frost]

I still don't have a functional upscaling workflow for Flux.2, but I've been experimenting with samplers & schedulers & prompting. My nightmare prompts node (vibecoded port from Invoke) told me a story and I thought Flux.2 did a good job of interpreting it:

I feel ya pal. In fact, the same exact emotion just sauntered my way at about the same time in the gym. I was gonin to lose my grizzs to a buttdark mutha-wey-hoa but I saw something funny in their gogurt yoga moves so I joined them! It may take me me and my friend Steve way too many days to get the buttdark mutha-wey-hoer weezer to understand what you're doin', kid but the kid in the back is gettin' friendly with you! (hearts you up) hahahahahaha haha! gettin your gogurt in the best quality so you can show me off after the gym party! (grabs gogurt from the back of the locker) ha, ha, haha, ha-ha, haha! a meatsauce party, you can have my belt! he snarls. and huffs, back to the sweaty gym

[jagged obsidian]

Is this more tuff 😈

[still rose]

The face structure looks a bit more masculine type than I remember her

But aye if that’s what ur into then it’s tuff asf

[viral frost]

Experimenting with a node for generating JSON structured prompts, paired with my random prompt generator. I was trying to see how successful it would be at pushing Flux.2 toward photographic output, which it seems was a moderate success.

One of the features of Flux.2 that BFL promoted was its ability to expand prompts using its own TE. Has anyone figured out how to do this in ComfyUI?

[jagged obsidian]

Yeah its different style, tone,

[quiet current]

Let s keep it clean from gore (and nsfw) stuff thanks. Others are ok but this one not so much.

[viral frost]

Sorry!

[still rose]

Goat

[jagged obsidian]

I was actually experimenting 3 different styles

Gritty realism/Dynamic realistic lighting/anime-semi realistic

[still rose]

Ik u have a “homework” folder somewhere i just can’t prove it yet 💀

[jagged obsidian]

BRO 💀

[still rose]

Real question is am I wrong

[still rose]

my attempt

[viral frost]

So, I went back to do some more experimenting with the tiled upscaling node I had been using with Flux.1 that I thought wasn't working correctly with Flux.2. It turns out that it actually does, but there must be some sort of memory overhead (not noticeable for me with Flux.1) that makes it use a lot more memory than the tile size would have you believe. With 512x512 tiles (0.25 MP), it is using about the same memory as the first pass does at 1 MP. It's quite a bit slower than Flux.1 and I'm still dialing the settings to get good added detail, but this is a preview of what I'm seeing.

[sterile kiln]

what if you pass it on Z-image upscale with low denoise?

[viral frost]

Then I'd need to have both models in memory!

One key aspect is that you CANNOT use the Flux2Scheduler node on the upscale, as for some reason it was built with no strength setting. I'm still trying to find the settings with the Basic Scheduler node that will produce the best result. Above is with the Simple scheduler with no shift at 0.2 strength.

[sterile kiln]

Simply make two groups.
Another group that does img2img with low denoise, with Z-image...

[viral frost]

I don't know what you mean by two groups. I have a Mac, so if the models overflow my memory, they get paged to disk, which is bad.

[sterile kiln]

You can simple do groups, enable/disable one of them to use it, empty cache, start a new gen

[viral frost]

Also, I don't have the patience to generate a bunch of small images, quit and restart Comfy, and then upscale them with a different model. I need something that works in one shot when I'm not there.

[sterile kiln]

patience pay 🙂
(i can't paste 4k version)

[viral frost]

You can transcode to WEBP or JPEG to upload larger images.

I had to transcode my previous image to JPEG at 99% quality, as WEBP was still too large.

[sterile kiln]

I know, don't worry, it's more of a problem of laziness.

how many gigs of unified memory you have?

@viral frost

[viral frost]

64 GB. I won't be able to upgrade for some years yet, but it'll be interesting to see what the M5 Max can do with 128 GB. Hopefully the neural accelerators will greatly speed up image generation.

[sterile kiln]

64 is not so bad

[viral frost]

It was great for Flux.1. Flux.2 is just too heavy to be comfortable at 64GB, though.

[sterile kiln]

yeah, flux 2 is too heavy.
But you can still play with Zimage, SDXL based like Illustrious/pony/others, Flux1, Qwen, others..

you're using quantified version of Flux2?

[viral frost]

Q8 GGUF. The fp8 version doesn't work on Mac.

I'm conflicted on Z-Image. On the one hand, it is light and fast and smarter than SDXL. On the other, it breaks in weird ways that make it hard to use for quality images. Like my sci-fi landscape images will look partly photographic and partly CG (with weird angular rocks and stuff). It just doesn't have the coherent detail that Flux.1 has.

[sterile kiln]

35 GB xD
Its super heavy man

[viral frost]

Due to the Mac's limitations on the number formats it supports, I don't think going to Q4 would actually help.

[sterile kiln]

Try on img2img with low denoise

images i posted earlier are made like this

[viral frost]

What, I2I at the same resolution?

[sterile kiln]

of course not x)

try the WF on this image

[viral frost]

I'll take a look this weekend, though I don't really do 1girls.

[sterile kiln]

its 1366x768 SDXL based > Zimage upscale x1.5 -> upscale to 4k

[sterile kiln]

it works for everything, the concept is simple, use a small model with more "inspiration", like SDXL based, Qwen image, and upscale with Zimage of Flux1

[viral frost]

Did you find a way to do tiled upscale with Z-Image?

Since it breaks down over about 1792 px.

[sterile kiln]

there is an upscaler in my WF, not sure it works with tiling, but i can pass higher than this easily on upscale, with 2 last groups bottom

[jagged obsidian]

She looks very gothtic here 👀

[whole apex]

Is Z image turbo highest resolution 1024:1024?

It is somewhat less sharper or lacking clarity compared to sdxl or illustrousxl

[sterile kiln]

no, you can easily 2k resolution

[whole apex]

You mean 2048px or 1536px?

[sterile kiln]

i mean 2560 x 1440.
If you go higher than 1080p (1920x1080), sometime you have small artifacts, not often.

scroll a bit, i put an image upscaled by Zimage

[whole apex]

For some reason I can't set 1080 number

It is either 1072 or 1088

I have tried it it doesn't morph the characters weirdly

It's good but sometimes it cuts off heads or out of frame

[still rose]

Yes I do like the gothic style

[stray jungle]

Its fun to play with low res

[viral frost]

Flux.2 upscaling is working, but I'm not super happy with the level of detail and there is patterned noise/stippling all over the image. I suspect it is due to the scheduling on the upscaling step, but I don't really know what to do about it.

Makes it look like a scan of a magazine print.

[olive anchor]

This looks great.
Are you running this locally or renting GPUs?
I’m curious because I’m testing some batch setups.

[rain gazelle]

I rent because it's just easier for me, but you could run local. (Z-Image)

[olive anchor]

Yeah that makes sense.
What part is the most annoying for you — setup, managing instances, or shutting them down after jobs?

[rain gazelle]

It's a lot of various software to maintain, with dependencies and stuff always changing or going out of date. (And despite being open source, some of it is a bit fishy when you get into the weeds.) I found myself spending too much time maintaining different tools, and only have a 3060.
I can rent a 4090 (or A100, H200) for extremely cheap... far less than it costs to buy a 5090 or whatever. Takes just a few minutes to spin up an instance and try out whatever I want, then just ditch it.
Keeps my PC build from earlier this year clean and crisp. 🙂

[olive anchor]

That makes a lot of sense.

If there was a way to just send a prompt or workflow, run it on a clean, pre-configured environment (SD/Comfy already set up), and get results back without managing instances or dependencies — would that be something you’d actually use?

[rain gazelle]

I feel like that exists on a lot of services already. like leonardo and such. I like the flexibility of managing the instance, being able to pull in specific models or add-ons or whatever. And that it is inexpensive. 🙂

[rare kestrel]

I'm looking for ideas on how to achieve this type of image conversion locally.
The example was converted in nano banana.
It's a satellite image and I want to remove the shadows.

also this type of conversion ould be great

[olive anchor]

That makes sense — we’re not trying to replace Leonardo.
The idea is more like disposable GPU runs where you bring your own Comfy workflow, models, and nodes, run it cheaply, then the environment disappears.
No long-lived instances, no babysitting GPUs.
Would something like that be useful for burst workloads or experiments?

[rain gazelle]

sounds a lot like runpod/vast

[olive anchor]

Totally fair — that’s exactly the comparison I’m trying to understand.

For you personally, what’s the one thing about RunPod/Vast that still annoys you or slows you down?

If nothing annoys you, that’s also useful for me to know.

[rain gazelle]

I use vast. Some of the instances simply do not work, or take a while to spin-up once selected. Others are fine.
Some have bettern internet speed than others, which makes model downloads faster.
That's what I can think of off the top of my head.

[olive anchor]

That’s super useful.
If you could trade slightly higher price for guaranteed working instances + fast startup, would you?

[rain gazelle]

I would not.

[copper matrix]

Again, scamerinos

[hearty violet]

guys, some advice to transform this cad plan into in a humanized floor plan based on a reference?

[trail cipher]

do you guys know any other small models to generate images like this

[noble sequoia]

[viral frost]

Just an update on this. It appears the patterning is being caused by the NN upscaling step before 2nd stage denoising. I'm kind of assuming that the increased VAE precision with Flux.2 is leading to it latching onto the upscaling patterns produced by the NN, rather than treating them as a noise source that gets morphed into extra detail like Flux.1 does. Using a simple Lanczos upscale leads to clean 2nd stage output, but it also has very little added detail. I'm trying to figure out what to do about this.

[still rose]

I have a question that another server couldn’t answer me with so I’ll try here as I don’t think I’ve asked but for upscaling models and stuff like seedvr2 for image and other tools which is better fal ai, wave speed and someone said Higgsfeild due to their tools for portraits and stuff since I mainly generate people and stuff but not sure so hopefully someone can help me here

[sterile kiln]

Hey Omnia. Pas mal, c'est quoi comme model?

[rain gazelle]

... :: runs through google translate :: -- ah... it's z-image

[wooden socket]

#💬｜general-chat a quite cat

[hasty karma]

I'm not happy with the results I got on SDNext with these two. Their facial expressions are not acceptable. I don't feel like I have the freedom that I want.

[sterile kiln]

I confused you with someone else here, who is French. No, you were the developer of PNG-SD.
Personally, I like to use Zimage for upscaling, it's awesome.

[rain gazelle]

Ha yep, that was me. I dropped it after everyone moved away from Auto and its derivitives. There's others out there that do cross-compatibility for several generation tools now.

[sterile kiln]

Yes, I imagine so.
And yes, A1111 ended up in a bad way.

[weak hill]

Using the SDXL model via SwarmUI

[brittle pier]

As I said, there are so many smart people here that I can't help but wonder.

[rain pecan]

/prompt lesbain kissing, passionate atmosphere, sexy, full body,

[rain pecan]

/generate photorealistic trans woman, nightclub background, neon lights, shallow depth of field, high detail skin texture, RAW photo

[viral frost]

So, SeedVR2 upscaling before the second pass does a lot better, adding fine detail for Flux.2 to work with while avoiding repeating patterns. However, there is some kind of memory leak with SeedVR2 (or Torch, or Python, not sure) that is causing high memory use on MPS, and I'm seeing a lot of writes to my SSD. I'm worried that this isn't very sustainable. (Note, I also switched to a different GGUF of Flux.2. The one that was initially available, from "Orabazes", apparently wasn't generated very well. It has Q8 precision on some key tensors that are left at bf16 in other quants, and also has a stray extra tensor that apparently shouldn't be there. Switching led to somewhat different output, but the text in my example image got quite a bit better.)

I am noticing a rather large increase in saturation and contrast in the second stage denoising process. I haven't been able to figure out why it's happening. I'm also still not super happy with the photographic output from Flux, as it often feels very CG or even like a collage, lacking consistent lighting.

This is sure to offend someone. 😆 Flux.2 did a good job on the logo, though.

I'm wondering what's up with the Flux.2 TE, though. It is apparently some kind of customized version of Mistral Small 3, as its size does not line up at all with other quants. The fp8 version that ComfyUI distributes through their HuggingFace profile is about 18GB, but all of the "compatible" GGUFs I've found are about 25 GB at Q8. Moving to a Q4 thus only brings a moderate size decrease to about 14 GB and, at least on MPS, results in the exact same memory use. So, even though it seems silly for the TE to be almost the same size as the image model, there is currently no benefit to reducing it.

[surreal plover]

How do you guys deal with teeth in your gens? Any positive or negative prompts you use? I normally just prompt "grin" and it drops me one of these, which look horrible. I usually just give it a little touch in photoshop to make it completely white like cartoon teeth but I feel like I'm missing something.

[heavy python]

I just noticed the cops in background on looking 🤣🤣

[noble sequoia]

[viral frost]

Merry Christmas, ya filthy animal!

[dawn seal]

Looks like some of the mountain roads here, in the Spring, when they reopen them.

[placid flint]

/me message:Generate image

"The rectangle box-like shape represents the vending machine itself, symbolizing structure and reliability. The letter 'R' stands for 'Revolutionary,' highlighting our mission to revolutionize vending machines. The curved arrow integrated into the right leg of the 'R' signifies progress and taking the system to the next level, emphasizing innovation and forward motion to features beyond vending. The color scheme uses unicorn silver (#E8E8E8) for the letter 'R' inside the logo and navy blue for the vending rectangular box.

[heavy python]

Not bad Sir David type

The hand gestures spot on 😂

[austere widget]

Prompt?

[short spruce]

https://marketplace.mintable.com/profile/detrabajo2

[clever oar]

Essentially made up of human bone fragments, each cell stores individual fragments of a biography. Style by Thomas Herrmann.

[crude walrus]

Latina súper girl

[sharp monolith]

Fait faire une prise de catch a ce doudou

[nova olive]

https://youtu.be/edvPrDCWwOk?si=_UrT9p4rkf3vowMJ

[noble sequoia]

[past nova]

Clean, natural portrait test.

[viral frost]

I wrote the prompt but wasn't super happy with the way the images were turning out.

So I made the prompt stupider and got better results!

[somber socket]

Someone posted this image long ago and i saved it. Do do any of you know who it was?

[heavy python]

Airbrush this on a van rn

[solar tide]

The 2nd image looks like what 5 seconds after the first image would be like

[viral frost]

I didn’t like how Flux.2 insisted on having the meteor floating there in front of the action, no motion blur or lighting to tie it into the scene, and that the building has developed a hole in front of the meteor rather than behind it.

[viral frost]

I tried writing my own prompts to reproduce a couple of the images here, but in photographic style: https://www.reddit.com/r/StableDiffusion/comments/1q7a36e/tensorart/. Flux.2 got the content, but I find getting the scale correct is always a challenge. There are only so many ways you can say "really big", and it's up to the model to interpret how big you mean. I'll try to introduce length measurements into the prompt and see if it can do anything with that, as well as inserting "aerial" so that it doesn't keep putting the view at ground level.

[viral frost]

On a whim, I decided to test Flux.2's knowledge and found that it seems to know minerals pretty well, including some obscure ones.

Tanzanite:

Larimar:

Diopside:

Azurite:

Not done with the hi-res image yet, but this is kunzite:

It gets the color and texture pretty well, but not necessarily the crystal habit.

[frank fog]

一只橘猫，可爱，细节清晰，毛发纹理清晰。像人一样站起来。使用手机拍摄的写实风格。

[viral frost]

Some more minerals. Flux.2 gets tourmaline pretty much perfect.

Fluorite. Color is plausible, but the model doesn't know the characteristic cubic crystals.

Vanadinite is just wrong. It should be red-orange-brown with flat hexagonal crystals.

Heliodor

Selenite (accidentally misspelled)

Shattuckite looks too much like azurite, but I don't fault the model too much. I had never heard of it before two days ago.

Sugilite is pretty close.

Celestite

[clever oar]

👋

sup

[dark halo]

is there anyone looking for devs here?
i am an senior full stack AI developer and have rich experience in LLM/SaaS projects

i can build Machine Learnig system, RAG system, AI agents, automation workflows, image and video generation tools, API integrations and custom AI tools using OpenAI, LangChain, Python, JS and so on.

please feel free to reach out to me if you are looking for a developer now. Thanks

[fleet goblet]

self promo

[jagged badger]

not what she asked for....

[viral frost]

Would you rather fight a chicken-sized dragon or a dragon-sized chicken?

[jagged badger]

chicken sized dragon is basically a gecko with a lighter

the other one is a t-rex

[restive pike]

Anyone into ultra-realism? I'm looking for someone obsessed with this topic to work together. Anyone interested?

[tiny hollow]

You can dm

[copper matrix]

Why not in here so everyone can learn huh

Sounds to me your trying to isolate people for advertising or scam purposes

[tiny hollow]

[noble sequoia]

[solar tide]

anyone got a good method to remove the white outline on the guys body? i used a background remover workflow but theres always a tiny bit remaining

[pure heath]

How did you do this?

[pallid ruin]

Nice

[clever oar]

thx hello long time dont see you

[pallid ruin]

Yeah I haven't used discord in months, good to see you

[still rose]

Anyone used new Flux 2 9 or 4b yet?

[tawdry badge]

hi any idea how to generate this style

[viral frost]

Attempt at using Flux.2 Klein 9B as a refiner for Flux.2 Dev. It's a lot faster, but I don't like the noise levels in the output. The subject also lacks some of the translucent quality and a lot of the fine detail from Flux.2 Dev as refiner, but gains some solidity to the form. I'll have to see if changing the scheduler helps at all.

[cobalt root]

https://civitai.com/images/118131941 for better quality

[wispy garnet]

Hh

[weak hill]

SwarmUI and "qwen_image_2512_fp8_e4m3fn" model

[digital moss]

Stable Foundation

[tawdry vector]

Hello everyone, I'm trying to find a consistent workflow of turning a simple 3d screenshot of a building into a realistic looking image using automatic1111. If anyone is willing to share, I’d really appreciate it. Thanks!

[tawdry vector]

Nice @nocturne oak ... this is what I'm looking for. Can you please let me know how can I try to do it myself.

[tawdry vector]

I struggle with sdxl controlnet.

[noble sequoia]

[tawdry vector]

@hearty violet I saw you had a great success with an interior scene. Can you please help me out.

[heavy python]

Official tourism ad for Hawaii 😂

[junior kestrel]

I just build an img2img oil paint style workflow, is that any thing that I can improve?

[hearty violet]

sure what you need?

[tawdry vector]

Trying to figure out how to turn a simple 3d screenshot of a building/interior into a nice looking viz

This is my progress so far

[devout plover]

i want this ingame

[hearty violet]

what is this problem? I'm trynt to generate a text2img with sd1.5

[silk sparrow]

Hi all, in this video the lip sync is in time with the music, i was wondering if they animated it in something like Kling, then used some other tool to add the lip sync? Any ideas?

https://youtu.be/ApmuSbAQ41M?si=PStWlCVUhGfl9AH3

[vague jewel]

what is your most photo-realistic portrait image (human face)?

[viral frost]

Tried to recreate a ZIB image from Reddit (using Flux.2 Dev). https://www.reddit.com/r/StableDiffusion/comments/1qp0lb5/z_image_base_is_great_at_abstract_stuff_too/

[tender kraken]

https://www.youtube.com/shorts/xewY6Bf15XQ

[noble sequoia]

[viral frost]

Interesting discovery: Flux.2 Dev is guidance-distilled, but still seems to work just fine with actual CFG. However, I don't yet have good evidence that style negatives make much difference in the output. This is the original image, CFG 1, Flux guidance 4.0.

CFG 1.5, 3.0, 5.0, 7.0 with style negative to try to push it towards photographic output, Flux guidance 4.0:

CFG 7.0, Flux guidance 1.0:

[cold shoal]

/imaginehttps://cdn.discordapp.com/attachments/1004159122335354970/1467336401459744800/Screenshot_20260201-021426.png?ex=69800303&is=697eb183&hm=0d87651a810f69d73f559d7664f78a352d8d58aaf78c52ea69e0e6468618bca6&
enhance photo, try to preserve original face as much as possible, photorealistic, natural lighting, slightly zoomed out

[quiet current]

not how that works ....

[finite crypt]

Render this hand-drawn image into a physical product, utilizing authentic high-end materials, elegant surface textures, as well as professional product lighting and cinematic depth of field

[quiet current]

Still not how that works

[viral frost]

Just an update about using real CFG with Flux.2 Dev. Here are some example generations (random prompts) with CFG 7.0 and Flux Guidance 1.0. I thought maybe it might be best to reduce Flux Guidance when adding CFG, but it sometimes seems to cause some objectionable patterning in the images (see the first image), as well as making the output "too crunchy".

These are with CFG 7.0 and Flux Guidance 4.

In general, I'm liking the output with CFG, though it takes longer.

[rich vessel]

#Ultra-realistic Indian woman standing in a traditional South Indian house doorway, wearing a red sleeveless blouse and white cotton saree with golden border, jasmine flower garland (mallipoo) draped on her shoulders, small red bindi on forehead, natural makeup, soft smile, hands raised behind her head, slim waist, natural body proportions, warm skin tone, cinematic natural lighting, shallow depth of field, highly detailed skin texture, realistic fabric folds, cultural South Indian aesthetic.

Background: a softly blurred bedroom interior with a man sleeping on a bed, natural indoor daylight, wooden door frame, authentic Indian home setting.

Camera: full-body portrait, eye-level angle, 50mm lens look, f/1.8, ultra-sharp focus on subject.
Quality: 8K, HDR, photorealistic, RAW photo, no distortion, no extra limbs, perfect anatomy.
Aspect ratio: 9:16

Generate an image

Ultra-realistic Indian woman standing in a traditional South Indian house doorway, wearing a red sleeveless blouse and white cotton saree with golden border, jasmine flower garland (mallipoo) draped on her shoulders, small red bindi on forehead, natural makeup, soft smile, hands raised behind her head, slim waist, natural body proportions, warm skin tone, cinematic natural lighting, shallow depth of field, highly detailed skin texture, realistic fabric folds, cultural South Indian aesthetic.

Background: a softly blurred bedroom interior with a man sleeping on a bed, natural indoor daylight, wooden door frame, authentic Indian home setting.

Camera: full-body portrait, eye-level angle, 50mm lens look, f/1.8, ultra-sharp focus on subject.
Quality: 8K, HDR, photorealistic, RAW photo, no distortion, no extra limbs, perfect anatomy.
Aspect ratio: 9:16

[copper matrix]

Scam

[lavish vault]

/imagije I need a realistic but a game-like look layout of aluminium die casting foundry.

[quiet current]

#artisan-faq

[verbal aspen]

Is stable diffusion able to generate images like this without restrictions?

[copper matrix]

Yes. You need proper prompts however

[verbal aspen]

Can you give me the official website for stable diffusion because there is so many on google and idk which one is the best one for generating those kind of images without having to worry about it being flagged

[quiet current]

1/ no there's restrictions whenever you generate stuff locally
2/ rule 2 #✍🏼｜rules-and-tos

[verbal aspen]

I dont intend on posting nsfw content on this server just needed some advice

But I need the official site for stable diffusion can anyone give me the link

[quiet current]

results on stability.ai 's web app (dreamstudio, https://dreamstudio.stability.ai/) should be censored for any political / nsfw / gore and whatnot so it would not be of any help.

I'm not gonna discuss it further.

[verbal aspen]

So which site can I use then

[copper matrix]

civitAI has some options to generate online but cloud based is not great as most people use it locally

[dawn seal]

Calvin & Hobbs!

[viral frost]

CFG does seem to increase contrast in a subtle way with Flux.2. That seems to work well with random prompts, as all of their low-probability and conflicting tokens tend to produce gray, cloudy results in Flux.2. But more defined, cohesive ideas may end up burning in a bit with CFG. This is the prompt (but with fixed shirt colors) with CFG 1.

[viral frost]

Never seen an anatomical model of a walnut before.

[sleek flame]

https://youtu.be/ATY0no0pbAo?si=CDbOfVdZyYPN9xRM

[copper matrix]

Would probably be very helpful to use the proper channel and include what ui you are using

But thats just my two cents

[night crescent]

my bad

[flat reef]

guys, help pls, what can I do about the fact that stable doesn't follow my prompts?

[copper matrix]

thats easy, better prompts

like " full view" is not a tag

using a anime model also limits your options

cat, licking, hand, 1girl,

[flat reef]

it doesn't work to me 🙁

[copper matrix]

try using it with controlnet, but thats a spesific prompt

and wayy too generic at the same time

its 4 words

[flat reef]

it's need a image reference, yep? and what if i want blue jeans and gray t-shirt, its will "blue, jeans, gray, t-shirt". will sd understand this?

[copper matrix]

with proper prompts, sure

1girl, charcoal tech pants with zippered pockets, standing, portrait, reference sheet, (also generic) but it says more of what i want it to have

or using a base model of like zturbo or sd3 allows for more natural prompts " a gnome sitting on a porch of a tiny mushroom house. its sitting joyfully reading a newspaper. titled "The Gnome Times" inspired by "the newyork times" "

[flat reef]

how will you describe my promt?

it's maybe hard to me

with low acknowladge english language

[copper matrix]

Too short & simple

Too short = random

[lethal dew]

👋 Hi there!

I train flux /sdxl lora for Onlyfans and patreon. If u need ur AI Influecer I'll be happy to help u with it.

🔗 Portfolio & custom LoRA with stable face:

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

[copper matrix]

Rule 5 spam

[simple mantle]

Can someone please please please help

[sterile kiln]

Screenshot of a CLI 🤮

[heavy python]

Youre missing requirements to build the wheel for openai CLIP?

Read that resources documentation. Youll need other dependencies likely.

pip install * where * is what youre missing. It could be a few in succession, one after the other.

[quiet current]

yeaaaah CLIP messed up their repo. they deleted some stuff they shouldn't have
1/ open a cmd in webui's directory
2/ run venv\Scripts\python.exe -m pip install https://github.com/openai/CLIP/archive/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip --prefer-binary --no-build-isolation
3/ close it and run webui-user.bat , it should work after that

[dim lark]

If not, what worked for me was skipping clip installation (editing webui.bat) and installing it manually with cmd in the stable diffusion folder. It should work 100% this way

[undone meadow]

Already up to date.
venv "Z:\automatic1111\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
Installing clip
Traceback (most recent call last):
File "Z:\automatic1111\stable-diffusion-webui\launch.py", line 48, in <module>
main()
File "Z:\automatic1111\stable-diffusion-webui\launch.py", line 39, in main
prepare_environment()
File "Z:\automatic1111\stable-diffusion-webui\modules\launch_utils.py", line 394, in prepare_environment
run_pip(f"install {clip_package}", "clip")
File "Z:\automatic1111\stable-diffusion-webui\modules\launch_utils.py", line 144, in run_pip
return run(f'"{python}" -m pip {command} --prefer-binary{index_url_line}', desc=f"Installing {desc}", errdesc=f"Couldn't install {desc}", live=live)
File "Z:\automatic1111\stable-diffusion-webui\modules\launch_utils.py", line 116, in run
raise RuntimeError("\n".join(error_bits))
RuntimeError: Couldn't install clip.
Command: "Z:\automatic1111\stable-diffusion-webui\venv\Scripts\python.exe" -m pip install https://github.com/openai/CLIP/archive/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip --prefer-binary
Error code: 1
stdout: Collecting https://github.com/openai/CLIP/archive/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip
Using cached https://github.com/openai/CLIP/archive/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip (4.3 MB)
Installing build dependencies: started
Installing build dependencies: finished with status 'done'
Getting requirements to build wheel: started
Getting requirements to build wheel: finished with status 'error'

stderr: error: subprocess-exited-with-error

Getting requirements to build wheel did not run successfully.
exit code: 1

[17 lines of output]
Traceback (most recent call last):
...

im loooooooosing itttttt

[copper matrix]

A1111 is outdated and i see your on amd, hop into #🤝｜tech-support and look in the pinned messages for a up-to-date guide or ask for help there

[quiet current]

check what I said like 3 messages above yours. Also yes automatic1111 is ANCIENT.

[restive trench]

@manic stone That's perfect for book covers. Subject on the right and back cover on the left.

[zinc spruce]

dm me

[quiet current]

@royal charm showing you were to look for admins/mods/etc
( cant send picture in #💬｜general-chat )

[lean socket]

good call, people always get lost trying to find mods

[quiet current]

pic with more steps

the more important the role, the higher its position in the list. (but it will only show connected people unless you scroll at the very bottom)

[royal charm]

Can I send you?

[quiet current]

some bench of the GTX 1650

@late elm

[late elm]

[honest frost]

create image of meditation

[quiet current]

not how that works pal.

[low juniper]

anyone want to help me? ill pay

[fleet goblet]

always find it funny when people respond to the bots

[feral sand]

Hi guys! I've been testing consistency with my virtual AI influencer, Riya. Let me know what you think! ✨ If you want to see her exclusive high-res collections, check out my official page here:https://dfans.co/riyaexclusive

[smoky vigil]

ban this sorry ass.

[heavy python]

🔨

[pliant jay]

I have some question on Stability.ai
I need help for api reference

[twin egret]

hello

[split gust]

Some hilarious (badly prompted) images from my earliest SD (late 2023) generations.

[hearty violet]

Hi guys, I made some studies but with Freepik, I think interesting so I will show here
for all these works I used LLM, I started use it now and is very powerfull
FLOOR PLAN:
keep the consistency very well. Some fine ajustes need to be made with krita

RENDER
keep the consistency very weel, some fine adjusted need to be maded with krita. Was hard to put the exaclty texture or ask to put the exact material on the right place, but LLM helps a lot

RENDER WITH A PHOTO REFERENCE
Made teh render looks like a photo! Looks awsome I need more control to change and I need to know how do it without photo, only by a 3d model, I belive that LLM is the secret
photo + 3d model + render

[gentle adder]

Hi everyone 👋
I’m a bit lost and would love some guidance.

I’m trying to reproduce a very specific cartoon-style graphic style using Stable Diffusion. I have a dataset of ~35 images.
So far I’ve tried:

• LoRA on Replicate → very disappointing
• LoRA on Flux2 (Fal.ai) → best so far, but characters are pretty bad
• PRUNA AI p-image with LoRA→ not great
• QWEN2512 on Fal.ai with LoRA → not great either

I’ve heard about DreamBooth, but I’m not sure if in 2026 it’s still the gold standard for this kind of task.

I’ll attach a reference image of the style.
Given this dataset and goal, what would you recommend today to properly learn and reproduce a precise cartoon style?

For context: this style was easy to generate back in 2023–2024 (even with ChatGPT), but it feels like it has completely disappeared from current outputs.

Thanks a lot 🙏

[shrewd falcon]

Haneda Tobari is coming

[tranquil condor]

#🏞｜general-with-images 精修这张钻石戒指图·做成商用珠宝摄影白底图，有淡淡的阴影，钻石切面明显，金属精修。

[fleet kiln]

#🏞｜general-with-images Create a wooden logo for Fuxiangji

[viral frost]

So I've been gone for a while generating because I found this LoRA for Flux.2 Klein: https://civitai.com/models/2384168?modelVersionId=2681004. Thinking the generations looked cool, but not wanting to use Klein or a LoRA, I decided to test Flux.2 Dev on the prompts (just prefixing with "Photograph of" to try to avoid illustration style). I think the LoRA results are still better in terms of style, but Dev seems to be more prompt-accurate. I tried first with CFG, but definitely saw it increasing the contrast unnecessarily, so I tried again without.

Somewhat cherry-picked:

There are artifacts here from my memory-limited upscaling procedure. If I had more unified memory and could do larger tiles, this should reduce.

Flux by default doesn't get the scale right in a lot of these images. The things are "big", but in terms of hundred of feet, not miles.

[heavy python]

You been on a roll

[viral frost]

Just want to see what Flux is capable of, and still feel like I'm scratching the surface. The hidden ability to use CFG opens up possibilities. I had been thinking that the regular output lacked the vibrance of other models. CFG fixes this, but seems to go too far. I'm trying to see if it's just my negative prompt now. I added some terms to try to crush the overexposed, hyper-contrasty look, but not sure how successful it is.

[fleet goblet]

imagine if it still had the wolf's noise, and that the tentacles were furry.

i bet it'd be a pretty cool concept animal

[heavy python]

Those are pretty cool

[upbeat canyon]

Hi fellas! Whaddya think of my Rainbow Dash movie poster?

@heavy python do you like it?

[restive trench]

From a grammatical point of view, your comma should be a semi-colon. Semi-colons in head lines are pretty ugly, so I'd go for the period. Make it two sentences.

[upbeat canyon]

@heavy python it's rainbow dash!

@restive trench do you like my poster?

[restive trench]

It has all the qualities of a good poster. Strong foreground character, side characters in background. The far background contains elements establishing location. The typography looks good, but I don't know that brand.

[upbeat canyon]

I made a theme song for the movie too. It's called Rainbow Dash the Pegasus!

@restive trench pretty awesome ain't it?

I even made a revamped version of the dinosaur crocodile hybrid monster named Dinocroc.

@restive trench cool monster picture ain't it?

I like it don't you?

Here's Kittybusters!

@heavy python hi deadbeef!

[upbeat canyon]

Is anyone there?

[upbeat canyon]

Hello? Anybody there?

[upbeat canyon]

@restive trench you like my ai pictures?

@viral frost hi fellas. It's me Matthew!

@clever oar

It's Gizmo and Pinkie Pie as news casters.

@wispy nest hi buddy.

Here's my movie poster and it looks great don't it?

[upbeat canyon]

pretty cool pictures aren't they?

[upbeat canyon]

I made them on Gemini 3 using Nano Banana 2.

[upbeat canyon]

Whaddya think about my pictures?

[upbeat canyon]

Whaddya you think of my Dinocroc picture?

[upbeat canyon]

Hello?

[noble sequoia]

hello

hi

[upbeat canyon]

Hello!

[viral frost]

😂

[upbeat canyon]

@viral frost I have my pictures here.

[viral frost]

Lol at Ben Franklin with Trump hair

[half hatch]

Web search crocodylomorpha, you might like the designs 😉

I dream of the Forest of Dreams

I dream of a land of pastel and acrylic

Also, "reflective salt flats" works quite nicely with Dreamshaper8.

[upbeat canyon]

It's dinocroc. I made this picture based on my drawing.

[silent rapids]

/upload

[atomic swift]

Hey everyone 👋

I built an AI tool called Ziraxo that generates 3D AI models from text prompts.

You just describe the model and the AI creates it automatically.

I’d really love to get feedback from the community 🙌

https://ziraxo.com

[heavy python]

Adreitz check out the cartoon Primal. If you havent seen it. It's by the creator of Samurai Jack. Think you may love the creative visual scenery & settings. Might give you more ideas.

3rd seasons still airing now i think

[summer yarrow]

yo anyone knows how can i fix the eyes?

[copper matrix]

adetailer / segment refinement depending on the UI will do the trick (or upscale, and inpaint the eyes only)

[noble sequoia]

👍

[rapid onyx]

Back in my day we made our own artwork

and we didn’t need no fancy computer to do it

I made this with grok i think

[smoky vigil]

https://www.youtube.com/watch?v=n233GPgOHJg

Hey guys, let's reminisce! Maybe one day they come out with a kickass SD4.5!

[dawn seal]

Apparently, I just have a dirty mind.

[shrewd falcon]

thunder raiko

[frail nebula]

Ltx does a great trump

Also pretty sure it can do joe pesci because one gen I made 'teeny tiny' pesci leaked in

[dawn seal]

That is HYSTERICAL!

[frail nebula]

Home alone parody

[smoky vigil]

https://www.youtube.com/watch?v=6Xa3t8spSYk

SD 4.5 when?

[maiden swift]

#American President Trump dancing in the Oval Office playing with a large inflated globe-shaped balloon in the style of Charlie Chaplin in the scene from the film The Great Dictator

[dawn seal]

What IS it about people and pictures of feet?

[rugged ice]

Ever felt like reality was watching you back?
My psychological AI thriller THE SIGNAL explores what happens when the world starts reacting to the way you observe it.
After a mysterious explosion outside Marrakech, Lara begins noticing impossible patterns — moments repeating, objects appearing twice, strangers who seem to know her before they meet.
The more she pays attention… the more reality begins to change.
🎬 My entry for the Higgsfield $500K Action Contest.
Would love your thoughts 👀
https://higgsfield.ai/contests/make-your-action-scene/submissions/7b55a320-863a-43b2-9e1c-ed10493958aa

[mighty wren]

Imagine I create an Anime movie with this quality

[blissful sage]

Hi guys... I'm working on an image management app called PixlStash (https://pixlstash.dev) which is meant to help deal with the sheer amount of images we create. It has ComfyUI integration so you can run workflows directly in the app and a basic plugin system, plus it tags images for you, including trying to spot typical AI malformations. I'd really welcome some feedback.

Good question.. currently I think I've been looking at around 20 images/s tagging with GPU inference. It has a task manager so you can see the processing tasks being done.. this is doing WD14 + running a convnext-base finetune for finding malformations and tagging images with things like "bad anatomy", "extra limb", "extra digit", "malformed teeth" etc. This is mostly 2MP images though

It will change a bit depending on the VRAM budget... but I know people are doing other things on their GPU so obviously you need to balance it with those other things

Takes a few minutes for 11000 images but I haven’t tested it with 70k yet

Tagging should scale linearly but it is the image likeness calculations I’d be interested in seeing with that many. I spent a bit of effort into reducing it from a n^2 thing to something that scales

[light raven]

Upscale

[copper matrix]

Scam

[mighty wren]

if you like my artwork support me on Sora:
https://sora.chatgpt.com/g/gen_01kkqr2kgcf91btvbkg3d8e6da

[heavy python]

This is my new religion

[heavy python]

Appreciate you 🫂

[ripe cedar]

What're the best models for flat 2d cartoons? The more detail the better, but as long as its not a style that gradients the colors

[fleet token]

Giving off that gta VC vibes

[wispy nest]

lol @manic stone i was testing loras

[copper matrix]

What model was used here?

[wispy nest]

z-image

[copper matrix]

Very nice

[sterile kiln]

FLux2-klein is awesome

[sharp fox]

Gere a imagem

[snow trench]

a blue moon

[viral frost]

Well, now. Ain't that something!

[upbeat canyon]

What did you think of my picture?

I even made the theme song.

[viral frost]

Have some beans instead.

[quiet current]

Not the place for that kind of topics.

[quiet current]

Just no, like really. NOT THE PLACE FOR THAT. Last warning before bonks.

[wispy nest]

@quiet current is that allowed?

[quiet current]

timed out and I m cleaning this whole mess

[copper matrix]

@wispy nest just dont engage

[coarse gust]

Hey everyone 🙂 I’ve been working a lot with Stable Diffusion setups recently, especially around LoRA training and ComfyUI workflows.

If you’re struggling with things like:
– inconsistent characters
– low quality results
– broken full body generations
– or workflows that just don’t give you control

That’s usually not just prompt related, it’s more about how everything is set up together.

I mainly help with:
• training clean, consistent LoRAs (face + flexible body, SFW &
• setting up smooth ComfyUI workflows for images and video
• improving realism and consistency without making things rigid

If you’re working on a project or trying to get better results always happy to point you in the right direction

[hollow cypress]

Animat this

[rain gazelle]

As requested. Animated.

[dawn seal]

It's OVER 9000!

[sterile kiln]

https://tenor.com/view/vegeta-scouter-dragon-ball-z-gif-7640583

Its an LoRA i just made.
I had already created the Toriyama-style LoRAs for the Frieza and Cell eras (for Flux1dev). Now I've just done one for Flux2-Klein. It looks really good.

[mighty wren]

he should have been only one timed out because he was the first one who started arguing racistly but I was just defending my religion artistically and I didn't say others should follow it but respect it and I don't accept to timeout for people who cannot have some trust in their religion so they attack others religions @quiet current

don't want to send a screenshot of what he said because you already know he said a racist expression.

[copper matrix]

I think its best to let the topic die since that guy isnt active anymore

[mighty wren]

just wanted to explain it because my art is not offensive towards any culture at all unless it is twisted

[marsh yarrow]

cute pixel art office worker character, 16x16, isometric view, simple and clean, pastel colors, transparent background --pixel

[blissful sage]

I don't know if AI-toolkit is the tool most people use for training, but I quite like the UI so I vibe-coded a way to browse datasets from my PixlStash image server... If it was to be included in AI-toolkit I'd need to refine it a bit but the idea was to make a browser for any kind of image server with plugins and supply a PixlStash plugin with the PR. I'd also be interested in integrating with other apps and will probably make some integration nodes for ComfyUI even though PixlStash already supports monitoring the output folder of ComfyUI and to run some workflows within the app itself.

[carmine garden]

This is interesting!! especially the part about identity consistency..

Have you found that dataset quality matters more than quantity for LoRA training, or do you usually aim for a higher number of images?

[pure mason]

Looking for an expert in realistic AI character creation (same face across multiple images). Need very natural, imperfect, real-life phone selfie style (not model, not studio). 20-50 images. Paid work. Please show examples of same character.

Looking for high quality work, budget flexible.

[manic stone]

Good idea to collect images for loras 🙂

[pure mason]

Send private

[quiet current]

chill down with those pictures please, #✍🏼｜rules-and-tos rule 2 no nsfw stuff

[slender cobalt]

Stop battling AI for consistency! 🛑 Mastered flawless Character Consistency with my new L-OSC Framework! 🧠🚀 #🏞｜general-with-images

[agile grove]

cinematic wide-angle shot of a skier descending a snowy mountain slope at sunset, dramatic mountainscape in the background, golden hour lighting, soft fog and snow particles in the air, dynamic motion, snow spraying from skis, subject slightly off-center, large negative space on one side for text, ultra-realistic, high contrast, depth of field, crisp details, cold blue and warm orange color contrast, premium travel photography style, 16:9 aspect ratio

[untold gate]

an interesting human

[viral frost]

What is even going on right now?

[heavy python]

You should do album covers for artists/musicians

[queen crystal]

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?

[flint cargo]

Hi, is there any chance someone might recognize the art style/model/checkpoint used for these images? I’ve been trying to find it, but I’m not sure. I tried looking at the metadata via SwarmUI and Forge Neo but they both came up blank.

[tight falcon]

Hey everyone,

I work as a ComfyUI specialist and LoRA trainer, helping people create high quality, consistent image and video content.

My focus is on solving common issues like inconsistent characters, unstable outputs, and inefficient workflows. I build structured pipelines and train custom models to make content more reliable, scalable, and visually appealing.

If you’re building an AI influencer or looking to improve your results, I can help you get there faster and more efficiently.

If you’d like to see samples of my work or learn more, feel free to DM me.

[subtle meadow]

Hi, today I vibecoded new funny nodes. 1 for "darkroom" solarization. I feel more realistic than Photoshop solarize effect. The best, if somebody already developed black-and-white photos (IRL) using short white flash diring process, the node simulate soft (B-type) and hard (H type) photopapers (if the strenght is 0 no solarize, just paper simulation). I feel its close to analog reality. The another node is Lightroom style clarity. Check the full images too.

[viral frost]

Just wanted to see how easy it would be to reproduce the thumbnail image from Dan Dingle's recent video with Flux.2.

[copper matrix]

@lunar aspen
anyways, the auto complete feature i was talking about

[lunar aspen]

ohhhh

its that thing

[copper matrix]

with images with +400 words i dont really need a lora IMO

(varies)

[lunar aspen]

also i want to generate some things..

will i need LoRa?

[copper matrix]

well a lora is only needed if its a weird concept. if its a NSFW question. shoot me a message request so i can awnser it but i prefer general questions or talk to happen here lol

[lunar aspen]

k

[last crater]

Yeah that’s true. For most cases, LoRAs aren’t really necessary unless you’re trying to recreate a very specific niche character or style. I also like using Swarm’s autocomplete to check if the base model already understands the character before adding a LoRA

[copper matrix]

Bot ?

[last crater]

Nope

[dense moss]

Me when bot

[grand basin]

i was trying to use openpose this error coming help

[quiet current]

most likely you're mix and matching different type of models

[grand basin]

Just using sd openpose model

[quiet current]

which ones

and with which model

[sterile kiln]

And what about very old friends ?

[cyan axle]

I'm trying to generate an image of a falling leaf via SD3.5 through ComfyUI. I've tried a ton of things to get it to work, but it always generates the leaf facing the camera, which just feels unnatural.
Do any of you have any suggestions for prompting that may break the thing away from all the macro-photography training it has of leaf faces, to just get a falling leaf that's actually seen from it's side?

Ran out of ideas, so got Claude to help write the prompt, but here's the most recent attempt, and the result.
Positive prompt:
Award-winning nature macro photography. A single dry autumn leaf is caught mid-fall between two trees in a forest. The photographer has captured the leaf at the exact moment it is tumbling, rotating on its horizontal axis, so the camera sees only the leaf's thin edge and the curve of its spine — the leaf appears as a narrow crescent or thin arc, not a flat oval. Golden backlit forest, god rays, bokeh trees. No ground contact. Photorealistic. National Geographic style
Negative:
flat leaf, leaf face, full leaf visible, oval leaf, leaf surface texture, leaf on ground, leaf on branch, twig, stem, sculpture, object, art installation

[sterile kiln]

Just stop use this deep shit of SD 3.5, simply.

https://huggingface.co/unsloth/FLUX.2-klein-9B-GGUF

[versed wasp]

generate image

[sterile kiln]

?
Would you like me to generate its prompt using Flux2-Klein?
Easy.
@cyan axle

[pale hawk]

“Wide cinematic shot, room mostly dark, focus on window light silhouettes of woman and young man close together, husband in foreground shadow, slightly out of focus, unmoving, atmosphere heavy with implication and psychological intensity

add cinematic text overlay:

‘Kabir: This is the line.’
‘Rhea: It was crossed a long time ago.’
‘Arvind (calm, resolved): I know.’

same characters, same faces, consistent look across all scenes, cinematic lighting, ultra realistic, 8k, professional photography, shallow depth of field, film grain, anamorphic lens, moody color grading, high contrast shadows”

[turbid root]

A high-quality, glossy Apple-style emoji image on a pure black background. The image features a yellow face with an embarrassed and awkward expression: an open MacBook near the left temple, a red exclamation mark near the right temple. The mouth is grinning, showing the awkwardness of work problems. Two full yellow hands (emoji hands) are spread out on both sides of the cheeks, indicating helplessness. Clear soft shadows, vibrant colors, 3D texture. --ar 1:1

[little vortex]

How can creat ai images here

Can anybody tell me that

[copper matrix]

#artisan-faq or not at all, most people here run it locally using their own computer

[unborn light]

I have no clue!

[nova olive]

Hello i know i do not say much here . But text to video is getting good . What would you say is a good one . but not Groc?

[unborn light]

Finally we know what the far side of the Moon looks like! Wow! They are not dead!

They also have a new way of talking!

"Planet Earth is blue - And there's nothing I can do" (David Bowie)

[unborn light]

Happy Easter Bunny! 🥚

[sterile kiln]

https://tenor.com/view/what-huh-gif-5238629

[restive trench]

To quote the crazy guy on 30 Rock. With God as my witness, there will be casinos on the moon!

[visual mango]

Hey all, anybody following the DLSS 5 situation? When I heard how hard people were coming down on the tech I felt like it was a ridiculous idea that it could only make things photorealistic. I did some experiments with Comfy Cloud / SD3.5 / ControlNet / LoRAs to look at what such a technology might be able to do if directed differently. I wrote an article, but just scroll through if you only want to see the images!
https://aitalesfromthefield.substack.com/p/nvidias-dlss-5-controversy

[wispy nest]

My ai video generator 4k 60 fps

#sorakiller

[viral frost]

Flux.2 Dev understood the assignment. Jerboa constrictor

[unborn light]

The Jerboa swallowed the snake?

[viral frost]

Inspired by the Mare Internum webcomic: https://www.marecomic.com/

[viral frost]

I saw this photo and just had to create my own version with the logical conclusion. https://www.reddit.com/r/StableDiffusion/comments/1sh4hpx/automate_text_replacement_in_images/

[gray axle]

Working on photo-realism.

[dawn seal]

How Vegans see meat. 😂

[twilit cargo]

design a smart watch

https://tenor.com/tTK7.gif

[undone onyx]

Black and white children's textbook cartoon. Two panels.
Panel 1: Boy with short black hair waves at girl with shoulder-length hair near school gate. Speech bubble: "Hello!"
Panel 2: They face each other smiling. Speech bubbles: "Hi!", "I'm Liu Tao.", "I'm Su Hai."
Halftone dot shading on clothes, round speech bubbles with tails, bold outlines, white background.

[maiden garnet]

Monster photo

[heavy python]

https://thehackernews.com/2026/04/over-1000-exposed-comfyui-instances.html?m=1

🫠😬

[distant cedar]

Animação futurista

[feral current]

Avez-vous besoin d'aide ?

[formal rampart]

@copper matrix

this tooo may be bad

[copper matrix]

Looks decent but what are you making them for?

[formal rampart]

fun

or may be professional later

[copper matrix]

Ah well it's decent. Most people here run it locally so there's more to tweak and edit

[formal rampart]

can i see ur designs

[copper matrix]

Sure let me hop on my pc

[copper matrix]

@toxic sonnet pure delusion

[copper matrix]

some of the more recent ones

[copper matrix]

[manic stone]

[formal rampart]

@copper matrix hi

@copper matrix hi look

hi

anybody there

[viral frost]

Calm down. This place isn't very active right now. If someone feels like reacting, they will, but nobody owes you a reaction.

[copper matrix]

@brave karma because i got this working with klein

woops he added a beard, didnt notice

oh well minimal details

[brave karma]

@copper matrix this is pretty close to what I'm trying to do

I have a pretty good idea on how this is done since I have experimented and done close examples

[copper matrix]

ahh feels like witcher style 3d models

could use a lora probably

[brave karma]

As far as I know, they take a picture of their character's face and give them a reference photo to copy the pose, environment and outfit

I've done similar stuff with Gemini but that's kinda it

[brave karma]

thats a super good way to explain it haha

[copper matrix]

hmm well depending on how bad your need is a lora is probably the way to go for something that spesific

[brave karma]

and i find that in civit?

oh and I have near to no experience with making workflows myself so that's why I was asking for a public workflow

[copper matrix]

ah a lora doesnt need a spesific workflow

either you train one yourself or be lucky and theres one availible

[brave karma]

oh shoot i dont know a thing about training a lora, i'll start searching for some tutorials

I appreciate it

[copper matrix]

hmm on civitai its not too difficult really

but locally? its gonna take some time (a lot of time using your gpu)

[brave karma]

oh you can train a lora on civitai?

or you mean finding one?

[brave karma]

oh but its a pro feature?

[copper matrix]

yes? but you can also train locally. but takes some effort

using their image labeling/tagging feature is free though

but for a style youd want 40-80 reference images ideally

[brave karma]

hmm i just might be able to get that

so let's say I did train it. then is it just putting down a module and making your picture go through it?

[copper matrix]

you just add a lora (node) to your workflow

kinda suprising that you went with comfyUI as your first UI

[brave karma]

I couldn't setup AUTOMATIC111, it gave out some errors so I just went with comfy

I like it though

[copper matrix]

ah A1111 is horribly outdated thats why

comfy will support most models though so thats a plus

ease of use however is not

[brave karma]

Eh yeah I'm getting used to it

But I'll probably get civit pro, the other way seems like a big hassle

[copper matrix]

well good luck making the lora. i think on fiverr you could also comission someone as it would be a lot easier if you dont wanna do it yourself

but never ever get someone in your dm's trying to sell you something or a service

always a scam

[brave karma]

got it, thank you

[austere rapids]

Hey guys, quick question — do you also get a lot of bad outputs like extra fingers or weird hands in anime images?

I’ve been working on a small tool that automatically detects these issues and filters out bad images so you don’t have to check manually.

If anyone wants, I can test it on your images and show results 🙂

[coarse torrent]

Hey folks — been deep in building mode lately and figured I'd share a couple of things I stumbled on this month that are actually useful:

GLM-5.1 just went live on BytePlus's ModelArk Coding Plan. Capability is aligned with the original full-strength model, and since it's running on BytePlus infra it's been stable and basically ready to go out of the box. If you're shopping around for a coding model (or just curious how it stacks up), here: :
https://www.byteplus.com/en/activity/codingplan?utm_source=External_Media_Agencies_Paid_Developer&utm_medium=External_Media_Agencies&utm_campaign=BP-Global-Codingplan&ArkClaw-publish-Q2APRFY26&utm_term=tegongyuzhou&utm_content=codingplan

Also — for anyone playing with video gen — Dreamina Seedance 2.0 is on BytePlus now too. BytePlus is the official API platform for Seedance models, so if that's your lane, worth a look: :https://www.byteplus.com/en/activity/seedance2-0?utm_source=External_Media_Agencies_Paid_Developer&utm_medium=External_Media_Agencies&utm_campaign=BP-Global-Codingplan&ArkClaw-publish-Q2APRFY26&utm_term=tegongyuzhou&utm_content=Seedance

Might poke at both this week — if anyone tries them, curious what you think 👀

[clever oar]

😄

🍗 eat healthy food bro

no scum

[abstract egret]

making diamond ring
engagement ring

How to generate with AI

#🏞｜general-with-images diamond ring,yellow hold,wedding

[heavy python]

💍

[viral frost]

Based off a dream I had, but I just couldn't figure out how to get Flux to look across the highway rather than down it.

[clever oar]

animal drone?🙂

[solar tide]

can anyone with an ltx 2.3 custom audio workflow animate this with custom audio? all my custom audios are poopoo now and idk what happened. it was working fine just a week ago then something changed. the custom audio can be anything can be any length

[honest storm]

：/imagine prompt: a cyberpunk cat with neon lights, wearing sunglasses, 8k, hyper-detailed --ar 16:9，

[manic stone]

[viral frost]

Didn't exactly work out how I imagined, but still funny.

[viral frost]

Aww, that spammy guy from yesterday got banned. And I made him the Christmas card he requested and everything.

[pearl citrus]

@solar tide

[solar tide]

@pearl citrus yo cool tysm man! did you use a custom audio for this?

[pearl citrus]

Thanks man, I used a new feature I'm adding to https://www.missinglink.build will send you the ui link once I finish it

[solar tide]

@pearl citrus not comfyui?

[pearl citrus]

no way man

way more control in pure python

also the dependencies and triton kernels I run are custom not sure how easy it would be to configure that in comfy-ui

[solar tide]

alright man well im only looking for a comfyui workflow that works cause thats what im used too. thanks for animating it tho !

[copper matrix]

Is it open source or cloud only?

[pearl citrus]

the dependencies you can download yeah

not entirely open source though

[copper matrix]

So its closed source beyond dependencies?

[pearl citrus]

yeah

[copper matrix]

Why?

[pearl citrus]

gots to get paid to build em

it gets expensive to build some of the combos and optimize them, flash attention for example is a nightmare to compile

[copper matrix]

So its a service your selling?

[pearl citrus]

yeah trying to start a business around it

not easy

[copper matrix]

I can imagine since comfyui already exists and people still gotta use their own hardware

But please do not advertise as its rule 5
I feel you on the struggle but its not up to me

[pearl citrus]

my bad, will tone it down

[copper matrix]

No worries, you seemed like an actual person so i didnt do a thanos snap i usually do with advertising bots/spams

[pearl citrus]

yeah please don't

[solar tide]

@vagrant moon

[manic stone]

If you have too many ai images and being bored...

[rocky otter]

What's the best one?

[copper matrix]

considering the outfit i think no 3? i dont know the character though

[rocky otter]

search

Five Hearts under one roof Gran

[copper matrix]

oo i sppose its accurate

a risque prompt, though lets not do nsfw images in here

[rocky otter]

sowwy

she had clothes tho

anyways

why is it like this

[copper matrix]

your prompt was very suggestive, prompt for clothes

[rocky otter]

okay

Why is it only working sometimes

[copper matrix]

your trigger word has a lot of similarities to like grandma, grandpa, grandfather, grandmother like i said earlier

[rocky otter]

do you think

i should

re train

it only took me an hour

[copper matrix]

you should do what you want to do 🤷‍♂️ if it works enough for you its fine no

[rocky otter]

would this be a good prompt

fiveheartsunderoneroof_gran

[copper matrix]

why not Gr@n

[rocky otter]

would that work?

[copper matrix]

ive seen it before, why not

[rocky otter]

okiii

thank youuu :))

btw

is wd14 better than blip

or is it acording to the model

[copper matrix]

i would not be able to awnser that, i use civitai's tagging service

[rocky otter]

okii

would Gran Dong work?

or Gran_Dong

[copper matrix]

second one might be better

[rocky otter]

okii

[copper matrix]

but have you read any guides on this at all

[rocky otter]

Do you recommend Regularisation images?

[copper matrix]

i useually dont.

[rocky otter]

oki

[rocky otter]

prompt
gran_dong lora:gran_dong-000001:1

prompt: gran_dong, 1girl, looking at viewer, lora:gran_dong-000001:1

tf

prompt: gran_dong, gran dong, 1girl, lora:Gran:0.7

why is it like this

@copper matrix

[copper matrix]

I don't make realism loras but you get the girl from the game you wanted no? Whats wrong

[rocky otter]

its not her

[copper matrix]

Ai shenanigans

Idk, probably not a good enough dataset, lora not trained enough or something else. Theres a million things that could be it

[rocky otter]

okay 😭

[copper matrix]

You really should ask training questions on the koyha/onetrainer discord as im doing 2d stuff only

3d allows for less leeway

[rocky otter]

i couldnt find

the server

[copper matrix]

For onetrainer?

[rocky otter]

both

[copper matrix]

On the GitHub page of onetrainer.

[rocky otter]

okii ❤️

[copper matrix]

Hope what you find what you need

[thorn tinsel]

@copper matrix this is illustrious plus wan

[copper matrix]

Yeah you could get the same image quality in anima easyly, just more writing/ prompt thu a llm

[sturdy gorge]

hi 😄

[copper matrix]

neat but rule5, thanks!@blissful sage

[blissful sage]

Ok.. Thought it said it was ok to share AI related tools but is it the community projects channel then?

[sturdy gorge]

hi am new and greet all 😄

here my welcome gift 😄 https://www.youtube.com/watch?v=i1TaIDEdWFo

[copper matrix]

would be better :-)

[pure frigate]

trying to get better took me while to get reintroduced to the tools

[sturdy gorge]

looks like a classic 8bf filter 😄

[sturdy gorge]

my newest one https://www.youtube.com/shorts/QLfPKXxiKO0

[viral frost]

Made some tweaks to my upscaling workflow. We'll see how well it works in the long term.

[halcyon magnet]

正向提示词 (Positive Prompt):
A breathtaking fantasy digital painting, a young girl standing in the center of an enchanted bioluminescent forest, giant glowing mushrooms, translucent ethereal jellyfish floating in the air, sparkling fireflies like stardust, magical atmosphere, intricate details, highly detailed, masterpieces, cinematic lighting, soft volumeric light, dreamlike, Unreal Engine 5 render, artstation trending, by Rossdraws and Alphonse Mucha.

反向提示词 (Negative Prompt):
(worst quality, low quality:1.4), deformed, bad anatomy, blurry, disfigured, extra limbs, mutated hands, ugly, text, watermark.

[copper matrix]

No nsfw @wise forge

[viral frost]

From my dream last night.

[copper matrix]

@ember trench no nsfw

[ember trench]

Okay

[mental geyser]

kayak

[clever oar]

👋

[flat pawn]

@copper matrix

[copper matrix]

Hey, What's up

[hardy vigil]

I'm trying to generate high-quality robot three-view image.

But it's very hard to keep perfect consistency.

[copper matrix]

looks pretty consistent! sadly perfection isnt possible yet but it looks nice.

[hardy vigil]

Thank you!

[dim grotto]

Hey I've created an automated ai workflow using a variety of tools and custom scripts. I use N8N to run a local ollama that runs of a set of randomized prompt guidelines that pipes into comfyui to generate random abstract art in large batches @ 10240x5760 but they aren't very crisp, I keep what I find visually appealing, then I refine 10240x5760 using tiling for my 6 monitor wallpapers using comfyui - then i convert those images into video - 961 frames first frame to last frame @ 960x540 - then upscale the video to 2k using rtx node in comfyui - then to 4k using a custom script and ffmpeg all locally on my 9800x3d with a 5080 16gb I started this project a couple months ago and had no idea about Linux - comfyui just basic computer knowledge and now I think I am creating some pretty stunning stuff with my end goal to broadcast it from my nas to a Samsung frame tv - the videos are 4k @ 60fps but ill create a script to play them slower and in forwards/reverse since I get a bit of a jitter when it loops on some of the videos so it will look like my wall art is alive.

Attached is a link to my amazon photos and github I am not trying to promote myself in any way I am a landscaper that is new and learning the hobby and want to share my tools and learn from the community.

Photos:
https://www.amazon.ca/photos/share/PjtiqJ3nQCJvhg7CDYcuBHogAJvuEG69rAGixKI3MDJ

Video Samples:
https://www.amazon.ca/photos/share/ReZIqZVf0Ln4vG8BLxZwJ2klxnt5s9Gy5WB5LtF2mmR

Project:
https://github.com/zantraxzantrax/AI_Image_Video_Tools/blob/main/README.md

You have to download the full .png about 75mb each image to get the full detail - right clicking and save as just downloads a low resolution web optimized version.

Looking for a way to host 100gb+ of my videos for free if anybody has a solution let me know.

Also need help with how to make my first frame to last frame video transition more smooth im getting stuttering on the last frame.

[sick ice]

Edit this

[quiet current]

not how it works... Also why would you edit some pic from some girl's social media

[fleet goblet]

Why do you even respond to the people who try to do that, its been done in this discord for years lmao

[quiet current]

No doubt people do it with grok other online tool or even on their own machine. (And personally I choose not to help people with reactor and other tool that like that).
But trying to edit pictures of existing people without their consent in broad daylight is not something to condone.