#šŸ’¬ļ½œgeneral-chat

1 messages Ā· Page 189 of 1

warm junco
#

Thats an Illustrious thing

#

Try out nova anime xl its really good

gloomy imp
#

Dang. And I did, can't mimic the blend of artists i like. Anilustrious and Nova3d can get there, but neither is for XL

rocky meteor
#

can i run comfyui and wan video in a 4GB nvidea graphics card?

tidal pebble
#

Hey people
Can anyone here help me with runpod, trying to create a Lora. Walking into too many problems. Especially when I do all the premilary things, press start training and nothing happens. Endless clicking

tidal pebble
pale wedge
tidal pebble
pale wedge
tidal pebble
pale wedge
# tidal pebble What do you use? If you do?

uhhh A111. pretty user friendly. i mean i also try to make sure everything is configured correctly, I usually begin by double-checking my dataset and configuration files. then, i monitor the logs throughout training in order to identify any errors early. also execute a small test batch aids in identifying problems prior to full training

rocky meteor
#

guys, when iam generating anime, there are some floating stuff, like a view of an internal organ

#

how to remove those?

steel prawn
#

Stop prompting for it? I dunno without seeing your prompt I can't tell you what's causing it.

atomic mortar
#

^, if you tried to use natural language in a anime model or use tags like "xray" "viscera" or other well grotesque tags it could be a cause

sour solstice
gloomy imp
hallow sequoia
#

Hello everyone is there a guide to jump on making LORA? I'm bored of seeing only the women from anime. Where are big muscles

sour solstice
#

Although it is true that it is less coherent than others

gloomy imp
#

Even adding in a style lora and it was bad

sour solstice
gloomy imp
sour solstice
gloomy imp
#

Itā€™s just a far more efficient tool imo. Setting up adetailer and control net on comfy was a pain

#

Also forgeui generates lightning fast

sour solstice
#

Speed was the same on my tests, but it is fair about efficiency. It is not really convenient for manual work

#

Invoke is cool for manual generative editing

gloomy imp
#

I think Iā€™ll probably just start training my own Lora for what I want. The few models I stick with for illustrious are so good that itā€™s just a matter of getting the odds and ends I want ported

sour solstice
gloomy imp
#

So far as wishlisting, I kinda wish that there was an equivalent to NovelAIā€™s vibe check, or their multi character prompt system

#

The former is such a reliable way to generate the exact same style and character parameters

sour solstice
#

maybe regional prompting and optionally quick sketching can help?

#

can't wait for flux kontext copium

gloomy imp
#

Maybe. For now Iā€™m just learning and getting my preferred stuff figured out. Iā€™ll be upgrading to a 5080 soon so thatā€™ll help run bigger stuff

sour solstice
gloomy imp
#

Iā€™m using a 4060. Not bad, runs more efficiently than I had thought

sour solstice
#

12gb is what saves me

gloomy imp
#

I need a better card though for more than just ai. High poly sculpting with blender

sour solstice
#

based

#

but isn't it on cpu and ram?

gloomy imp
sour solstice
#

got it

valid aurora
#

Hey guys, apparently fkunn1326/openpose-editor has been dysfunctional for some time now. The tab won't show even if you install it, and I've seen people with similar problems. I've also tried downloading the sd-webui-openpose-editor in the Extensions > Available tab, but I don't know if it didn't work or I can't find it. So, are there alternatives for SD 1.5 AUTOMATIC1111?

I also tried to open Latent Couple, but I'm having the same problem.

gloomy imp
#

Anyone got a link to a decent tutorial on LoRa training?

forest delta
#

Hey guys, I would like to morph one face into another, generating e.g. five intermediate images. Is there a model that you know of, that could perform this task? Many thanks, chillbird

valid aurora
#

@tawny drum search not4talent on youtube hes the best ! helped a lot although im too untalented to make a decent lora ive made a barely passable one with his help. its for sd 1.5 automatic 1111 tho

tawny drum
#

šŸ¤”

atomic mortar
#

I think he meant to ping missalys but accidentally pinged you somehow?

dawn mulch
#

Is anyone using AI to write custom nodes I just tested how easy it was and I was able to get a node for rotating image for box selecting an area of an image and another loading multiple images and making them work in various ways etc its basically as easy as 'one too three'. It takes just 1 prompt to create costume nodes now days and in 1 to 3 minutes. So anyone using AI to create costume nodes?

#

Anyone have ideas for custom notes?

keen niche
wary tangle
wary tangle
#

um is this a scam?

#

i guess it was xD

valid aurora
# wary tangle is a1111 still a good interface to use? i just now got it working with zluda xD

im too noob to give you an answer myself, so I'll copy and paste a discussion from another group chat "Don't use A111, it hasn't been updated in nearly a year. Since extension authors are focusing on support for forks that supersede it, a lot of them won't work with main anymore.
I suggest giving forge or forge classic a try. (edited)
NEW

@Equinox Psychosis
Don't use A111, it hasn't been updated in nearly a year. Since extension authors are focusing on support for forks that supersede it, a lot of them won't work with main anymore. I suggest giving forge or forge classic a try. (edited)

Kylada ā€” 5/31/2025 1:45 PM
but i am so used to a1111...
NEW

@Kylada
but i am so used to a1111...

Equinox Psychosis ā€” 5/31/2025 2:59 PM
In that case I'd recommend forge classic. It stays more true to the original a1111 webui while keeping things optimized and up to date. It does remove some of the outdated stuff tho, like the old textual inversion training and the old model merge menu(You should be using sd-scripts and super merger instead) (edited)

1
NEW
June 1, 2025

@Kylada
but i am so used to a1111..."

wary tangle
#

i think ill stick with it until i learn how to use sd

#

and then will see about forge

valid aurora
#

thats my plan too

wary tangle
#

it already feels a bit overwhelming by itself, i dont need more stuff to worry about while im learning it

#

and using an amd card is basically a gamble getting anything ai to work on it

#

i swear, anytime ive tried to do anything ai on amd, ive ended up spending a weekend looking for workarounds, not just stable diffusion, but i was playing with openai whisper at one point and it too was a pain in my ass

valid aurora
#

ohh coincidentally i also used whisper for one of my college classes xD

wary tangle
#

i used it for creating captions for a japanese concert

valid aurora
#

good. i used the clour version but it was too slow, so i ran it locally no problem

#

it seems they discontinued the v2 onwards large models so i used v1 i found a workaround eventually but there was some weird stuff at the start

wary tangle
#

hm, it was pretty slow for me local, but not too bad, took like an hour for a 3 hour video

valid aurora
#

i recorded my lectures then summarized in chatgpt lolol

wary tangle
wary tangle
valid aurora
#

korean šŸ˜›

#

korean to eng

wary tangle
valid aurora
#

oh ya it had weird shit but as long as i got the general concept it was fine

wary tangle
#

true

#

mine had music in it so that didnt help, but also, i think long pauses really screws it up too

#

it would say some crazy stuff

#

and some times it would just say the same word like 20 times in a row

#

lol, man, i spend so much time learning how to get this stuff to work, use it one time, and now i dont even remeber how i ever got it to work

#

i even installed ubuntu just for whisper

#

i guess i learned how venvs work tho

#

its funny tho, i couldnt get whisper to work in windows, and couldnt get a1111 to work in linux

wet grotto
#

Hi guys, I'm finishing to record my sd course for achitecture in portuguese -Brazil
I'm recording this for 1 year, not all time but was when I started. It's 10 hrs of practical video class and 3 hours about historie, isntallation and interface.
The course is 100% focused on architecture with A11 and SD1.5
https://www.youtube.com/watch?v=bSaYeWK83wk&t=2s

#

I want to thank the server and the whole community, many of you have helped me by answering questions, giving attention, thank you very much.
I would especially like to thank @oblique elk , @warm junco and @fervent thunder . I would like to ask you to mention your name in the acknowledgments of the official course material, may I?

warm junco
wary tangle
steel prawn
#

and reported

grim aspen
#

Has anyone experience with doing motion capture with auto1111/forge/comfyui and using that data in blender? I've seen videos where they make the opposite like use a rig to create openpose images. But I'm looking for smth I can run locally that allows me to capture motion data from images/videos and use those in blender.

hot ember
#

hiii

pale wedge
silent gazelle
#

um hello guys I'm a computer science student who is into DeepLearning, I have took machine learning and neural networks courses and I want to expand my knowledge in that domain, I know few deep learning topics, didn't really wrap my head around RNNs yet how does they get trained, also same with the attention, so does anyone knows is this server a good place to do deeplearning projects ?? I know that stable diffusion is a pre trained generative model

grim aspen
deep narwhal
#

any 5060ti guys in the chat

#

is it worth getting over the 5070 (12gb) for the extra vram

fervent thunder
#

ngl comapred to veo 2 and others Stable diffusion picture generation hurts my eyes

#

its a failed abomination

#

doomed to die

pine path
#

the XL finetunes are still pretty good for stuff that isn't complex

sweet pivot
#

hello

jaunty anvil
#

Hi, everybody

#

I'm an AI developer with specialized in generative AI and llm, langchain and RAG

static mortar
#

Hi guys

crisp gyro
#

Hi folks. I'm trying to figure out how to generate video using a starting frame on ComfyUI, like how I'm able to do with Klang/VEO/etc. I tried WAN Image to Video, but it's not adhering to the reference image. Any help?

gloomy imp
#

Are there any decent methods to make backgrounds and characters look appropriate in size compared to eachother? Like making a bed look realistic and not like a toy next to them, or have them not be the same height as a door?

shrewd lichen
#

yo

pale wedge
stark scroll
grim aspen
pale wedge
pale wedge
ancient mauve
#

@abstract quarry do you know any Ui for multimodal models? lm studio cant handle safetensors it seems

oblique elk
fierce crag
#

hey guys im like day 1 of stable diffusion and had some questions

#

can someone help me?

atomic mortar
fierce crag
#

i got the help now

atomic mortar
tribal scaffold
gloomy helm
#

I've been using FOOOCUS UI for a long time now. I really wanna try using Flux but I'm not exactly sure what's the best way to use it locally. Do you guys have any tips?

still glacier
#

what about no ?

#

if someone joined that discord, leave it. It's all scammers trying to install ransomware / get your wallet

#

@gloomy helm

gloomy helm
tame vector
#

can someone recommend a lora for adjusting clothes that is not Pony?

tribal crown
#

hello guys, seem to be getting problems with Wan lately. I am running on 4090 but when I run it, it is getting stuck at ksampler part, disconnects and tries and reconnect but will not finish my prompt. I am also getting this error, but not all the time. Any help? Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.

normal ridge
#

Was hoping to find some activity in here for some help too

#

My SD runs but the art is horrible

#

Iā€™m trying to use dreamshaper but it looks like itā€™s not even switching to it

still glacier
normal ridge
#

Bruh

still glacier
#

there is no "external support discord" or whatever

normal ridge
#

I shouldā€™ve know agony

#

Yeah luckily I didnā€™t give them any info

still glacier
#

it's just a bunch of scammers trying to get you to install ransomware and get your wallet

#

because "it looks bad" is not very descriptive

#

what software are you using, which settings, give some output examples, etc

normal ridge
#

Gotcha Iā€™ll collect some now

still glacier
#

gotta go to work for now, I might help if I have some time.

tribal crown
#

fixed it

#

fixed it, but now am getting this error

#

Given normalized_shape=[1280], expected input with shape [*, 1280], but got input of size[1, 257, 1664]

lean timber
#

Hey guys, I was wondering if anyone knows if there are any models that have better prompt understanding than FLUX. I really like how FLUX understands the prompts so well, but I don't like the appearance of it that much.

pale wedge
lean timber
#

I've tried Juggernaut and RealVisXL, but they also use the older tag-cloud style right? FLUX seems better at understanding a long descriptive text rather than a list of tags.

pale wedge
lean timber
#

I can't find this savro anywhere

#

Oh Savro.ai. But it has no local model or?

pale wedge
abstract quarry
#

there are many loras and checkpoints for flux

#

if you are more into artistic stuff, PixelWave is a great flux checkpoint for arts

#

then there are a lot of loras for photographic styles in flux

livid gale
#

does anyone know what the hell is DPM++ SDE CFG++

#

this sampler even exists?

dawn mulch
#

It's not exactly simple but it's the most straightforward way to fix such issues effectively.

muted kelp
#

Hello..

real path
#

hey is here a lora trainier for realistic ?:D

naive cipher
#

Scam

#

No one click that

#

No one is going to send you to a different discord server to give you help

tame python
#

Question so I want to make ai art again i previously used forge im told that's no longer being supported? So should I use comfyai automatic1111 or swarm?

still glacier
#

(sorry, was watching their messages history didn't realise this message was days old)

dawn mulch
#

@still glacier They already scammed me but they only took $55.5 billion, from me so it wasn't too much..

dawn mulch
real path
#

Yeaaa but first of all you have to train a good lora for the beginning of comfy ui but I have big big trouble with KoHYA GUI

dawn mulch
#

kohyass is terrible last time i tried to use it.

oblique agate
#

something I noticed with most gen ai model that works with wan 2.1 is they associated long hair as blonde (when you do image2vid) so the vid always turns out as a blonde long hair girl

pale wedge
livid gale
#

guys
my flux is consistently making one image twice as long as previous one

#

1st image takes like 60-70 sec, second like 140 sec

#

consistently

#

why

bold trout
#

is a model required if you are just creating ai images from existing images you have on your local folder?

#

and how does one accurately tell the ai image to image to use the picture included and just add a simple thing, like a hat to a bald person in the picture

#

while still keeping the picture the same

valid aurora
#

@bold trout img2img > inpaint > only masked area. im using automatic 1111 sd1.5 tho. if its an anime style, i can try and do it for you

bold trout
valid aurora
#

im noob and i havent tried realistic but if you sned it in the general-with-images section i can give it a go and show you my process

bold trout
#

are existing images from the web allowed?

valid aurora
#

erm im not sure tbh

#

i think they are

bold trout
#

I'll share an example from the web in that gen with images

bold trout
#

Also can you add prompts in complete multiple sentences like for ex: make a stickman wearing a boonie helmet. Also have him carry an ak47

bold trout
#

@warm junco nsfw is not allowed to be talked about here in any form? right?

warm junco
bold trout
#

Oh ok then. I might have to take it to the reddit then since it sems that nsfw is (somewhat) allowed there regardless of the rules? I think they allow discussion but no pics

warm junco
dawn mulch
# pale wedge heyy, when you say building your own nodes, like does this apply in the same way...

I've seen people create all sorts of nodes and workflows for all sorts of things beyond just image generation. Co mfyui can handle more than just stable diffusion. It can make video, 3d models, audio, etc. you can even talk to LLMs in comyui and have nothing to do with image generation. Being able to create your own nodes, if you are willing, or at least create your own workflows with pre-existing notes, allows you to customize most anything in comfyui. But to answer your question yes it does make it easier to customize 'ai influencers ' or anything that can work in comfyui.

steel prawn
#

just started my comfy journey after learning forge and A1111. Steep initial learning curve, but its well worth it. And i havent even touched video gen or any of the other stuff.

dawn mulch
steel prawn
#

had a buddy show me his flow, and he explained what each portion or group did and why. the connection dependencies im familiar with since ive done some light automation for work with n8n and powerautomate. but theres a bunch of stuff forge does under the hood that you have control of with comfy and i didnt even know where to begin looking for all of it.

dawn mulch
#

If you don't know you can also use LLMs to help you create nodes yourself, if you dont want to learn how to do it, and do all sorts of other things. I havent tried asking it to build a workflow for me, but i suspect, it can actually build workflows. The only other thing is that you do have to have the models that have the ability to do whatever available for the nodes to work with

steel prawn
#

yeah, assuming its a reasoning model or can search the web, im sure it can. i had them teach me how to make wildcards when i was using forge so im sure it could probably teach me to make custom nodes. but that sounds like a lot of work. might look into it if pressing generate starts getting stale though.

cunning walrus
#

Hello

dawn mulch
steel prawn
#

huh, might have to try it out

dawn mulch
#

I want people to realize how easy it is because I want people to start building a bunch of awesome nodes.

steel prawn
#

haha theres already so many to pick from, it would be hard to stand out and have your nodes be seen and used lol

valid oar
#

I also tried Comfy for the first time yesterday, after using A11 for a year and a half. I can see where people come from regarding the "flexibility" for bulding the pipeline, but regretfully for SDXL most of the extensions that do matter for me still only work properly for A11 like Multidiffusion tiled upscaler. Just like Forge's, Comfy's version of the extension remains broken

#

I will still have to get into comfy one way or the other if I want to start practicing with Flux and SD3.5 though

steel prawn
#

hmmm multidiffusion tiled upscaler eh. thats something i never really messed with. im surprised it isnt already something someone has fixed up on comfy though

bold trout
#

so does the interrogate feature describe the image source in question? when doing image to image generation?

steel prawn
#

it will attempt to tell you the prompt, but it wont extract metadata if thats what your asking. though i do believe there is a way to do that if the image has metadata attached to it

#

deepbooru also can interrogate and convert to danbooru tags, but its not super reliable

bold trout
#

so it will tell you the promp of the estimated image or what the prompts could be for the image?

steel prawn
#

it will try to. its not 100% accurate, youll often have better luck with sharing an image with chatgpt and asking it lol but if your dealing with nsfw you can give it a shot. i know comfy has a metadata extraction node that can pull full prompts from images.

bold trout
valid oar
# bold trout so it will tell you the promp of the estimated image or what the prompts could b...

they are designed to be used for Lora trainners and checkpoint refiners in order to automate the very labor intensive metadata filling for each picture on the dataset. as Retsubu mentioned they are not meant to be fully reliable so afterwards the user is expected to clean up the false positives themselves. but the bottom line is, that as someone who is just generating and not developer you shouldn't look at it as a way to find out tags or what to write, it takes sometime but in my opinion the best aproach is to just bookmark the danboru wiki and search things up as you need them, with time you will memorize which tag correlates with which concept

steel prawn
# bold trout got any link for that?

the clip interrogator is built into forge, and deepdanbooru can be downloaded through the extensions tab in the webui itself. it connects directly to github

bold trout
steel prawn
steel prawn
bold trout
bold trout
steel prawn
#

in the comfy ui manager, you can look it up and install it there

bold trout
#

also how does one save prompts in stable diffusion in both positive and negative prompts

steel prawn
#

create a .txt file and save the metadata from your gen to it, or go into your settings and tell it to copy a file each time you generate an image of the metadata, which will contain your prompts and settings/parameters

valid oar
# bold trout also how does one save prompts in stable diffusion in both positive and negative...

if you mean how to save them into the picture's as metadata, that happens automatically upon generation. if you mean how to save them on the workflow that happens automatically as you save your workflow before closing the program. if you mean how to save the prompt as a stand alone text so you can use it later... well, it is is literally just text, you can copy paste it into a .txt and open it later to use as a reference

bold trout
valid oar
# bold trout so for someone new to donboru wiki. how does one use those images to create imag...

my apologies, I should have been clearer, this is something specific to anime oriented models. SDXL is trained on I think 8Billion pictures, those pictures would carry asociations of descriptions of their content, often in the form of short verses. The developer of SDXL fined tuned for anime models realised they could use this idiosincracy to train their fine tuned checkpoints over some hundred thousands anime pictures carrying those tags so their models would be most likely to respond while using those "tags". if you are not planning on using an anime model this is straight up useless for you

bold trout
#

yeah no anime models just realistic ones.

pliant moss
#

Does anyone know of SDXL finetunes which have been trained for native 1536x1536 generation, like some Illustrious models have been?

bold trout
#

so when it comes to prompts, is there a priority on which one gets first? and how does one tell the ai there are 2 characters in the picture and how to designate an action to each one? ex, vietcong soldier and american soldier. american soldier weiling an m16 while vietcong weilds ak47.

torpid tendon
#

No. You just have to hope that the model has been trained enough on images + captions that have this level of detail. The more recent models like whatever is in current chatgpt and things like Flux Kontext are improved over earlier models like Flux and SD3.5. And those (Flux and SD3.5) are improved over earlier models like SDXL. For now, a prompt like your example may or not be correct. Although in general I would expect these models to be bad at differentiating between specific gun types.

stone sinew
bold trout
vast furnace
#

Hi everyone, Iā€™m using a new RTXā€Æ5080 (Blackwell, sm_120) and Iā€™ve tried to run AUTOMATIC1111 WebUI with SD.
I get errors like ā€œno kernel image... sm_120 not supportedā€.
Has anyone managed to get Stable Diffusion running on RTXā€Æ5080? What steps or builds did you use?
Thanks for any tips šŸ™‚

#

Subject: Stable Diffusion compatibility with RTX 5080 (Blackwell architecture)

Hello,

I'm trying to run Stable Diffusion (SDXL/WebUI) on a new NVIDIA GeForce RTX 5080 GPU, but the current version does not yet support this architecture (sm_120 / Blackwell).

PyTorch nightly builds with CUDA 12.8+ still fail or require manual compilation, and several components (like xformers, flash-attn, or torch-sdp) are also not working.

Can you please let us know:

  1. Is support for RTX 5080 (sm_120) planned?
  2. When approximately do you expect compatibility with Stable Diffusion (esp. SDXL)?
  3. Are there any known workarounds or official plans to address this?

Thanks in advance for any updates. Many users with new GPUs are currently blocked from using SD.

Best regards,

oblique elk
pale wedge
valid oar
#

let it run for the first time and test that it is working fine. Xformerfs are functional but don't appear on the list by deffault. so after running it for a first time close it and add this to the webui.user.bat:
set COMMANDLINE_ARGS= --xformers --force-enable-xformers
afterwards launch A1111 again and go to settings > stable diffusion > optimizations and you will see xformers once again in the list, select it, and aply settings. From here on you have a fully functional A1111 for Blackwell

reef plover
#

What is the best image to video tool?

floral umbra
fallow veldt
#

do you know how to make it do a shot where it doesn't move the camera much or not at all?

upper plinth
#

Slow motion video of a woman gently kis the viewer while she slowly tilts her head to the sides with pucked lips and a feminine expression. The soft diffuse light highlights her detailed skin textures and delicate contours of her nose and face in full detail

#

Adapt that prompt to whatever you are trying to create and see if it improves

fallow veldt
upper plinth
#

yeah its a common problem

floral umbra
oblique agate
#

tea cache speed up vid generation by a lot

fallow veldt
floral umbra
# oblique agate tea cache speed up vid generation by a lot

teacache, if you got 30 series or better, sageattention, skip blocks node, speed loras, so much that can boost it by sacrificing accuracy. Or if you're lucky enough to have 50 series from nvidia, nf4 which can speed it up even more.

What i wish existed, was a tesnorRT converter for wan/comfyui that splits all the components into smaller ONNX, then once converted, pack it back up into the final .trt which can speed up generatio for up to 3x.

oblique agate
north night
#

hello

oblique agate
vast furnace
oblique agate
valid oar
vast furnace
#

Yes. that was the first thing I did after I created the folder where the SD will be installed.

valid oar
# vast furnace Yes. that was the first thing I did after I created the folder where the SD will...

did you use GIT? all you have to for blackwell to work is to install the specific version I told you
just go to your folder (eithera new one or delete the contents from the previous install) and run this in CMD/powershell

git clone --filter=blob:none -b dev https://github.com/AUTOMATIC1111/stable-diffusion-webui.git

once you have it running, give me a heads up and I will guide you on how to activate xformers

wet grotto
#

guys what is the difference between SDXL and Pony?

valid oar
# wet grotto guys what is the difference between SDXL and Pony?

SDXL is the base AI model developed by Stability AI, all currently popular models, such as Pony, Illustrious, Noob and Animagine are simply fine-tuned (retrained to a lessser degree) versions of the very same SDXL. Pony in particular forked very early with its idiosyncrasies. so amongst the rest it is the one that feel more different to operate, but it is still SDXL under the hood, like the rest.

next gen illustration/anime models are already looking into DIT for their next projects but progress has turned difficult, allegedly. the next animagine is gonna be based on SD 3.5, Illustrious has already been cooking with aura flow, and who knows what else they modded under the hood which allows them to aim for natural language prompting just line NoveAI, Noob introduced for the first time (I think) a V-prediction tweak into their training. but as of right now they are still SDXL

vast furnace
valid oar
#

cool. is it working properly?

vast furnace
#

I don't know yet, it downloads after launch

valid oar
# vast furnace I don't know yet, it downloads after launch

Yeah, don't worry, it does that. The one thing you will notice is that by default it will not recognize or list xformers (the best memory optimization setting) however, it is absolutely installed and it works just fine, activating it is just a matter of adding a couple lines into a text file, don't worry we will get there when we get there

vast furnace
#

it installed and it works. Thanks

bleak igloo
#

Algum brasileiro para me ajudar? Estou tendo problemas onde meu stable diffusion demora 15 minutos para gerar uma imagem com cfg de 5 steps de 30 e seed fixo sem hire Flix sem Adtailer e ControlNet

spiral totem
#

Help me generate some illustrations of Jiangshan City's intangible cultural heritage projects

wet grotto
valid oar
# wet grotto thanks! Idk about it all! What is DIT and version V-prediition? I will check abo...

The short answer is that DIT is a new architecture used on newer generative AI, very resource intensive, it is for users but even more so for trainers, that's why it has stagnated on gaining adoption on the open source community. I am told it is very difficult to research and train it without the resources of a big company. V-prediction is simply a different math formula being applied for the noise scheduling part of the process, however this is not just some component you can hot-swap for any model, it has to be involved during the training for it to also work a intended during generations

wet grotto
#

@valid oar I'm making a presentation to explain what is SD and controlnet to a portuguese course that I'm recording here in Brazil, using A11 and SD1.5 in this first module. Can I DM and show you to get your opinion?

valid oar
#

Sure, no problem. I don't consider myself an authority in this or anything but I'll be glad to share some thoughts

worn condor
#

Hey will upgrading from 16GB to 32GB RAM make a difference?
I generate with illustrious model in 832x1216 with rx 6600

low tulip
#

If my goal is to produce photos of myself or some other person (have a library of training photos data) is there any advantage of training a full model from a base model or does a Lora basically give full power at fraction of training time/size?

#

Also are there any new techs for achieving this?

#

I need to be able to produce consistent images (body and face)

#

I actually don't need hundred percent adherence to the original person, but consistency is more important

#

Second question, what is the best uncensored open source models nowadays? For photorealistic images

river talon
#

i think realistic vision v6

#

or maybe cyberrealism

#

i never used them locally but as far as i see they are most realistic for me

low tulip
#

Those are based on sd 1.5?

#

What about flux?

steep remnant
upper plinth
#

anyone gotten dreambooth to work with 50 series yet?

simple trail
#

I used Google Labs to generate an image; there is one image it generated I like I'd like to change the rotation of the image but keep everything else ?

oblique elk
simple trail
#

@oblique elk It's a bottle and every thing about the bottle I want to keep but change it's angle and rotation

steel prawn
#

he wants to change the composition but keep the bottle (subject) the same

oblique elk
#

Well if you look at the bottle from a different angle and rotation the subject (bottle) will change as you see another part of it or less depending on the rotation etc. Like this would be the bottle from above "o" while this "I" from the side. So you would need to change the object and the background elements. Totally not worth the work compared to generating a new one in the correct position

simple trail
steel prawn
#

save the seed from the original might do the trick. depends on how strong your prompt was

oblique elk
floral umbra
#

No idea if it'll work, but currently with git's copilot, i'm gonna attempt to make a script with tensorrt's sauce code on git to convert wan/hunyuan checkpoints to .trt for a 2-3x generation speedup

simple trail
#

It's like with these image generators, they can generate what you type but if you get the result you want but need to do some changes. That's when it creates a whole new image and you can't get anything close to the previous image.

floral umbra
#

On my 3090, it can generate such images in 300 millisec or so.

simple trail
#

Lock Seed

#

how do I do that with Google

floral umbra
# simple trail how do I do that with Google

Didn't know google had one. I've blocked off it's annoying LLM popup from my search results if that's what's doing it, so i don't know how that one works.
I'm using comfyui for all local generations with a library of models and loras on my server.

upper plinth
#

Which is the best service for renting GPUs for SD?

pale wedge
# low tulip If my goal is to produce photos of myself or some other person (have a library o...

hello, yeah, LoRA often gives you most of what you'd want from a full model but with way less compute and training time, especially if your focus is on one person and you just want consistency. If you want a consistent body and face outputs, good image curation and tagging matter just as much as model choice. And for your second question, there a bunch. one is RealVisXL and juggernaut. did not work for me though hahahahah tried to use savro (a plug and play) and results have been good so far :))

hearty fox
#

anyone have a link to a decent img2vid (ideally wan 2.1) colab notebook they use?

upper plinth
#

dreambooth is still king

#

but a pain in the ass to get working

#

ideally dreambooth + lora will give the best possible results

supple meadow
#

Hey guys im new in comfy ui. And im having this problem: ā€œ./ComfyUI/output/2025-06-11/ComfyUI_0001.txt' is a write-protected path.
Please add it to the whitelist fileā€. Can anyone help please?

low tulip
#

Is it better to use rundiffusion or spin up a runpod? For Lora training and generating images

#

Also is a MacBook pro m4 suitable for all this stuff or do I need Cuda enabled Nvidia chip?

floral umbra
#

If a lora has "unknown", or "other" as the base model, how do i find out what base model it's for? Would only way be to run it through every base model to see what won't look like colored static/deep frieed image?

oblique elk
pale wedge
floral umbra
#

I'd almost wanna make my own trained image analyzer model based on deep fried images and use non deepfried to sort them that way kek

magic spade
#

Hi i am complete new....what shoukd i install first to get on stable diffusion....i am broke with midjourney paid version....heard stable diffusion in github.....

warm junco
magic spade
#

Hi....i got lenovo legion laptop

magic spade
# warm junco Hey, what's your GPU?

Does gpu matters....no idea but its decent laptop.......one of my friend uaing lower specs than my laptop ...bit hia image are very good....he runa python etc like that....he works from that

#

Its RTX 3070 8gb

brave vigil
#

hey everyone

#

I'm looking for somewhere to ask questions about a diffusion model I'm trying to build

#

I need some human feedback

floral umbra
#

@vapid dove scammer above. As we just filtered that scam in a server I staff in. Who added it as a link filter. the server ID that starts with 1294 and ends with 6045 in the first numbered after attachments/

magic spade
floral umbra
magic spade
#

Did that too

#

I did good choice doing this than getting midjourney right ?

magic spade
#

šŸ”§ RECOMMENDED SETUP (Stable + Tested):

Component Recommended Version

Python 3.10.6 or 3.10.11
PyTorch 2.7.1 + CUDA 11.8
CUDA Toolkit āŒ NOT needed (PyTorch comes prebuilt with CUDA)
Git for Windows āœ… Yes (needed to download WebUI)
GPU āœ… RTX 3070 is perfect

floral umbra
magic spade
#

Oh....how to upscale image caz in it we can get 512x512

keen cove
#

hello

bold trout
#

are you able to create ai videos by default on stable diffusion?

#

I didn't know this was a thing lol

upper plinth
#

sort of.

#

animatediff

#

but it sucks ahh compared to sora

#

sora has a few jailbreaks for those wanting to do some semi-erotic stsuff

livid reef
#

has anyone uses 1.2?

#

and manage to get xformers working?

sturdy cobalt
#

Hello

#

[NEWS] Lockheed Martin Corp. - The Most powerful arms corp in the world invest in First Quantum Currency ever made - QRL (Quantum Resistant Ledger)

floral umbra
floral umbra
#

Lol. No idea what that was. I don't use stock client on my phone, so whatever ammo discord thought was a good idea to add to scammers for instance wouldn't work on me, all I see is raw code omegalul

past knot
#

Iā€™ve been working as a software engineer for over 7 years, mostly focused on web and mobile. Recently, I built a time management app ā€” so I even managed my own time pretty well! I also have experience building apps in travel, news, POS systems, and integrating with weather APIs. I'm currently looking for new opportunities. Thanks!

magic spade
#

Hi

#

This is my issue

#

Can anyone help which version to download and step by step process....the one who did recent ?

#

šŸ§µ Stable Diffusion + Karras Incident Summary (with GPU info)

  1. Installed Python 3.10.6 after issues with newer versions.

  2. Set up Stable Diffusion WebUI (v1.10.1) successfully.

  3. WebUI worked fine with model (dreamshaper_8.safetensors) and default samplers.

  4. Wanted DDIM++ / Karras samplers, so tried adding a custom sampler.py script or extension.

  5. Script needed xformers, so I tried installing xformers==0.0.37, but it doesnā€™t exist for CUDA 11.8.

  6. Tried various .whl files from GitHub and VastAI ā€” none worked.

  7. My GPU is NVIDIA RTX 3060 12GB (CUDA 11.8) ā€” compatible but still faced issues.

  8. Accidentally pasted multi-line shell commands ā€” possibly corrupted environment.

  9. Got error: sampler_names[0] IndexError ā€” UI failed to load due to empty sampler list.

  10. After multiple failed fixes, ended up formatting my PC to start clean.

astral sierra
#

Hey guys, this might appear basic as hell. I just started with sd yesterday and here is my situation. I have real images/backgrounds that i took in real life. I also have created some characters/models in stable diffusion and I would like to place them in the very real backgrounds that I took without altering them. How do I go about doing this? I have tried inpainting, imgtoimg with my background but my character never comes out quite right. (I only have faces for the characters) any help would be greatly appreciated. Thanks,

vast topaz
#

how to use stable diff

#

i have a gpu GTX 1650

warm junco
astral sierra
warm junco
#

In img2img

#

Oh you already tried it

astral sierra
# warm junco You can try Inpainting

Do I inpaint over the background and then generate the character in there? Ive tried that but then how do you inpaint the right area where they would actually be, the face for some reason never matched the face i had in the ip control net. It was always too deformed to be able to tell anyway

warm junco
#

You make a mask where you want to have the character and then prompt for it and play around with the denois.
But I think the whole thing is to much to ask for SD and can be better done in Photoshop maybe

magic spade
#

Installation on forge ia same as auto1111

astral sierra
warm junco
astral sierra
warm junco
#

Np šŸ™‚

magic spade
#

Many here use auto or forge ?

#

Curioua to know...heard forge has limited resource than auto...is that true

vast topaz
warm junco
warm junco
#

Auto1111 has a better extension support but forge nearly supports the same extensions. Also Auto1111 hasn't been updated in months.

tribal scaffold
vapid dove
#

anyone here have much experience with Invoke?

oblique elk
#

@wicked badge is as far as I know an invoke hero.

thorny rose
#

I haven't ran SD for well over a year. Last used on Automatic1111.
Is there an easy to use face swapper that's decently good these days? I have a decent amount of RAM and a 3090.

clever yew
#

Do you have any tips for avoiding extra fingers for characters? No matter how much I specify it in the negative prompt, when my character's hand holds an object, his hand always ends up with 6 fingers instead of 5

wicked badge
#

I'll say one thing about these bots. I admire how wide of a net they are casting. Don't see this much coordination usually. Shame it's such a shit tier obvious scam that they're peddling.

abstract quarry
hard blade
#

anyone know how to get SD working with a 5090?

#

geting errors, something about Cuda

lucid gate
#

whts the prblm

astral sierra
hard blade
abstract quarry
abstract quarry
gloomy imp
#

Ok, got a question. Does anyone know if there is an ai program that can generate normal maps based on a given mesh/UV? Say I have a low poly count mesh that is UVā€™d, is there a program yet that could read that and then generate a normal map?

still glacier
#

get lost

still glacier
#

There is no "external official discord support" or something like that

#

they'll jus try to get in your wallet or install ransomware

tiny spear
#

I'm interested in finding an expert to pay for some consulting. I'm using COmfyUI, ipadapter, controlnet

gloomy imp
young storm
#

which one is better

#

stability APIā€™s or Runway API?

#

Any dev who has experience in both ?

pale wedge
# young storm stability APIā€™s or Runway API?

not anaa expert haha but an avid reader on reddit. people say Stabilityā€™s API gives you more control if you like to tweak things, while Runwayā€™s is easier and smoother to use. It depends on whether you want flexibility or simplicity. Some folks mix both.

#

So, if you're still exploring, savro.ai is also a solid middle ground for this one and playground.ai for experimenting

lofty basin
#

where to download stable diff?

warm junco
#

Scammer

fading shell
#

Hello, sorry to bother you, I am pretty new to stable diffusion, I followed some tuto, but it seems to not like my prompts cause I never get what I want, do you know where I could get some help?

thorny bloom
#

gguf models are slower? i mean in wan, vace, fusion x etc

young storm
ebon agate
#

hello guys

grim tendon
#

Hello server

rancid stirrup
#

anyone able to help me figure out what was used in a certain set of images? I've not found the lora or anything

pale wedge
tropic frost
#

you guys know any good illustrious checkpoints or loras that help enforce an 2d, anime screenshot style?

#

or do you just rely on an artist style that is focused on that particular aesthethic?

gritty lava
#

i luv gemini

versed rain
#

Hey everyone! šŸ‘‹ I'm an AI art creator experimenting with future human and city concepts using AI tools. Looking forward to learn and share ideas here!

pale wedge
wet grotto
#

guys, can I understand that nowadays to may be possible to use a open source model, the model needs a coder and decoder? is it?

#

or the coder/decoder isn't from model, but is a tool from the interface?

#

is this what is called "architecuture"? the way of the model is trained? Or the tools of the interface?

#

and that model (checkpoint) training with denoising is called Unet? is it?

lament kindle
#

a

still glacier
# wet grotto guys, can I understand that nowadays to may be possible to use a open source mod...

Lots of words in there I'm not sure you understand all of them.
Understanding how AI work can take months/years. It's definitely not something that you can master just by reading my over simplifying rambling. But here we go !

Usually coder/decoder/VAE in the context of stable diffusion refer to the part of the pipeline which transform "pixel information" into "latent information. What it means is that instead of working on an image made of pixel, you're working on an image made of "latent information". What is that ? Let's see.

The core of stable diffusion/AI models being a "neural network" meaning it's a huge graph made of (likely) trillions of neurons. Each neuron having the role of recognizing a specific feature, for instance : circle shapes, fur, nose, whiskers (in reality they recognize things so abstract we don't have proper word to identify it...). So if you feed all of those neurons the picture of a cat. The neurons in charge of recognizing cat features will lit up much more than the other ones. And throught the magic of math, the model can recognize cat. The map of all those neuron being lit up is a description of your image in "latent information".

But how does it make picture ? Simply put, we feed it a picture made of noise (random pixels) and ask the model "hey do you see a cat there ?" The model answer "dunno maybe there if I squint my eyes really hard I can see a cat in the top left". Then we modify the "picture" (in latent space) to reflect those results, ask again, etc... Until we have a proper picture in latent space. Then we ask the decoder to transform it into pixels.

If you're into music, think of "pixel space" and "latent space" as when you're applying a Fourier transform to a song to get its "frequency domain / space". Some operations are much more easier in "ftt space", same goes for "latent space".

#

Usually "open source model" refers to model with "public source". We know the code behind the model, how to run it, how it works and we can run it ourselves to replicate the results. (But it does not necessarily mean that the data used for its training are public tho)

#

it's really difficult to explain U-net properly tho.....
U-net is not "one thing", it's the assembly of multiple blocks working in latent space.
If it helps, think of its job as follows;
U-net segment/split/reduce the """""resolution"""""" (in latent space) of the image, then work on it much more easily because of its low """"""resolution"""""" then upscale it back.

#

I'm so sorry for any AI scientist reading this....

#

Not sure if I helped or made your headache 10 times worse.

wet grotto
#

@still glacier can you give me a time to reply you? I will see the unet video and answer you. I'm working now, I need time to reflect about all that you say!

remote brook
#

Hi, does anyone have any tips for generating images with sunglasses that are completely opaque?

ancient mauve
#

anyone here knows about inpainting?

#

I cant manage to get good inpainting with my pics

#

is there any trick to it?

oblique elk
# ancient mauve anyone here knows about inpainting?

It starts with good inpainting models. Not each model that exist can be (sucessfull) used for inpainting. There are flux inpainting models, SDXL etc.
Next thing is masking, feathering, etc. to get the right elements good embedded etc.

ancient mauve
#

I cant just use the lora for inpainting

oblique elk
ancient mauve
#

they are trained to maintain the artsyle of the image or does it give problems with less common artstyles?

oblique elk
ancient mauve
#

what do people use for non realistic images? anime, cartoon, comic...

#

@abstract quarry whats the model everyone's using for inpanting nowadays?

ancient mauve
#

@oblique elk do you have any model you like?

oblique elk
ancient mauve
ancient mauve
tiny spear
#

does anyone here use flux Kontext in comfyui?

#

If you want to do a stly transfer, how do you refer to the second image to transfer the style?

#

it looks like it uses an image stitch. I think the model only takes one image input

wet grotto
ancient mauve
#

sdxl basically

wet grotto
#

With realistic vison (sd1.5), the inpaint model don't create, just remove! the normal model create. This make confusion for a long time, I don't know if it's only for realistic vision, or is commom on all inpaint model. If it happens on sdxl can be it

#

In theory, what I undestand is that inpaint models are better to remove and substitute, and normal models to create

#

about mask options

ancient mauve
wet grotto
#

must be, juggernaut is huge! I like this model, Is my favorite from sdxl, but I used more sd1.5

#

but this is a normal model, not a inpaint version

ancient mauve
wet grotto
#

you kow civit ai? is easily to find models

wet grotto
ancient mauve
#

what model

wet grotto
#

I will show you

wet grotto
# still glacier https://www.youtube.com/watch?v=NhdzGfB1q74 This explains it way better than I c...

hi, Aretys, I saw the Unet Video, it is very cool! Is what I was thinking, about coder and decoder, noise and denoising... I like when you talk about what is open source, this help me a lot to understand better what exaclty is it, what is free, and how I said, about neural netwrok. Was very good the example. I talk with chatgpt and understand better what is architecture on sd1.5 sdxl / sd3.0, sd3.5/ flux šŸ™‚

#

can I DM you to show you one thing?

sturdy relic
#

I think im gonna ditch flux and go back to Stable Diffusion.
Working with that model in WebUI is absoluteley impossible

sturdy relic
#

Is there any way to use OpenPose with Flux in WebUi???

pale wedge
sturdy relic
#

I have the integrated controlnet, but the generation doesnt follow it. The file in the model section of OpenPose is only for Stable Diffusion and there doesnt seems to be a .pth file for Flux

pale wedge
#

hmmm.. have you tried generating the oamge separately then uploading it as a reference instead? not quite sure why that happened but maybe because flux is expecting a format that it can read?

sturdy relic
#

oagme?

tender minnow
#

Is there a way to inpaint without masking and using only prompts?

frosty tide
#

Hello, please I'm new to stable diffusion. Please let me know how to install SD on my pc.

warm junco
wary flint
#

Does something good come from here?

gentle prairie
#

does swarmUI automatically updates when i launch it? Didn't use it for a long time

tame vector
#

does someone already have a Yunara (new LoL champ) LoRA?

still glacier
gentle prairie
#

how do i download stuff on huggingface? All i see is "clone repo" and i have no idea how to do it...

still glacier
#

files tab, then click the arrow next to the file you want to download

#

or click on the files then on the download button

gentle prairie
#

thanks

gentle prairie
#

i'm trying to download the wan.21 I2V model
i have 10 GB ram
there's a 480P model FP8 and a FP8-scaled, which one should i pick?

#

Also how do i save a video in SwarmUI? I see no option to download or save anywhere

#

i see so many conflicting and cryptic info about all the img2video everywhere, i can't find a definitive answer (sorry i'm bad at this, never did video before, only used txt2img on a1111)

sleek fable
#

hey,
im using A111 with SD but want to get into Flux locally....
Which is the best way to start? ComfyUI or Forge? I m looking for best support also for LORAs.
I have a GeForce RTX 2060 Super

#

I also dont like switching between stuff often, so I would prefer the environment that is best for long term use

gentle prairie
#

is it normal that img2vid takes way longer that text to vid with wan 2.1?

wet grotto
#

someone already hear about Mellon UI?

sleek fable
wet grotto
#

I don't know, I see something about it but I don't undertand if is a future project

#

looks like ComfyUI

sleek fable
pine path
#

@vapid dove support scam

pine path
desert dagger
versed rain
#

Hey everyone,
I recently used AI to generate some weird futuristic designs ā€” imagining what humans might look like if they lived underwater.
The result is strange, otherworldly... but also kind of beautiful.

I made a short 15-sec YouTube video of the result here:
https://youtube.com/shorts/WLOVT8nHYT4?si=arc-d4cl3H7W5zPF

What do you guys think?
Would you survive in a world where humans evolved to live underwater? Or what would you change in this design? šŸ¤Æ

Any creative thoughts or weird ideas are welcome! I might generate new ones based on your comments. šŸ˜Š

fathom meteor
#

Any reliable website where u can hire people to train Lora for you?

still hare
#

Anyone have open webai installed locally? I got ollama and a couple other models working on it and I heard theres a way to integrate stablediffusion into for it images inside chats. Not sure how to go about it doe. Thanks for any help

oblique elk
#

You will need to have comfyui, auto1111 running to generate images.

#

Will get busy on your host as language model and a model like flux would fill up the vram very very fast.

urban summit
#

Hello, I just purchased Stability AI credits and I'm looking to test out thei video API.
Is there a place I can request accesss to it?

#

Also, hello everyone. šŸ–šŸ»

wet grotto
wet grotto
#

can you talk more about it?

#

they are the same person, I just saw his github

bright ember
#

guys what are the resolutions to use on forge for photos?

spiral badger
#

Hello

oblique elk
bright ember
#

ok,thx

desert dagger
#

He's tired of the issues comfy has and is writing his own interface

wet grotto
#

ginius*, I hope he finished his project! S2

desert dagger
#

So does evrryone else. Hes got a channel for updates for it on L3

wet grotto
#

interesting, I will take a better look on that! I will take attention

sullen vapor
#

Call for Participants: Help Shape the Future of AI!

My name is Elie and I'm a researcher at Stripe Partners, a London-based research agency. Weā€™re preparing a study on how people use agentic AI tools in their personal lives ā€” tools like AI agents, operators, browser extensions, or chatbots that can complete tasks for you, such as planning a trip, organizing your calendar, or building a website. We want to understand how people are using this new technology, what they like about it, and any challenges they have faced trying to make it useful.

Weā€™re looking for participants who use these kinds of tools outside of work to take part in a paid research study. This will include:ā€Ø
A short digital diary over several days
A focus group with others exploring similar tools

If youā€™re interested in how AI fits into everyday life ā€” and youā€™re using it to get things done ā€” weā€™d love to hear from you!

jaunty hinge
thorny bloom
near silo
#

I'm a super veteran for image gen and ESPECIALLY text gen, but I'm so new to video gen. I've been using text and image generation locally since basically the day they both started, and I've seen the growth trajectory of both technologies, and then I look at video generation, and there is just so much competition, and rapid improvement. It boggles my mind and overwhelms the crap out of me!

I don't even know where to begin šŸ˜­

I've been using frame pack for like 2 months, but I took a big break in the middle. I just updated it and started using it again and I'm getting some pretty cool results, but I know that the results that I'm getting are nowhere near what you can get now with newer models, but I just cannot for the life of me figure out where the hell to begin

Does anybody have any suggestions or resources that they can point me towards so I can try and dip my toes into more advanced local video generation?

inner latch
#

Hey folks! šŸ‘‹
Has anyone tested Stable Diffusion (especially SDXL or video models) on the new AMD Ryzen AI Max+ 395?

Iā€™m particularly curious about:
ā€¢ Performance when generating video (e.g. using FLUX)
ā€¢ Compatibility with ComfyUI, FLUX, and video workflows
ā€¢ Any issues or limitations with drivers / acceleration on AMD?

I know itā€™s a very new chip, but if anyone has hands-on experience or even rough benchmarks ā€” please share! šŸ™
Would love to hear your thoughts before investing.

Thanks in advance! šŸŽ„šŸ’»šŸ”„

barren juniper
#

Hey im trying to use the hassakuXLpony model in python with the diffusers library and i want to increase the token limit for clip so i can pass more tokens, how can i do this?

covert sand
#

does forge or reforge have controlnet built into it?

jaunty hinge
covert sand
jaunty hinge
urban summit
#

Anyone wanna tell me how I can use my Stability API key with credits to generate videos from my discord bot?
I would really appreciate it.
I can get the image generation figured out, but not the video.

versed rain
#

Hey everyone,
I recently used AI to generate some weird futuristic designs ā€” imagining what humans might look like if they lived underwater.
The result is strange, otherworldly... but also kind of beautiful.

I made a short 15-sec YouTube video of the result here:https://youtube.com/shorts/5JCtGOwgixU?si=wS0IoUL-nHzSjr19

What do you guys think?
Would you survive in a world where humans evolved to live underwater? Or what would you change in this design? šŸ¤Æ

Any creative thoughts or weird ideas are welcome! I might generate new ones based on your comments. šŸ˜Š

warm junco
warm junco
plush roost
#

Hi all!
If you work with AI or large language models, Iā€™ve built a free open-source toolkit (WFGY) that can instantly boost reasoning accuracy and stabilityā€”no extra setup required.
Details are in my profile/bio if youā€™re curious. Feel free to check it out!

wet grotto
#

"NVidia Cosmos Predict2! New txt2img model at 2B and 14B!"
I saw this on reddit
It's correct I assume that the only open models now are SD, Flux and now Nvidia?

wet grotto
thin maple
#

how to genrate text to video

grand rain
#

my A1111 randomly screwed itself up...

#

i didnt run git pull myself, but now infinite image browser and pretty much nothing else works šŸ˜¦

#

what could have happened?

#

does it auto update?

warm junco
echo jackal
#

does anyone know if i can use a111 with my 5060 ti or not

#

why is it showing me thats it not compatible with pytorch

valid aurora
#

hello guys, does anyone know if there are significant differences between a 32gb vs 24b vram in image generation? maybe in video/gif generation too

sonic forge
#

Hello, everyone!

I'm an AI Full-Stack Developer with over 8 years of experience in the software industry.

šŸš€ Services I Offer:

Automation: I specialize in automating tasks using tools like n8n, Zapier, and Make.com.
NLP: I handle advanced NLP tasks with models such as GPT-4.5, GPT-4o, Claude 3-7 Sonnet, Llama-4, Gemini2.5, Mistral, and Mixtral.
Model Deployment: I assist with the seamless deployment of machine learning models across various platforms.
TTS / STT: I implement both TTS and STT solutions for interactive and conversational AI experiences.
AI Agents & Chatbots: I develop custom AI agents, Agentic AI, chatbots, and VoiceFlow applications for diverse business needs.

šŸŒ Portfolio: Check out my portfolio in my profile

If you have an innovative project idea, feel free to reach out. Letā€™s bring your vision to life!
Thank you.

tame vector
#

which ComfyUI Sampler is the equivalent to the Euler A Sampler on Forge?

trail matrix
#

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

tame vector
# trail matrix hi, can someone help me. i generated good picture but it have bad eyes, someone ...

You can use ADetailer to give specific features more detail (for example face, eyes, hands...)

Guide for installation and usage: https://youtu.be/Gxj_bYoZ4dg?si=O_bwyt9_ui8z577Q

If you want to download custom ADetailer Model, you can find on Civitai.
For eyes i recommend following: https://civitai.com/models/150925/eyes-detection-adetailer
Dont get confused about the words "finder" or "detection", the ADetailer simply locates the specific features and adds details to them.

Also, if you're wondering where to put your downloaded file so you can select it in Stable Diffusion under ADetailer:
...\stable-diffusion-webui-forge\models\adetailer

#

.

Anyone have a ComfyUI Workflow i can use for Character Illustrations?

I used Forge up to this point, there i used some LoRAs, ADetailer and sometimes i used ControlNet.
I've tried just drag and dropping a Forge generated image into ComfyUI to "recreate" the workflow. A Workflow appeared, but once i started generating, the image was just random noise...

oblique agate
#

on my whatsapp volunteer group, a girl is that ai is about screwing over the little guy. I told her ai is about control and if you can't beat ai join them. Classic simps jump in that's defeatist altitude. Another you have to sacrifice yourself so others can have a good life altitude.

inner latch
#

Hey folks! šŸ‘‹
Has anyone tested Stable Diffusion (especially SDXL or video models) on the new AMD Ryzen AI Max+ 395?

Iā€™m particularly curious about:
ā€¢ Performance when generating video (e.g. using FLUX)
ā€¢ Compatibility with ComfyUI, FLUX, and video workflows
ā€¢ Any issues or limitations with drivers / acceleration on AMD?

I know itā€™s a very new chip, but if anyone has hands-on experience or even rough benchmarks ā€” please share! šŸ™
Would love to hear your thoughts before investing.

Thanks in advance! šŸŽ„šŸ’»šŸ”„

maiden violet
#

I downloaded a detection model for adetailer but I have no idea where to put it so I can actually use it

left acorn
#

Bit of a noob question: I want to take image A e.g a landscape, and reimagine it in a different time. What would my best bet be to produce that? Control net is good but a bit too restrictive I think

#

something like what chatpgt will do, where itā€™ll draw from the image when creating the new image, but it isnā€™t restricted to the depth map

stiff nexus
#

Hi guys, I'm a beginner who wants to generate celebrity faces for my new yt shorts channel. I own a pc with gtx 1060 6gb and 16gb ram. I'm seeing if there are solutions where the model can be run locally. Where do I get started?

tame vector
languid panther
#

hey so I wanted to create my first lora and I was wondering if any of you in here have any recommended parameters for training. I have read a lot of pages but on each one they say something different like the recommended number of steps and epochs. My lora would be based off an artist style and so it's supposed to have more steps, but I can't come to decide what numbers to use. Any help would be thankful

warm junco
# echo jackal why is it showing me thats it not compatible with pytorch

You have to manually switch the pytroch version;
try this:
Go into stable-diffusion-webui folder
Open cmd prompt in that folder by clicking into the adress bar and type cmd and hit enter.
Then copy and paste these three commands one by one:
venv\scripts\activate
pip uninstall torch torchvision torchaudio -y
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

Then relaunch the webui-user.bat

craggy rune
high flint
#

Hi, everyone. I'm a web developer. I find someone who can lend me a real account. I'll pay you 20% of my salary. If you want this, please contact me.

keen saffron
#

A Hollywood friend of mine is messing with stable diffusion and green screen for high level production. He made this concept piece to hopefully pitch to the big wigs. I am just showing my friend who is wildly unappreciated some love. Thanks for your patience, guys!

https://youtu.be/yyx1eXcqrdc?si=mM6yFVSWmBTi_WY9

sonic forge
#

Hello, there!
I'm an AI Full-Stack Developer with over 8 years of experience in the software industry.

šŸš€ Services I Offer:

Automation: I specialize in automating tasks using tools like n8n, Zapier, and Make.com.
NLP: I handle advanced NLP tasks with models such as GPT-4.5, GPT-4o, Claude 3-7 Sonnet, Llama-4, Gemini2.5, Mistral, and Mixtral.
Model Deployment: I assist with the seamless deployment of machine learning models across various platforms.
TTS / STT: I implement both TTS and STT solutions for interactive and conversational AI experiences.
AI Agents & Chatbots: I develop custom AI agents, Agentic AI, chatbots, and VoiceFlow applications for diverse business needs.

šŸŒ Check out my portfolio in my profile

If you have an innovative project idea, feel free to reach out. Letā€™s bring your vision to life!
Thank you.

jaunty hinge
#

ummm interesting

#

Dm please

sweet gulch
#

hmm..

wet lantern
#

Yo guys, My cousin is looking for a EU Boyfriend

fervent thunder
#

Hello there

candid gulch
fervent thunder
#

I'm new here, I was referred to come here to learn how to set up and configure SD to generate on my machine, but I can't find much on the channels.

floral umbra
floral umbra
warm junco
floral umbra
#

@fervent thunder 1 crucial point to start out with from the pinned messages:

Always check gits for what each SD generation tool requires of python, pytorch and cuda versions. And install cuda toolkit for each version of cuda you use. And make sure they are compatible with eachother. As that's one point i struggled with most when i started out :P

#

The biggest annoyance, is when i need to have several comfyui venvs as some nodes are too outdated, so can't use the latest stuff on those lol.

worthy zephyr
#

Hi! I got a quick question, is there an open source version of Flux Kontext that I can use locally for free?

fervent thunder
fervent thunder
floral umbra
#

Plenty of workflows online you just drag and drop into the webui for easy recreate images thinky

floral umbra
worthy zephyr
#

Ok thanks for the info!

burnt moat
#

Hey everyone! Iā€™m Blackemarlin ā€” big fan of cinematic AI art and post-apocalyptic vibes. I mostly use Sora + Photoshop and love storytelling through images. Looking forward to seeing your work and sharing ideas!

oblique elk
visual pond
#

Yo guys is ComfyUI better for realism or auto1111, or is there any diffrence

oblique elk
visual pond
#

this is what ive watched in YT

upper plinth
#

Are stylistic loras for WAN 2.1 a thing?

#

dont click that. its an obvious virus.

storm cape
#

@karmic brook

#

@hallow edge

alpine bear
jovial ruin
#

Hello
I've question I want to test stable diffusion.cpp with small size model
which works

  • sd3_medium_incl_clips_t5xxlfp16.safetensors
  • sd3_medium.safetensors
    ?
#

would sd3_medium works with prompts ?

sinful seal
#

hey guys im new to stable diffusion. Any advice on getting to to not generate people with multiple limbs or deformed bodies?

upper plinth
#

me

valid aurora
#

does anyone know how those video generations are made? ive only used stable diffusion but would like to try it one day. im looking to upgrade my pc rn so i wanna know what specs is needed for those video generations. is it just vram like SD ?

warm junco
bold nova
#

Hello!
I need help activating the free trial version of Google AI Premium. Due to regional payment restrictions, I canā€™t complete the required verification step myself. If someone can assist me with this, I would be very grateful! Please DM me if you can help. Thank you!

warm junco
bold nova
marsh forge
#

how to use the bot?

faint timber
#

@sinful seal Make sure your output resolution is not exceeding your model's resolution. Controlnet canny can help maintain a pose or composition from an image. Revisit your prompt and or add negative prompts.

last lodge
#

hi

spare flicker
echo jackal
#

Is there a way to save your previous prompts? Or at least access them once you've closed stable diffusion?

warm junco
echo jackal
brisk grail
upper plinth
#

sure, let me know specifically what youre struggling with

#

Has anyone tried to use IP Adapter or Reactor with WAN yet? If this is possible, this would be huge

#

VACE/WAN is already insane at replicating subjects with just a reference image. Using IP Adapter or Reactor alongside would be perfection

bold nova
#

Hello!
I need help activating the free trial version of Google AI Premium. Due to regional payment restrictions, I canā€™t complete the required verification step myself. If someone can assist me with this, I would be very grateful! Please DM me if you can help. Thank you!

round cosmos
#

Hello!

tired nacelle
unborn hedge
#

does anyone know of a Wan VACE Video-to-Video Workflow download anywhere? im not familar with comfyui and i dont wanna spend hours tryna figure out how to make this work, and i cant find it anywhere online?

fervent thunder
#

is there anyon who can solve my error in comfyui

summer sedge
#

What should I use for video generation, WAN, LTX or FramePack?

So far, I got WAN to run in comfy. LTX is just giving me errors most of the time (also in comfy). FramePack works fine but feels kinda limited in its UI-options and mostly produces completely still backgrounds (in non-realistic pictures).

vast furnace
#

I'm trying to create a character and once I describe everything and have it generated, the result is completely different from what I wrote. Where am I making a mistake?

faint timber
#

It's hard to say, without seeing the prompt and knowing what model you're prompting against.

still glacier
#

and the tools / settings used

upbeat gorge
#

Hi guys, im back at ai image generation, and i want to make product images. I have my products and i need to keep the image and make different scenarios.
Theres some like stable difussion GUI. like the NMKD but "2025" where i can use controlnet?

Thanks in advance šŸ™‚

rain spindle
#

Hi, i try the first time Stable Diffusion and download SDXL 1.0 as BaseModel. But i have some problem with rendering things without "anime/Manga/illustration" style. It would be not generate things in photorealistic style. Can someone help me to fix that missue? I dont add an Lora or Refiner.

abstract quarry
#

just use prompts like "35mm photography" or "digital camera photo". However, SDXL base is a weak model. You should use models specialized on photography

#

for sdxl these are models like realviz

#

or use flux. But flux base also has some kind of plastic look and needs custom loras for perfect photorealism

rain spindle
#

i tried on 512x512. 1024x1024 is better i read here in some channels. I create atm some pictures, but it still a long time because atm i work on "only cpu" mode (Intel Mac). But i think, this way with 1024x1024 is very better. Thank you for our help and sorry for my broken english. Thanks a lot.

#

99% of bugs sitting in front of the system...šŸ¤”

worn aspen
#

Hi all! What's a good tool chain to generate a variety of consistent images for games, sprite sheets? I haven't played with stable diffusion in a year. I think I'm using the comfyui interface. Based on my memory if what I was experimenting with, I can't see a path to solving my question, but maybe someone more experienced has a suggestion.

valid aurora
#

Does anyone know whats the best tool for animating an anime style SD image ? Like if I have a character and I want to make him move. From what Ive heard, ppl say Luma is good

valid dirge
#

āœŒļø

lucid violet
#

Hi

spare flicker
pale wedge
round cosmos
#

Hello!

floral umbra
#

Do you guys know if there's any comfyui workflows, or just a git repo with code designed to create lora preview images? As lora manager fetches what's available of images from civitai, but a lot of loras fetched elsewhere has no images.
As i tried to make a automated script, but it doesn't fully give the output i want it to have.

valid aurora
#

@pale wedge thanks dood

hollow fog
#

Hello!

tropic frost
#

anyone know a good tutorial to makes loras? Not necessarily character loras, art style loras would be nice

merry stump
#

Hello!!

unique ravine
unique ravine
# tropic frost anyone know a good tutorial to makes loras? Not necessarily character loras, art... 
To train a GLora, you'll need Jelo's fork of Derrian's trainer
This will break a library, to fix it go to the trainer's location and run these commands:
.\backend\sd_scripts\venv\Scripts\Activate
pip uninstall numpy
pip install numpy==1.26.4

Starter Settings
Gradient Checkpointing: True
Gradient Accumulation: 4
Batch Size: 4
Number of Repeats: 1
Max Training Time: (Epochs) 50
LoRA Type: GLoRA
Network Dim/Alpha: 18/18
Unet Learning Rate: 2e-5
TE Learning Rate: 1e-7
Optional Args: "weight_decay=0.07", "betas=[0.99, 0.999, 0.99995]"```

https://github.com/Jelosus2/LoRA_Easy_Training_Scripts
#

aint gonna get better than that

serene shoal
#

Nice to meet you, all friends. I am here to help you.

unique ravine
#

@steady parrot still need?

steady parrot
unique ravine
steady parrot
sonic void
#

I would like to ask you how to achieve the effect of stable diffusion patching pictures or eliminating parts of pictures for appropriate filling.

sleek fable
#

hi,
i have a RTX 2060 Super with 8GB Vram and while this is ok to run, I wonder if I can, for a couple hundred bucks, aiming 500 EUR max, increase the Power of my PC for Local (Forge Based) AI generation. I saw there are some weird cards that basically pack VRAM on top, but I m not familiar with that kind of stuff.
TLDR can I somehow buy a dedicated card ON TOP of my RTX 2060 Super that helps with local AI?

#

I want to primarily be able to load big models that require 16, 32 or 64 GB

warm junco
sleek fable
#

oh, that one is cheaper than I thought

#

thx!

warm junco
#

np šŸ™‚

warm junco
#

also if your case has the space go for a 3 fan design

sleek fable
tribal crown
#

haven't been here for a while here and would like to know what's the easiest to install comfyui i2v generator/workflow that has loras and all of that stuff? thanks

mellow meteor
simple trail
#

Is there an ai service that you can make text to speech but change many aspects of the voice; for a character voice over ?

oblique agate
#

I am waiting for 5080 ti/super with 24gb vram

low tulip
#

What exact tech do companies that take like 10 selfies of you and generate professional headshots use?

#

Do they have some stable diffusion based pipeline they optimized, or do they have their own foundational models?

abstract quarry
#

who knows. But it's so easy to do that I guess they just use Stable Diffusion or Flux or similar models

low tulip
#

I tried creating a Lora with stable diffusion and it was pretty awful

#

i guess i just suck at it?

abstract quarry
#

Flux is really good for that. But SDXL is decent, too

low tulip
#

how many photos are optimal? Also if i managed to get some ai generated photos that look nice and look like me, can i include those as training data too? why not, right?

abstract quarry
#

however, you can also use FaceID or IPAdapter and stuff like that

low tulip
#

What are those? tools to just replace face?

#

but i'd like for my body shape to be preserved too, if possible

#

I'll look into them too though

abstract quarry
#

they act a bit like controlnets, but different. You give a face photo, they transfer the facial features into the character in your image

#

of course, for best results you should train a lora

#

my discord profile photo is a Flux generated image lol It's so fucking realistic, every single hair of my beard is correct

low tulip
#

oh wow

#

so you trained your own lora

#

on flux

abstract quarry
#

yes. Flux gives best results, in particular for photorealism

low tulip
#

how many photos did you use? what platform did you use? did you spin up some machines on runpod or something?

abstract quarry
#

my own machine, a lot of photos (probably 100...) and a lot of time X_x

low tulip
#

ohhh

abstract quarry
#

but you don't have to

low tulip
#

wow im jealous

abstract quarry
#

10 images are enough

low tulip
#

i don't have a good machine

abstract quarry
#

in particular if they are high quality

#

my own photos were all a bit shitty cause my smartphone is so old and the camera is not very good

low tulip
#

what constitutes high quality? can i just do 10 selfies?

#

Did you follow any lora training guide? or do you already know the best settings and stuff

#

and what exact model did you use? flux base?

abstract quarry
#

yes, Flux base

low tulip
#

i tried using runpod as well as rundiffusion for training a lora with sdxl i believe, and i just hate working with cloud machines :/

abstract quarry
#

lora rank 24

#

I think learning rate around 1e-4 to 5e-4 depending on how much time you have

#

I used SimpleTuner, but other tools are probably also okay

low tulip
#

Ah never heard of it

#

i used the standard lora training thing

#

i forget the name

abstract quarry
#

there are many xD

low tulip
#

starts with k

#

lol

abstract quarry
#

kohya probably

low tulip
#

it seems a bit like automatic 1111

#

yeah that one

#

i didn't like it too much, too many settings

#

i wished there was something way simpler

abstract quarry
#

I used kohya a lot in the past. It took a while until they had Flux support, thats why I switched to SimpleTuner

low tulip
#

I see. I'll def take a look at simpletuner thanks

abstract quarry
#

its not an easy to use tool xD More complicated than kohya

low tulip
#

Did you use the BLIP captioning

#

or any captioning? or just trained on one keyword

abstract quarry
#

no, never do that xD

low tulip
#

Ah lol

#

why not though

abstract quarry
#

BLIP is very old

#

I would use modern multimodal llms nowadays

low tulip
#

like chatgpt?

#

lol

abstract quarry
#

yeah. I prefer local ones, though xD Otherwise all my images are on ChatGpt

low tulip
#

ahh

#

like which ones can you run locally?

#

open source?

abstract quarry
#

Gemma and Mistral are quite good. I would use Gemma if you have enough vram

low tulip
#

gosh.. i think i need to get a machine with graphics card

abstract quarry
#

yeah, so annyoing that all these old gpus are so expensive still

low tulip
#

if i use my friend's macbook m4 with 24 ram can i run this stuff? or do i need nvidia CUDA?

#

i want to use cloud machines but they are SOOOO annoying to use

abstract quarry
#

you can probably, it will be just super slow I guess

low tulip
#

ah you think they'll be slow?

#

what if the macbook had like 32 g ram

abstract quarry
#

Gemma quantized runs well on 24gb vram. You might even quantize it down to use it with 16gb vram, but thats all possible

#

(not all possible, but all that makes sense)

low tulip
#

But ultimately the photos should be captioned right? it's not a good idea to train with just the main keyword?

abstract quarry
#

hm, my feeling is that captions are better in the long run

#

when training only one a keyword, the model might later have problems if you have a long prompt

#

but you can train only on a keyword and it will work

#

it will just get worse results if you use it later on a very long prompt

low tulip
#

Ah, i see

#

btw if i have like a 5 year old desktop but with decent cpu and ram

#

and i upgrade the graphics card

#

will it be sufficient for this stuff?

#

or will there be a bottleneck

#

it's probably more like 10 years old but its still like an i7 cpu

#

with like 8 gb ram

abstract quarry
#

uh no, you need more ram and a good graphics card

low tulip
#

oh

#

normal ram matters too?

abstract quarry
#

the model should fit into your normal ram

low tulip
#

i see

#

ugh i feel like i should just learn to use cloud machines well

abstract quarry
#

yeah, its probably much cheaper than getting new hardware

low tulip
#

cool cool. I think i have a much better idea of where to go. Thanks a ton for your advices. I sent a friend request in case i have more questions later. Please accept šŸ˜€

stone tangle
#

Hello

wet grotto
#

can someone sugest me a site to rent a pc to me use my architetural programs and play games? My pc is a potato

last merlin
#

post: young girl anime showing tidies and vagena
civitai: looks good to me

me: semi-2d girl cycling bycycle with all skin covered
civitai: realistic image of minor is not allowed, you dumbb-ass pedd

simple trail
#

anyone know any things for voice generation where as I can change many options of the voice for text to speech ?

floral umbra
#

Always fun when you have 100's of loras without pictures, and you have to generate previews for every single one for lora manager eugh

magic spade
#

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

shrewd drift
#

Hey! I'm looking for a skilled Veo3 user to collaborate on an IG video project (Not freelance work ā€“ long-term partner setup.) ā€“ DM if interested.

stone tangle
#

Hello everyone!
I am š€šˆ & š†šžš§šžš«ššš­š¢šÆšž šŒšžšš¢šš š„š±š©šžš«š­ | š‚šØš¦š©š®š­šžš« š•š¢š¬š¢šØš§ | š•š„šŽšŸ‘ | š€šˆ šˆš¦ššš šž/š•š¢ššžšØ/šŸ‘šƒ š†šžš§šžš«ššš­š¢šØš§
I can help you.
please DM!

#

šŸš€ š€šˆ & š†šžš§šžš«ššš­š¢šÆšž šŒšžšš¢šš š„š±š©šžš«š­ | š‚šØš¦š©š®š­šžš« š•š¢š¬š¢šØš§ | š•š„šŽšŸ‘ | š€šˆ šˆš¦ššš šž/š•š¢ššžšØ/šŸ‘šƒ š†šžš§šžš«ššš­š¢šØš§

šŸ‘‹ Hey creators and innovators - I'm your partner for AI-driven image, video, and 3D generation, computer vision, and custom machine learning solutions.

šŸ§  What I Do:
ā€¢ Generative AI - AI-powered image, video, audio, and 3D generation using VEO3, Stable Diffusion, ComfyUI, and Hugging Face
ā€¢ Computer Vision - Object detection, pose estimation, and video analysis with YOLO, OpenCV, TensorFlow, and PyTorch
ā€¢ AI Model Development - Fine-tuning, training, and deployment of custom models for vision, media, and generative tasks
ā€¢ AI Video Generation - Leveraging VEO3 and diffusion models for cinematic, high-fidelity AI videos
ā€¢ Automation & APIs - Backend pipelines, API development, and full-stack AI-powered tools

šŸ“Œ Highlights from My Work:
ā€¢ quickpose.ai - AI-powered real-time pose detection for fitness and sports
ā€¢ www.hellohistory.ai - AI chat app for conversations with historical figures using LLMs
ā€¢ kyzo.ai - Computer vision system for movement tracking and analysis

āœ… What You Get:
ā€¢ Fast execution
ā€¢ High-performance AI pipelines
ā€¢ Creative and production-grade generative models
ā€¢ Transparent communication and daily updates

Let's create next-generation AI tools together. šŸš€ DM me anytime!

mighty glacier
#

Hy anyone's have hey gen ai membership contact ya who know how to do ai voice dubbing contact me

tiny zealot
#

Hello

low tulip
#

For training LoRa's on people, is it best to have simple selfies / headshots, or is it better to have a variety of photos?

grim stirrup
#

hello hi

low tulip
#

@abstract quarry should i run preprocessing on the training images? should i resize them? should i crop them to show mostly face? should i upscale them?

raw hemlock
#

Hello guys, I'm new here and trying to understand everything. I've heard Kohya is the best LoRa trainer in term of quality and professionalism, is that true ? (better than Flexgym but harder to use ?)

Also, what tool you use then to generate the images with the LoRa trained with Kohya ?

AUTOMATIC1111 Web UI
ComfyUI
InvokeAI

Which one of these is the best according to you ?

Thanks a lot

low tulip
#

@raw hemlock Hey, im kind of new to and trying to learn

#

I had SimpleTuner recommended as lora trainer

#

you can take a look too

#

and i think ComfyUi is generally considered best for generating

fervent thunder
#

hi

woeful saffron
#

hi anyone on?

spare flicker
woeful saffron
#

@spare flicker Can u help me

spare flicker
woeful saffron
#

whats this issue

#

pls help me somekone

warm junco
woeful saffron
#

control net

#

i want to use like image to video

#

u help me na pls

#

i cant find a fix of this

#

@warm junco

warm junco
woeful saffron
rain spindle
#

Need someone who can merge me a model with 8 Loras.

I cant to it because I have only CPU work with my SD. (MacBook Pro 2019. šŸ« )

And now im Not able to merge this.

So, if we found someone, ill pay them some Coffees!

rain spindle
#

@stone tangle you?

abstract quarry
# low tulip <@321288280651857922> should i run preprocessing on the training images? should...

don't upscale them. Downscaling is okay, upscaling will decrease quality. You want to have a good mix of training images with some showing face as portrait and some show upper body or full body. Face is most important, though. You can crop an image to only show your face, but this only makes sense if the image is super high resolution, otherwise, the face will be too lowres. Making a good portrait photo is better, probably.
Flux wants your images dividable by 16 pixel.
Most tools automatically crop your photos such that they have the right resolution. You have to be careful though, as sometimes they do a lot of "preprocessing" that rather damages your data

#

like most tools come from a time where you HAD TO train on very specific resolutions and so they expect you to provide the resolutions as parameters somehow

#

for flux, you can train on any resolution you want, so just make sure that your training images are not accidantelly cropped or resized in a weird way by the training tool you use

#

you might also add some low res photos (e.g. 512x512, 768x768) to your train data to speed up training. Just downscale your highres photos (1mpx) down to the 1/4 mpx)

#

one more thing: add group photos to your training data, i.e. photos where you and 1-2 other people are clearly visible. This is important for the model to learn that not everybody is you šŸ˜‰

abstract steeple
#

hello

low tulip
abstract quarry
#

yes, flux can

#

just caption them as you want, e.g. with "photo showing [trigger word] and another man with black hair" or "photo showing [trigger word] on the left and a woman on the right"

#

it's just important that the method has some training images where it sees your face together with other faces

low tulip
#

Ah I understand. but it's also interesting, first time hearing something like this. Have you done comparisons of LorAs trained just one solo pictures and LorAs trained on solo and group pictures and noticed a difference?

abstract quarry
#

yes. Without group photos it will transform any face into your face

#

its even enough to add collage photos

edgy shard
#

has anyone managed to make chroma work in ForgeUI yet, now its been added/

abstract quarry
#

with collage photos I mean: just add two portrait photos next to each other into one image

#

the important thing is that the model learns to recognize your face from other faces

low tulip
deep slate
#

is there a starboard here still?

abstract quarry
#

Flux Kontext is out! šŸŽ‰

drowsy robin
#

Is Flux Kontext better than dev at text to image

#

or is it only better at image and text editing?

floral umbra
drowsy robin
#

I assume nobody knows whether it is or not.

#

šŸ¤·ā€ā™‚ļø

raw hemlock
floral umbra
desert dagger
#

Luca Taco has several very good trainers for different models in his replicate space

crisp cobalt
#

Do you use the discord for rendering or some kind of app?

still hare
#

Hi, I have a 4080 super and got a1111 working w some random model. What's the best model I could reliably? For text to image? Or if possible text to video?

valid aurora
#

does anyone know if you can use blender or any vfx tool to polish anime videos? my plan is to create an anime clip then polish it with blender. is this feasable? from my knowledge you need to have a 3d model to use blender not a video. (havent tried any of them yet, so im clueless)

fervent thunder
#

What are some good UIs for Stable Diffusion nowadays? (No Comfy)

granite zephyr
#

Hey! I just launched a blurred vault called SceneSnatch ā€” would love to join any creator network servers or vault collabs youā€™re part of. Open to swaps or promos!

pulsar night
#

guys can i make a job post here?

pale wedge
#

there's also invoke ai. cleaner buttttt also has a higher vram gluck w the gpu memory huhu unless u invest in good GPUs then ur set w this

#

there's also savro which is a plug and play, so it's easier to use. i use this nowadays since i make the model myself and there's consistency

#

in general there are a ton aside from comfy, u can use whatever suits your liking. i personally prefer savro thoo

inland wyvern
#

Hey. Can you make images based on an input image using SD?

#

Like a reference image to copy artstyle from?

abstract quarry
#

there is flux Kontext that can do that

#

released yesterday

#

for stable diffusion there are several ipadapter

cloud flax
#

Not sure where to ask. Are there any decent low-profile GPUs for stable diffusion?

empty pond
#

Looking to create a series similar to black mirror. If anyoneā€™s interested hmu.

low tulip
#

with Flux Kontext, can you do a controlnet to only modify certain areas?

#

I'm guessing yes?

lavish olive
#

Hello pals,
I don't know literally anything about AI at this point, but I am an expert in traditional art and VFX. I want to be able to take a fairly simple physical object as a video of all angles of it (or whatever) or photos of it from all angles and compose it into an AI created video, or in-painting it into something else etc.. Would this require training LoRa for my specific object? I am interested in consulting with an actual expert (at least that can do what it is I am talking about and understands the comfyui pipeline or whatever for it), paid of course, or interested in someone here (or anywhere, or a recommended platform to hire) an ai artist to do this for me as a example as I can't find similar thing online really.

Anyone know or is interested please let me know. I'm going to start learning about this stuff full time from this week on until I get it sorted, find someone to do it for me, do it myself, or learn what I want is not possible.

Thanks, and please to meet you all.

lofty wraith
#

hi

wet grotto
#

guys, what is the difference from Forge to ReForge?

abstract quarry
low tulip
#

Also is there a comfyui node for comparing different outputs? maybe with a slider

oblique elk
low tulip
#

Does Flux respond better to a different prompting style than stable diffusion? i can't remember if i read something like that or not. Like more prose natural language prompting style vs SD keyword style prompting

oblique elk
low tulip
#

Can you use multiple images as input for kontext? like 5 instead of just 1.

oblique elk
stoic mica
#

Any advice on how I can make the videos I generate on the Image-to-video model better?

Unsure of how to go about this. For instance the image is fine, I have even created a separate AI agent node in my workflow to create a video prompt and advise on how the video should be brought to life, though the rendering isnt great - clearly shit

low tulip
# oblique elk yes you can

And how do you feed that? I see there's a image stitch node i can use to stitch two photos together. But more? Do i have to stitch the photos manually first?

#

wait, comfyui is open source right? i want to contribute a node that stitches arbitrary number of images into a grid

#

Should i? lol

oblique elk
# low tulip Should i? lol

Well there is a image stich node which can be used for it and repeated use leads to the same result, so well it might not be required.
Feeding depends on your scenario. If you want to make a group picture etc. you would stich all the images together and then prompt them for whatever they should do.... The people on the picture are sitting at a restaurant,....
If you have a special background etc. you would use the background as another condition and merge the group and the background.

#

But as the model work best with step based editing i would think about adding steps between to avoid glitches

low tulip
oblique elk
low tulip
#

so i can iterate without having to download and upload over and over?

oblique elk
low tulip
#

A question not related to Kontext.
I'm trying to create a Lora and lets say i have 10 images. I've already created workflows using Lora or otherwise (using Kontext or other models etc) and generated maybe 20 more images, based on the 10 original. Of course the 20 were handpicked to be the best representations, out of lets say 100s of outputs.

Now, if i train a new LoRa on the 10 + 20 = 30 images, will there be an increase in quality over a lora trained on the 10 original, even though the 20 extra are "based off of" the original 10?

winter cloak
#

How do I enable Xformers in ComfyUI?

abstract halo
#

I am kinda nooby here is there any resource where installation of latest model is shown?

warm junco
dawn mulch
#

Does the flux context dev make IP adapter and maybe controlnet more or so obsolete when it comes to editing images? flux context dev seems to be much better at generating the same face and such. IP adapter face ID seems to max out at about 80 to 90% but flux contacts dev seems closer to 99%

hexed barn
#

is there discord community for automatic 1111 user?

warm junco
velvet grail
#

hey hey everyone, Elios is the name, animation is my game! ā¤ļø

bleak anchor
#

Hey

lofty sapphire
#

hey

abstract quarry
#

ipadapter works by processing the image by some vision model and extract information out of it. These information are then injected into the diffusion model. The important thing here is that your input image is not given 1:1 to Flux, but only some kind of description or summarization

#

that's why ipadapter can never use all details of the image you give it

#

flux Kontext in contrast is putting the image as it is into the diffusion model, so it can use all details of the image

#

(one reason why ipadapter is no doing this, besides its computational more demanding, is that if you give the diffusion model all details of an image, risk is high that the diffusion model just ouputs the very same image. Like you say: give me a similar image like this and the diffusion model just outputs the image itself. Flux Kontext works because it was trained very long on different tasks, very likely MUCH longer than ipadapter

dawn mulch
#

Controlnet definitely has more fine control especially compared to ip adapter as far as I know and I assume flex context dev doesn't provide that level of control yet.

chilly storm
abstract quarry
#

control nets and kontext work differently

#

Flux Depth(and flux-fill, same both technique) place the control image and the generated image on top of each other. So each patch in the generation image gets the exact corresponding patch from the control image

#

so this makes it easier to transfer shapes, depth maps, contour lines and so on. It's also more compute efficient than Kontext

#

in Kontext control image and generated image are placed "next to each other". This makes it easier to change the generated image and more difficult to keep it exactly the same.
Nevertheless, you could train flux kontext on depth maps and it will probably perform almost as good as Flux depth. It's just: why would you want to do that, Flux kontext is slower than Flux Depth.

dawn mulch
#

I mainly use IP adapter for face transfer and flex kontext dev seems better for it so ip adapter is going to the basement for a good while.

abstract quarry
#

flux Kontext can also be finetuned

#

as to be expected, many people currently waste their compute on unclothing finetunes and stuff

#

bit I think/hope we will see cool and interesting finetunes of Kontext in the future

chilly storm
low tulip
#

How doees flux kontext decide the output image size if i didnt set it? it doesn't necessarily seem to match the input

abstract quarry
#

the same way as the base model

#

if you use comfyui, the default workflow is to use the conditioning image as input for img2img. just change the workflow to use an empty latent image in arbitrary resolution

pale wedge
chilly storm
inner sedge
#

Hey guys, I've been doing some SDXL in diffusers as my first AI project. If you wouldn't mind I'd love some input on my code! I have a local branch trying to configure the 2x latent upscaler instead of the beefy 4x
https://github.com/luckielordie/sdxl

river spade
#

Anyone train on Massed Compute?

#

just wondering if anyone uses it

quiet finch
#

yeah

outer patio
#

Holoi

#

Ambeeer here

#

Im looking someone to train a aai with my image [paied{

unreal vortex
#

Hello @outer patio

#

Iā€™d love to help you train an AI model using your image.

#

I have hands-on experience with custom AI training pipelines such as DreamBooth, LoRA, and Stable Diffusion, and I can generate high-quality, personalized outputs whether you're looking for realistic portraits, stylized art, avatars, or any other use case.

outer patio
#

Just dmed u

oblique agate
sleek fable
#

anybody here using Microsoft Designer (Dall-E 3 via Bing Chat?)

outer patio
#

Pwo

rocky rivet
#

Who sell ig followers

outer patio
#

If that works

#

Hahah

floral umbra
#

Wish me luck. Middle of installing rocm via distrobox to run comfyui on steam deck xD

#

Hell, maybe even attempt with the highly unstable zluda lol

floral umbra
#

Rip, with rocm in distrobox, it got as far as loading model, and halted on ksampler HAhaa

warm junco
floral umbra
wet grotto
#

guys, how much gb is flux kontext?

#

and Flux or black forest labs have discord?

tender minnow
#

is there a way to mask images just using prompts? Is there any extension i can download for that?

unreal vortex
faint timber
pale wedge
manic aspen
#

hello newbie here catlurk

oblique elk
tender minnow
#

is there a way to view the previous results? isnt it saved in a different folder? does anyone know where it is?

low tulip
#

Maaaaan i can't get good results with Flux Kontext (using their api for their pro model even)

#

I ask it to change the background of a photo of a person, and the person definitely looks slightly different

#

Disappointing...

low tulip
#

Might have to go back to LoRA training.. i was hoping Kontext could be a hacky alternative to using LoRAs

coarse flicker
dusky thunder
#

hi

sacred verge
#

hi

abstract quarry
errant bronze
#

Hi, all.
If you have a project in mind or if you're just exploring potential enhancements to your website or web application, I'd love to chat with you as a full stack developer.

latent bolt
#

@stanle

tropic igloo
#

hello gamers

outer patio
#

Hoooi looking to someone to help me with some generations

#

Paied of course and willing to ask for more if its works uwu

#

šŸ—£ļø

#

One answered me but just ghostee dhdhd

lavish meteor
#

does stable diffusion have the capability to keep an images overall composition and such but update the colors/rendering to improved styles?

dusky thunder
#

HI

sly peak
#

Hey

unreal vortex
#

Gm

still glacier
novel tide
#

Hi

unreal vortex
#

Hi @novel tide

tepid kindle
blazing violet
#

Hello everyone! We need person who understand how stable diffusion works and can help us with workflow. Write for more details, payment is negotiable!

waxen halo
#

Random can somebody please tell me why there are so many tech layoffs.Yet accounting seems to be more stable right now.I know things in the future will eventually be taken by ai but would like to know why software engineering jobs seems to be more impacted by ai then accounting for example.

valid aurora
#

more forgiving*

oblique elk
# waxen halo Random can somebody please tell me why there are so many tech layoffs.Yet accoun...

Would say that the optimization for accounting has been done in the last 20 years will all kind of tech. OCR Document, Fraud Detection, ... The potencial is not very high. The "mass" Jobs which need not much skilled labor are already done by machines. The Workforce is below the line cheaper and better available.

IT on the other side has not been much automized. Still there are enough easy tasks that can be done by ai instead of junior devs. And even junior devs are kind of expensive, so the potencial cost cut is kind of large. Additionally in some countries skilled engineers are rare, so finding another solution is more urgent

lavish meteor
oblique elk
lavish meteor
#

try (as in to try and upgrade the rendering/style etc) on some of my design

oblique elk
#

You could post it in general with images maybe someone will pick it up

lavish meteor
#

I posted a thread to "community projects"

ebon shore
#

Hi can someone help me choose a model for generation of creatures?

bright ember
#

A curiosity when creating the Lora. Is it better to use images of that character without the background (transparent) or with a black and white background?

abstract quarry
#

there is no transparent

#

diffusion vae does not support alpha channel

#

if you have a transparent background it will usually just get black

dusky thunder
#

HELLO

wind tundra
#

hi there!
Ive got the dumbest question in the history of ComfyUI but I got to try.. I had this green dino in my comfyUI that could save workflows and clean cache or Vram. Now I had to re-install the whole thing and its gone.
Anybody knows the name of the plugin/node?

dusky root
thin vessel
#

hello from India

misty berry
#

hello

sleek hound
#

Hey

mild sedge
#

heyyyy

#

which one gave you the best results so far, comfyui, pykaso, or savro?

low tulip
#

Do you guys know if stable diffusion or flux models understand colors in terms of hex codes or HSL or whatever?

bright cargo
#

Hey is anyone using fooocus ai on amd?

dusky root
#

heha

mild sedge
tender prairie
#

Do you know any (free) program to remove people from a vacation photo? I usually use affinity photo but it's too hard to do it by hand

pale beacon
#

Hi, there

peak cypress
#

Hi, there.
I am an AI/ML developer with over 8 years of experience solving real-world challenges across multiple industries like healthcare, law, education and so on.

I specialize in developing AI-powered solutions such as chatbots, AI agents (MCP/Agentic/Voice), Prompt engineering or implementing RAG system, and LLM models (training, deployment, and fine-tuning).

With a deep understanding of both AI/ML and web technologies, I provide end-to-end solutions from conceptualizing AI models to integrating them into practical, scalable applications.
My track record of successful projects across multiple industries ensures that I can deliver high-quality, tailored solutions that meet your specific needs.

Services I Offer:

Automation: I specialize in automating tasks using tools like n8n, Zapier, and Make.com.
NLP: I handle advanced NLP tasks with models such as GPT-4.5, GPT-4o, Claude 3-7 Sonnet, Llama-4, Gemini2.5, Mistral, and Mixtral.
Model Deployment: I assist with the seamless deployment of machine learning models across various platforms.
TTS / STT: I implement both TTS and STT solutions for interactive and conversational AI experiences.
AI Agents & Chatbots: I develop custom AI agents, Agentic AI, chatbots, and VoiceFlow applications for diverse business needs.

** Check out my portfolio in my discord profile**

I always try to learn new and cutting-edge technologies, and I place great importance on collaboration with team members in development.
If you have an innovative project idea, feel free to reach out. Letā€™s bring your vision to life!
Thanks.

round siren
#

Does anyone know how to automatically get the pictures for models and loras on Forge? On Sdnext there is a button to press that likes goes on websites and autodownloads pictures etc. Wondering if it can be found on the other sd types

split haven
#

Hey, I am a ai automation developer if you need any help or service just drop me a msg @split haven

lean field
#

hey guys im superr stuck, i successfully saved 2 workflows. I'm trying to save my 3rd but... its not saving to my workflow folder, i am clicking save as, and when i try to save again it is complaining that the file already exists

valid aurora
#

what benefits does comfyui nodes have over forge/automatic1111 ? is it worth to learn it even tho its harder ?

pale wedge
pale wedge
tame thorn
#

Excuse me, where should i text prompt to generate image?

exotic condor
#

hello!

sullen reef
#

https://crowdllm.ct.ws I found this tool to crowd build datasets for LLMs the concept is interesting. you put you create a task and people can contribute by adding prompt answer pairs, just putting prompts or just answering prompts without answer.

low tulip
#

Which model do you guys think is best for "X artist style of Y" like "Picasso style cat"? SDXL, SD1.5, Flux

west lark
#

I would try SDXL checkpoints like Juggernaut XL or Realities Edge XL, Dreamshaper isn't bad either.

devout remnant
#

hello

#

I'm sorry im dumb can som one walk me through start up

upper plinth
upper plinth
#

there are certain workflows that are only available through comfy

#

and most of the newer stuff is going to be comfy-only

#

100% recommend learning comfy if you truly want to reap the benefits of AI