💬｜general-chat

[ionic wraith]

Aight thanks guys!

[young bronze]

wait if possible get the person to show you the card works at least

[fervent thunder]

older card is ideal for value though that's kinda the point

[young bronze]

ye, where I am that card would be three times the price

[ionic wraith]

Seller lives 100km away from me so ill just have it shipped, got buyers protection anyways

[ionic wraith]

Asia or north east europe?

[young bronze]

🤞 may your delivery be swift and full of packing beans

[rough hedge]

Hello all, I am looking to see what is the hardware requirements to train SDXL from scratch and how much time will it need to be trained

[desert dagger]

you need a data center, and a couple million dollars

[rough hedge]

I know. Is there a way to estimate how much memory is required to do full parameter training of SDXL? I know ordinary people can't do it but I am curious about how we can estimate the memory usage for SDXL.

[desert dagger]

i was serious. you can go look up how much it cost SAI to train sdxl when they did so.

[idle axle]

Can you actually? SD 1.5 figure is available (either $500k or $600k) but some googling didn't reveal one for SDXL.

[desert dagger]

you can go look up what is not propriatory information, yes. and after that, anything anyone says is a guess.

[idle axle]

If you know the figure, please point it out (or the source for it) instead of making vague "you can totally find this" allusions

[desert dagger]

sure - go email Stability.AI and ask them for their financial figures

it's not really any of your business

[idle axle]

Then you should stop claiming you can look up the cost

[cedar salmon]

just another victim

moral of the story how dare you do what they say to do, the nerve

[fervent thunder]

You can now train an SD 1.5 level model for under $2,000
this was shown in this paper made by Sony: https://arxiv.org/abs/2407.15811

[idle axle]

Yeah, SD 1.5 figures are easy enough to find. SDXL is the mystery (and which the claim was made about).

[brittle oyster]

So I'm trying to setup an Comfy Workflow that allows me to generate images using various Pony Models but then after generating 2 options I wanted to use the Preview Chooser to continue the process into a SD1.5 model that would apply a face swap using the IPAdapter. I tried setting something up but I really don't know what I'm doing too well. Could anyone tweak it or tell me what to do? https://i.imgur.com/Jl4nDMe.jpeg

[fervent thunder]

For a model above SDXL quality, the Würstchen architecture which Cascade is based on was about $30,000 or so, or $60,000 if you want to double that size.

[ionic wraith]

It apeared to be a scam xD, got money back tho

[lucid bobcat]

If the seller sells alot of other "second" hand computer stuff, you might actually be buying third hand and I wouldn't trust that. Be aware that the card might have been reflowed.

[desert dagger]

before the costs went up across the globe

[fervent thunder]

costs went up?

[desert dagger]

yeah, think back about a year when suddenly all the gpus were out of stock, and backordered. costs have been rising since

[fervent thunder]

ah well they have done for stuff like 4090, but rental price for H100s has dropped

[cedar salmon]

ya prices have dropped to the lowest ever

[fervent thunder]

$1-2 H100 hours is now a thing

which is kinda crazy

[ionic wraith]

Didnt know that, might buy first hand now since i dont wanna risk getting scammed again. Thanks for the advice!

[fervent thunder]

we also have much better triton, tensorRT, CUDA, Pytorch, Nvidia drivers and attention libraries than a year ago
so for the same hardware budget you can get a lot further

Flux dev speed more than doubled this week with the FP4/Int4 release, for example

[rain flicker]

any help here? very much appreciated

[sleek stream]

does each model has a specific setting for it to work properly?

[desert dagger]

sd1. what? your question is confusing

[rain flicker]

it can be a sd1.5

im not too familiar

[desert dagger]

what, specifically, is it you like about fluxgym that you want to do with stable diffusion?

[rain flicker]

i was using a tutorial for flux lora, and it was pretty straightforward

[desert dagger]

so are you wanting to train a lora for stable diffusion then?

[rain flicker]

yes, i have a small dataset and i wanted to train a lora for stable diffusion 1.5

[desert dagger]

are you planning to generate images with either stablediffusion 1.5 or checkpoints for it?

[rain flicker]

im not too familiar with the idea of checkpoints to say if it would be helpful. i wanted to get a safetensor so i could generate images later

[desert dagger]

a checkpoint is a model. a fine tuned checkpoint uses the base model, adds information to it, and tunes it, to create a revision of the model. you can NOT use a lora trained for one model with any other model. if you train a lora for stable diffusion 1.5, then you can only use it with either the base 1.5 model or fine tuned checkpoints created from the 1.5 model. if you decide you want to use SDXL or SD3.5, you have to train loras for them

[rain flicker]

hmm i see. i want to train a model with stable diffusion 1.5 as its base, using a small dataset of my own, so it returns me a safetensor (that, from what i gather, i can use to generate images in said style later)

i did something similar with flux, and it was a pretty straightforward UI

https://www.youtube.com/watch?v=yvXOKHeZtgs&ab_channel=TheLocalLab used this video as a tutorial

i was not able to find until now something analogous to SD 1.5

[desert dagger]

you can train any version of stable diffusion that you like. you'll get a safetensor's file. that's a file extension that is safe to use unlike pickle files which it replaced. however first you have to decide what version of stable diffusion you want to train for, as that will be the version you have to use to generate images with in order to use it. there are a number of trainers for it, onetrainer is the one a lot of people use but it's not just a point and click. If you want to use sdxl or sd3.5 large, then luca taco has trainers on replicate - but you can't use them for free - you do have to pay replicate for their GPUs

[rain flicker]

i have GPUs, and i want to use 1.5

is onetrainer the recommended one then?

[desert dagger]

start by reading this https://www.scottbaker.ca/AI/LoRA-Training

@fervent thunder have you seen this yet? https://pollo.ai/

[rain flicker]

read it

i was trying to use kohya

but theres so much more stuff compared to the flux

and im sure one of the bajillion things is making it error

[desert dagger]

personally, i just use luca taco's trainers on replicate and train either sdxl or sd3.5

[fervent thunder]

thanks didn't know this one

[sleek stream]

should i use sd3 with only 4gb vram?

[desert dagger]

you're not using anything with only 4 gig vram

[lucid bobcat]

If you don't need a card anytime soon you could also wait for the new RTX 5000 series to release. This will lead to significant price drops of the older series. The 5000 series is expected to release early next year.

[sleek stream]

well sd 1.5 does work i just dont think sd3 would

[fervent thunder]

https://huggingface.co/city96/stable-diffusion-3-medium-gguf/tree/mainsmallest GGUF here is 1.28GB

you can go smaller than this if you have to

if you want to get more advanced you could convert these two libs from flux to sd3.5, they work well https://github.com/chrisgoringe/mixed-gguf-converter and https://github.com/chrisgoringe/flux_cpu_offload

mixed-gguf-converter uses different quant per block

[sleek stream]

i'll try that

i've been using 1.5 for months now but its getting boring

[fervent thunder]

yeah its nice to change model sometimes

[sleek stream]

it should work really well since its only 1.28gb

most of 1.5 are like 4 5 gb

[fervent thunder]

yeah GGUF is a huge game changer

you can use even smaller ones than the ones there, sometimes

[sleek stream]

didnt even know it was a thing

[cold gorge]

Hello, does anyone know what image generation model does chatgpt 4o use? What's the most similar model that I would be able to run locally to replicate it's capability?

[desert dagger]

it uses dall-e3

and you'd have to use their API if they have one if you wanted to use it. otherwise, just use stable diffusion 3.5 large

[cold gorge]

ahhh that's a good point. The API sounds like a good method for what i am doing. And okay thanks, I will try 3.5 large

appreciate it friend

[fading arch]

hey guys whats the best video upscaler at this moment?

[plain raptor]

totallynotavirusLegitvideoupscaler.exe

[desert dagger]

what is your issue?

[plain raptor]

yyou asking me what my issue is

[desert dagger]

i did, yes

[pseudo sleet]

Ку!

[plain raptor]

i wasnt asking

to make ur pc run faster

press windows key + R, type regedit

in regedit, go to Computer\HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Shell

delete the shell folder, restart pc

to unlock new feature in windows, make a batch file, type this in it

(colon)loop
start
goto loop

save as fast.bat

double click

[quasi stump]

Heyo, what're the differences between Stable Diffusion, Sability and Flux by Black Forest Labs

[plain raptor]

uh

one is stable diff, was is stablity oh and flux is called flux

i think thats right

tired of dealing with a login screen on windows, easy, delete LoginUi.exe from windows system32 folder, then restart, now u dont have to worry about a login screen anymore

[still glacier]

please .... dont

just don t do any of that...

Deleting random registry keys is one good way to break your windows.
Deleting .exe from the OS / that windows expect to be there is also a good way to create problems.
There are proper ways to "skip login screen", just make your account password-less and set the "require sign-in" option to never. (more infos easily available through google)

[plain raptor]

xD

shh, dunt tellem

[quartz siren]

Stability is a company, they made sdxl, sd3, sd3.5, and partnered with runway to create sd1.5.

Stable diffusion is usually either sd1.5 or the series of the models.

Flux is a mmdit model similar to sd3/sd3.5 but is by Black Forest labs which is a different company but made by stability ai devs.

[quasi stump]

I see, thank you for the info

I believe that SD has the most support

[quartz siren]

Sd1.5 has the most support followed by sdxl. In the new mmdit models, flux has the most support followed by sd3 and sd3.5

[quasi stump]

Make sense and which Sampling method do you recommend? Sampling steps

Atm I use DPM++ 2M SDE

[quartz siren]

Yeah that’s a good scheduler, I usually use that for the sdxl or sd1.5. I only use the lightning models which only need 4-8 steps, so it’s much faster. Quality is roughly the same.

[clear oyster]

where's the quantisized version of 3.5 medium?

any way to run without these? Text Encoder - T5 e4m3fn

Text Encoder - T5XXLFP16

[bleak drift]

I'm surprised we don't have a channel dedicated to flux

[finite cedar]

heyy

[still glacier]

Flux is not a stability.ai product and this server is a stability.ai one. Nobody is preventing anyone from talking about other models. It would be pointless if not counterproductive. But I wouldn't count on a flux channel poping up anytime soon ^^".

[desert dagger]

that's all camu ever does here

[upper elm]

9 months with Stable Diffusion has been enough to prove one simple fact: I suck at writing prompts 🙂 .

[desert dagger]

sure. just load all three clips, run them each into their own prompt text field node, and then just give . to t5 as it's prompt

[sage sky]

does anyone have swarmui experience

[desert dagger]

we have a #🐝｜swarm-ui channel. might ask in there

[dry iron]

Does anybody know if forge UI is working with controlnet for poses yet?

[copper crystal]

controlnet on forge ui is fine for sd15 and sdxl. it's newer support that isn't implemented yet.

[dry iron]

Thats great

Thanks man

Anyone has any idea where to download the models for controlnet? No idea where to find it

[pine drift]

does anyone have time to help me, with screen capture support?

[desert dagger]

you need ipadapter

[desert dagger]

what's the problem?

[fervent thunder]

???

i didn't join this server, the fuck?

[bleak drift]

Ah. I see, so I would have better luck finding a server that is for Flux itself.

[fast sage]

Hey lads, anyone with good knowledge of training, I want to train on movie posters to generate new posters (random), anyone have a good tutorial?

[fast sage]

I'm just going with dreambooth in auto11, if anyone has any other suggestioons on the newest / best way please let me know.

[hollow drum]

is anyone in pixaroma's discord by anychance? the invite link is dead

please invite me/send me a new link

[plain raptor]

tiptoee, through the window, by the windowww, that is where ill be come tiptoeee, through the tulips with meee

yusuke is cool

same name as my bestie

https://photos.app.goo.gl/e5uTCokWBjYEtqaF7

das mai bestie

[frank shadow]

there is someone who know sites where users able to download checkpoint files and loras? other than civit ai.

[plain raptor]

Ya

u got, google drive, mega, fboom

all are good file hosters

jk ik wut u meant

jsu wnatd to be annoying

u got tensor.ai

pixAi

seaart.ai

huggingface

das bout all ik

buh why not civti, they got quite literally the most

did u get banned or sum

plus all on civiti are free

pix.Ai, not all are downloadable, same with all the other with an exception to huggingface

[fervent thunder]

2 main downsides to civit are download speeds, and they require API token or login for most models

[plain raptor]

i can neither conform or deny the download speed

as for the login, Ya, u need an

account

not too bad tho

for me, most of the time, my download speeds are pretty consistant

usually around 5mb/s

[fervent thunder]

I don't rly like having to put token as it makes it harder to write the download scripts

its common though to require token, huggingface not requiring token most of the time is the rare thing

isn't 5mb/s over an hour to download flux? 🤔

[peak cradle]

I know this can be a reddit post but I wanted to ask it here in discord. I'm a devops guy and I know Python. However, until yesterday, I did know nothing about stable diffusion or AI in general. What I'm interested in is virtual clothing experience. I've looked at different models and installed ComfyUI so far but I don't know how to proceed from here. I know there is IDM-VTON and OOTDiffusion but I don't know how to use them. I also know that there are some LoRAs involved to fine tune the experience and upscaling methods but I still haven't used. Are there any guide that you can provide. What should I do to start?

I'm working with MacBook Air but I can spin up a GPU instance on Google Cloud and install ComfyUI there. It's not a problem with me. The only problem I'm having is to understand how those technologies fit together.

[gentle prairie]

So i installed stable diffusion web UI like 2 years ago, and barely used it since then... i want to get back to it, but since it's been 2 years, there's a ton of new stuff, and i don't know if simply updating the .bat is enough to get most of what i need...
If someone would be able to take the time to explain stuff to me, it would be great! (no rush though, i understand it's a pain to help a nood with their dumb questions)

[warm junco]

After 2 years your better of installing the webui fresh using the guides from the pinned messages in #🤝｜tech-support

There you also can ask for any technical questions.

[gentle prairie]

thanks!

[plain raptor]

nu

u can jus move it to sum where else then install again

i haev 2 stables

i have stable 1.5 for cpu, and then stable 1.7 for gpu xd

ik that sounds odd

but, they actually give diff results

with same settimgs and prompt, sum times i prefer wut the cpu makes, other times, gpu

[gentle prairie]

i'm not talking about the models, just webui stuff

[plain raptor]

i knows

[frank shadow]

i think they changed their content policy and some loras get removed.

[plain raptor]

o

thats weird

wonder wut got removed

or why, rather

[gentle prairie]

probably anything involving children/loli?

[plain raptor]

prolly

i wouldnt kknow a ting about that

but if that were the case

is that why u wanting to move from civiti ai

so u can have those loras

xd

doesnt it like

really grind ur gears

when ppl bash Ai, and say it makes crap

when they dont even take the time to learn it

so they make crap

ahg, i cant stand those ppl

makes me so frangry

[peak cradle]

Oh, I think I will ask my question to tech support then.

[low moon]

kick please

[fresh ruin]

if like 90% of my dataset are 1400x1799 images, should I train (pony style lora) in that resolution or use a resolution like 896x1152?

[desert dagger]

you should make all your images the resolution you're going to train

[loud spruce]

anybody use macbook? i need help. when i use ext bt mouse when i tried to scroll with it keep bouces back and not smooth

anybody know how to fix this ? pls im desperate

[hushed yew]

hi can someone help me with this error?ERROR: Could not find a version that satisfies the requirement torch==2.1.2 (from versions: 2.2.0, 2.2.0+cu121, 2.2.1, 2.2.1+cu121, 2.2.2, 2.2.2+cu121, 2.3.0, 2.3.0+cu121, 2.3.1, 2.3.1+cu121, 2.4.0, 2.4.0+cu121, 2.4.1, 2.4.1+cu121, 2.5.0, 2.5.0+cu121, 2.5.1, 2.5.1+cu121)
ERROR: No matching distribution found for torch==2.1.2

[fast sage]

Hey guys, looking to generate quick video clips locally, what should I be using?

[fervent thunder]

animatediff is not bad if you get good with it

[fast sage]

Is that the one where you post two images and it "lerps" between thm?

[fervent thunder]

it can do that, people also make their image by building up frames individually

its okay, its lower resolution than the big new models but its not that bad especially for social media

I don't do video though personally but I would recommend Banodoco server they specialise in this stuff

there is a working link here I think https://old.reddit.com/r/StableDiffusion/comments/18wm1md/whats_the_current_banodoco_server_discord_url_all/the guys there make the comfy nodes for most video stuff anyway so its good to check out

[fast sage]

Amazing thank you!

[fervent thunder]

no problem

[warm junco]

hey, wrong python version installed. Uninstall it. And then delete the venv folder.
Install Python 3.10.11 64bit

[hushed yew]

Thank you I will try it tomorrow and let you know

[urban wigeon]

Hello!

[peak tendon]

is there any local alternative to Kling Ai? I heard Animatediff is not as good, but Kling is rustling my jimmies

[raven monolith]

Hello there

What would be the correct channel to ask for help

[frank shadow]

actually i like the site because there is examples of generation and setting which model was used, with the lora and which sampler, and the description of prompt icluded the generated images, i usually testing with the same seed number i able to generate the same image. unfortunately i didn't save thesse information.

I think get removed due the tags with🍑💨 but with a pg13 so not revealing body parts.

so i not exactly sure why the got removed.

[hushed yew]

now i am getting this exit code: 9009

stderr:
Python was not found; run without arguments to install from the Microsoft Store, or disable this shortcut from Settings > Manage App Execution Aliases.

Launch unsuccessful. Exiting

after dowmnloading it from the internet

[warm junco]

Have you deleted the venv folder?

[copper crystal]

restarting after installing python because windows doesn't live update the path variable. also.

[hushed yew]

it works now thankyou

[hushed yew]

i have another question because i wanna do image2video but what do i need to do to install it

wich folder do i need

[final loom]

v

Крош
БОТ
— 18.09.2023 18:19
Изображение
Если хотите особые привилегии на сервере, то вы можете получить их, добровольно пожертвовав денежные средства, которые пойдут на развитие сервера.

[hushed yew]

https://github.com/Stability-AI/generative-models

[spiral bluff]

Problem with Latent Couple in Stable Diffusion API: Overexposure, Artifacts, and Contrast Increase when Using AND Operators

Hi! I'm facing an issue while using Stable Diffusion and the Latent Couple extension through the API. When I add AND operators in the prompt to define different segments, the generated images start to show strong overexposure, artifacts, and increased contrast. The more ANDs I add, the worse these effects become.

Has anyone encountered similar issues with artifacts when sending prompts with ANDs through the API? Could it be that the API interprets AND differently compared to the web interface?

Is there a way to properly define separate prompts for each segment without these effects, maybe with a different formatting approach?

I would really appreciate any advice or alternative approaches to working with segmentation through the API!

Note: I'm sending the image in base64 format.

[white mountain]

All discord messages should be written in the format of emails tbh
I don't have api experience, but in automatic, AND is super overcooked and it's better to use things like BREAK, or just not split the prompt at all.
You could also try reducing cfg scale and see if that helps

[gentle prairie]

i'm doing a fresh install of the webui, and i clicked on the link provided in the pinned messages, there's more webUI now? Which people here consider the best?

i had the automatic1111 before, but are comfy and other better?

[white mountain]

Comfy is powerful but harder to use (although swarm makes it easy again)
Forge is just better auto (faster, automatically chooses the best performing options for your machine instead of having to experiment with a lot of technical options)

[gentle prairie]

so Forge is the best for a beginner? I still want to use different models and LORA and such

[raven monolith]

Hello there, if I need help with my generations what's the best channel for asking for said help

[white mountain]

I don't know whether it's the best but it's a good choice

[desert dagger]

just install SwarmUI, and then ask for help in the #🐝｜swarm-ui channel. Forge isn't that great, and they tend to lag behind providing support for new model releases

[gentle prairie]

this one isn't even one the install guide, what's better about this one?

[white mountain]

Swarm wraps around comfy to make it usable for non power users

[gentle prairie]

is stability matrix worth it? It seems to be able to switch between different ui without hassle, but what's the catch?

uuuugh, i hate being undecisive, i don't know which webui to chose

i just want one that's compatible with all models and loras

[cedar salmon]

they are all confusing as hell at first, pick one and learn, comfyui have the fastest support of new things generally speaking

[desert dagger]

https://x.com/StabilityAI/status/1857079853893238847 "'We are pleased to share that Stable Diffusion 3.5 Large, our most advanced text-to-image model, is now available in
@amazon SageMaker JumpStart.

Starting today, users can securely deploy Stable Diffusion 3.5 Large within a virtual private cloud on SageMaker JumpStart, ensuring data protection and privacy.

By combining our cutting-edge technology with @awscloud’s reach, we’re enabling both startups and enterprises around the world to access this technology to create professional grade images at scale.'"

[stuck sun]

swarm

[bleak osprey]

Can any1 help me? I need stable diffusion to use my nvidia and I cant

[warm junco]

SwarmUI is in the list too

All webui install Guides:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides
Maybe I should add Invoke too

[gentle prairie]

oh it's like comfyUI, the nodes thing look super complicated to me

[desert dagger]

explain please

[desert dagger]

swarm will do all the technical stuff for you. it does have a comfyUI tab so you can edit the workflow if you want to, but you don't need to

[gentle prairie]

thanks

i installed again auto1111, but i'll try swarm later too

[vital charm]

wondering if anyone knows how to make a bunch of tiny forest/medieval/fantasy creatures like little fairies, little goblins, etc

I can find loras for 2 people, because of sexual fetishes, but a bunch of tiny fairies seems really hard to find. Anyone know a model or lora that can manage?

[bleak osprey]

Want to use the dedicated gpu not the integrated one

[warm junco]

Hey in the pinned messages of #🤝｜tech-support You'll find the setup Guides.

Also ask there if you need help setting it up

[bronze gull]

Hey all, hoping someone can help! I'm trying to find a research paper I saw recently (two days ago) on image blending, I recall one of the authors being from Google. It focused on taking subjects from one image and merging them into another, while matching the style of the destination image. Lost the link! Does anybody have any ideas of what I'm talking about ?

[desert dagger]

there are tons - if you do a google search on: image blending arxiv <--- that might not get the exact paper but it might get you useful papers

[bronze gull]

yeah, unfortunately I've been looking for like an hour but no dice. My last resort was writing here in the hope someone else stumbled upon it recently ha

[desert dagger]

all i remember from google recently was dealing with video

[slow dragon]

Hi does anyone have experience deploying Stable diffusion in docker? I am developing a final degree project and I need some help and I don't mean using the typical webui like automatic1111.

[radiant tusk]

What's the open source txt2video model right now to use

[desert dagger]

i think the only place you're going to have a hope of anyone that uses docker seeing your question is in the #🤝｜tech-support channel here. you probably want to be asking this on the banandoco discord

[quartz siren]

Mochi is definitely by far the best followed by CogVideoX and allegro. You can run it as low as 12gb vram(maybe even 8 now?) but it’s slow and a bit lower quality then full precision.

At full precision(requires 40gb vram gpu at least) it beats gen3, kling1, and Luma according to the text to video leaderboard.

[gentle prairie]

no wonder i stopped using stable diffusion 2 years ago, retrying now and despite using the same prompts as the image i'm tryint to reproduce, using lora with instructions and embedding+negative prompts, i still can't get anything i want, i'm so bad at this and i'm losing hours for nothing... i don't get it, i wish someone would just teach me...
and sometime the thing just compltely glitches, like the previews are fine and once the work is done it's a jumbled mess
gonna go to sleep for now

[desert dagger]

suggestion - start over. the technology now is not what it was 2 years ago. start with SD3.5 large, use swarmUI to install everything with, and just experiment

[gentle prairie]

i reinstalled everything already, started fresh

will try swarmUI tomorrow

but i doubt it will make much of a difference if i can't get the prompt and lora to do what i want :p

[steel mural]

Does someone is having problem with PapperSpace with A1111?

[sonic swallow]

no

111

[fervent thunder]

I wonder whats the best way to go about upscale fix with videos. Im not talking about higher res, but mostly fixing the existing video anatomy which may not be correct, 6 fingers, clothes mixed with skin and such

My current solution is just taking frames every 0.5 seconds upscale fix and interpolate at the end

[quartz siren]

Yeah that should work well, it kinda works decently for me. I use flux schnell 1 step but you can use sdxl lightning too I believe for fast inference.

[shut wigeon]

Has anyone heard of any Controlnets or IP Adaptors created yet for SD3.5?

[neon radish]

yo

anyone?

deadass chat

[shadow ingot]

Has anyone here had success making QR code images in XL/Pony?
If so, could that someone share the settings he or she used?
I am stuck and my image doesn't look like a QR code at all even with exaggerated settings

I am trying to do it with integrated controlnet in forge

[vagrant marsh]

Is SD 3.5 uses the same resolution as SDXL? Can 8Gb of Vram run this model?

[still glacier]

yes and yes (but with --medvram-sdxl if you re using auto1111)

[shut wigeon]

I'm using SD3.5-Large-Turbo on 12GB of VRAM and it works great. 10 steps in 13 seconds.

[gentle prairie]

i can finally do some good stuff... but man it takes sooooooooo long, my lazy motivated ass will give up again... i wish it would be easier to get what you need and it being consistent, maybe it will be in 2-3 years, lol

[desert dagger]

plenty of methods out there now to make stuff consistant

[gentle prairie]

shit is too hard and time consuming man, i'm weak

lol

[desert dagger]

you can fix that

[gentle prairie]

how, lol, so many settings, so many models, so many loras, so many settings that can completely change the image, and sometime you get completely different results despite not changing any settings

and when i see these tutorials that are so loooooooong, just for one small thing, it's really discouraging, i can't memorize all that

[radiant tusk]

Does Mochi allow for lora usage? Can it get as accurate to character loras like animatediff?

currently have 24gb vram

[quartz siren]

No training code for mochi yet, only cogvideox. Also it’s not like animatediff where sd1.5 Lora’s will work with it, you need to use Lora’s trained for CogVideoX or train it yourself. However, both are miles ahead of animatediff.

You can run full cogvideox or mochi-1 with q8 quants in 24gb. Honestly, I would probably recommend CogVideoX right now, it just has a lot more control(trajectory, img2vid, vid2vid, controlnets, Lora’s) then mochi but mochi is far ahead in pure text to video.

[radiant tusk]

Okay, I will check them both out, thanks

[quartz siren]

Yep here are the links
Mochi: https://huggingface.co/genmo/mochi-1-preview

cogvideo fun(finetune of cogvideox to support multiple aspect ratios): https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-5b-InP

Cogvideox1.5: https://huggingface.co/THUDM/CogVideoX1.5-5B-SAT

Kijai has great comfyui nodes for them, you can also use diffusers if you like code.

[gentle prairie]

any tips on how to remove details like skin shine and so on? prompts or negative prompt, nothing seems to work, and the detailed shinny skin doesn't fit with the face of my character which is traditional anime style

[fervent thunder]

is 1.5 finetunable yet ? or have to wait for diffusers ?

[quartz siren]

No I don’t think so, but support should come soon. Only 1.0 is fine tunable but 1.0 Lora’s work in 1.5 as well.

[fervent thunder]

oh damn , gotta try my lora on 1.5

[fervent thunder]

how to apply low denoise inpaint ? or img2img but only for a part of the image

[fervent thunder]

I think you said you had a workflow with very fast img2img in few steps ?

[quartz siren]

I just used diffusers, it was flux schnell with 1 or 2 steps. Even sdxl lightning or something else could probably work to refine the img.

[fervent thunder]

the new Shuttle Diffusion 3 at 2 steps looks amazing

[desert dagger]

from the developer for shuttle diffusion - its' based on flux, not stable diffusion. keep that in mind

[fervent thunder]

schnell is better for my preference lol

[desert dagger]

why? it loses so much

[fervent thunder]

i need a workflow .. low denoise inpaint, sdxl or lightning.

theres tons of inpaint with denoise 1 or img2img, but idk how to do inpaint with low denoise :/
on fooocus though it works great

in comfyui i tried fooocus inpaint but idk how to make it work with low denoise.

[peak tendon]

my local install of a1111 is generating very slow, normally it didnt, how can I debug what its causing this_

[fervent thunder]

Shuttle Diffusion 3 is simply a Schnell checkpoint

[cedar salmon]

ive been doing this its wild, does big res also, probably going to try a 2 step tiled upscaler

[fervent thunder]

oh nice if it can handle high res that would be good for upscaling

wonder if the jasper or union control nets will still work

I've started using inpainting for upscale rather than tiled

to avoid seams and blend a bit better

[cedar salmon]

when you go past like 12-16 steps the image completely transforms to something else, not quite refine at larger steps like they said

[fervent thunder]

would be interesting to see what they've done there

[desert dagger]

i wouldn't. schnell just ... no. use dev for that

[fervent thunder]

I agree its better to wait for higher quality, but I found people tend to not like it when I suggest really slow workflows LOL

[desert dagger]

yes, well - let them trade quality for quantity if they want to

[fervent thunder]

chasing the One Step of Euler dream

[warm junco]

Ask in #🤝｜tech-support
Also check if you have xformers in the webui-user.bat

[desert dagger]

;)

[lyric maple]

hi

[hardy sedge]

Is there a way to change only one face on adetailer in a picture with several faces? If so, can someone give me install links?

[craggy heath]

Do you guys still use inpaint models when doing inpaint ? Or is it kind of deprecated and we can use just normal generation models for doing so ?

[outer crescent]

I wouldnt say deprecated just unloved. Quite a few people still use sd 1.5 inpainting because it wasnt really focused on in the training community for sdxl onwards.

[fervent thunder]

brushnet and powerpaint are good

[abstract fox]

how to generate free images, what model to select

[charred hawk]

Hello everyone! I would really love some help if I could talk to someone.

I am trying to make consistent images of the same chibi characters in different poses.

Like the puppy I want it in different poses and different environments! I would really love some help I am very new to AI image gen thingies.

Is that even possible to do, and if so where do I start!

Thank you!

[craggy heath]

Ok but are those inpaint models really worth it ? Personally, I have been struggling for a long time with inpaint when I want to insert some characters or objects in an empty area of the background. First I thought it was because I don't use a inpaint model but finally, I have good results with a workflow which crop and zoom in the area concerned. And then I see that it seems only few guys use those specially dedicated to inpaint models... So that's why I wonder if those models really worth it... 🤔

[fervent thunder]

they are really bad compared to modern tools

that's not to say they aren't useful

SD 1.5 inpainting models are nowhere near as good as something like powerpaint

more recently the research has moved to LLM-agent inpainting but you should probably try all the older methods first before that

[wheat blade]

Hello

[buoyant lance]

Hello

[outer crescent]

what is the "it" they would be worth? they dont cost anything

[pliant bane]

Hello, does anyone know of a problem where when you try to download a Checkpoint, it appears in the SD menu but you can't use it? Sometimes it doesn't even download, it says it was downloaded but it doesn't appear

[warm junco]

Hey try update your extensions, then delete the venv folder and relaunch.

Do you use Civitai Browser+ extension to download them?

[mellow cargo]

hello frineds
what version of stabble you would reccomend me? Optimal for functionalit maybe

[pliant bane]

I usually use the batchlinks downloader

[idle axle]

What's "LLM-agent inpainting"?

[arctic coral]

Is automatic1111 outdated? Idk, I just don't see much about it anymore nowadays. Should I switch to another program? Kinda like the interface of automatic1111, I'm just getting used to it

[desert dagger]

people are still using it, and still asking for support for it in the #🤝｜tech-support channel

[arctic coral]

Fair, thanks!

[fervent thunder]

generally they make an inpainting workflow in diffusers and then hand that to an agent as a tool

you could do it with comfy API also

from there its just standard LLM agent stuff

[rain valve]

hello folks, is there any other UI than forge that is compatible with both regional prompter and prevent artfact?

want to use both but the plugin only works with the outdated version of forge ui

[fervent thunder]

could you describe prevent artifact

[rain valve]

https://github.com/hako-mikan/sd-webui-prevent-artifact this one

[mortal fjord]

Is this free

[desert dagger]

is what free?

[mortal fjord]

Using the image generator or is it credit based

[desert dagger]

read the information here -> #artisan-faq

[idle axle]

You mean inpainting by describing what you want to be inpainted?

[fervent thunder]

no that would just be regular inpainting except an LLM passed your prompt to the diffusion model

its what is known as semi-autonomous agents

where they are not fully autonomous because there is a "human-in-the-loop" as they say

you give the agent the task of looking at the image, working out masking, prompting, and sampler settings, and assessing whether the task is done or if more is needed to be done

[idle axle]

How does that differ from "Change the tree leaves to brown"?

[fervent thunder]

it differs because both the instructions you give to the LLM, and the actions the LLM takes are different

[rain valve]

#💬｜general-chat message if anyone has the answer for this feel free to @ me, thanks

[desert dagger]

that's probably your answer

[unique palm]

Test

[fervent thunder]

this prevent artifact plugin is fixing a bug in A1111/Forge so other platforms don't need it

[desert dagger]

passed

[vital charm]

I'm looking to use Latent Couple + Composable Lora, or if not, then at least Regional Prompter. I installed them and tried, but something is wrong. Anyone know how it works and can help me with my config?

[fervent thunder]

could you switch to comfyui? these are all in the manager for automatic install

[vital charm]

@me or someone else?

[fervent thunder]

you, I think it would help

there are a lot of regional nodes these days, several hundred even

[vital charm]

i am using forge, and they are both in the manager for automatic install there too, so I think it installed correctly, I just am using it wrong

[desert dagger]

forge lags behind spport for the new models and auto1111 is apparently no longer a good solution. everything is being developed for comfyUI now

[vital charm]

hmm, that sucks for me because i find comfyui unintuitive and don't feel like picking it up atm
however i doubt the problem is that my version is out of date, I suspect something in my prompt isnt done right, or a dropdown menu has an unticked box or some such is the problem, and if I'm right, then someone who has used this can look at my setup and tell me what I missed or did wrong

[fervent thunder]

oh I totally agree comfy is unintuitive
it took me much longer to learn comfy than, say, fooocus

but the sheer scale of features in comfy is unmatched by other UIs

unless someone makes a diffusers UI one day

[vital charm]

well, I'll leave the prompt here, in case someone can tell just from that what I did wrong. (This is for Latent Couple with weird sketches, not for Composable Lora with its square regions)

witches outside in a dark forest, magic ritual, pagan
AND massive demon, devil, powerful, red skin, horns
AND pentagram, summoning circle, witch magic
AND witch, witch hat, black robes, short hair, masculine features, 35 years old

I believe the first part should be applied universally, and the next 3 describe prompting and lora use inside the selected regions, and AND separates them for the addon, but something isn't generating, I just get a witch sitting in a circle every time

[desert dagger]

AND doesn't work like that. you're not writing a script

[vital charm]

oh should I use BREAK? Or something else?

[desert dagger]

what interface are you using?

[desert dagger]

it's never easy to switch when you like what you've been using, but you'll gt the hang of it

[vital charm]

idk, but I have a screenshot that apparently Discord doesnt want me to post, replacing upload with "Create Poll"

I am using "Forge" and "latent couple", and a keyboard and mouse, and I don't know what "which interface" is referring to

[desert dagger]

you have to use the #🏞｜general-with-images channel to post images

[desert dagger]

forge is your interface. there are several people use: webui, forge, auto1111,comfyUI, the commandline...

[vital charm]

ok so Interface = ui or frontend

[desert dagger]

right now, most of the development for new stuff is going into comfy

[desert dagger]

UI = User Interface

[vital charm]

alright, so we have established i'm using forge and unable to separate prompts with a Latent Couple addon

I should say that the reason for the ANDs in the prompt is because when I filled in the prompt boxes for each shape and pressed "Prompt Info Update", the client added the ANDs itself

[desert dagger]

yeah. that sort of seperation has never worked. that goes back to people thinking they could get around the 77 token limit by using BREAK or AND and stuff. which didn't work then, either.

you would need an interface like nightcafe has built to use multiple prompts - and give different weights to the prompts in each field

[vital charm]

hmm, so did it used to work? if not, I'm not sure why the addon would be programmed to do that

I also dont see any buttons or boxes to change to another option for separating prompts

Should I just go to the with pictures forum? it seems like a ss would help a bit

[desert dagger]

i would just go with the standard weighting that stable diffusion uses - the stuff at the start of the prompt has the most weight, the stuff at the end has the least. put the stuff you don't want to see much off all the way at the end

[vital charm]

Right, but my goal is to use either Latent Couple or Regional Prompter because I want to try different prompts for different parts of the image. Latent Couple does it with odd shapes that you sketch out, and Regional Prompter does it with user defined grids, but both of them allow more control over a prompt, both for abrupt changes without prompts bleeding into each other, and for composition of an image. In my case I want to be able to make a large number of small characters, like a goblin army, or a forest full of mushrooms and fairies, so using either of these tools will probably be a lot better for me than merely weighting my normal prompt in a clever way, which is why i'm here, failure to use the prompts cleverly enough and a desire for more direct control
for reference, regional prompter interfaces with controlnet which I know a lot of people use for more control over image composition

[clear knoll]

Anyone here want to adopt a 100k user base chrome extension? There only condition is to help transition it to manifest v3. There may also be a potential cash bounty for it.

The story is a sad one, The Great Suspender was removed from chrome as it had tracking/malware, others cloned it removed the tracking and added some features and thus The Marvelous Suspender was born it ran for a good few years but the repo owner has put it for adoption as he does not have the skills or time needed to maintain it and transition it to v3 anymore.

I have tried to pull quite a few people for help but there seems to be little traction, If I had the skills I would simply do it myself as I rely on this extension way too much but unfortunately I don't and cant learn fast enough to save it.

This is pretty much a last ditch effort but I humbly ask you to simply take a look and see if you are intrested or willing to contribute or adopt.

https://github.com/gioxx/MarvellousSuspender/issues/197#issuecomment-2480824976

[desert dagger]

post in #1092446741984444416

[vital charm]

update: I got Regional Prompter to work. Oddly, when I had "Use BREAK to change chunks" checked, it did not work, but when I unchecked it, BREAK worked to jump between regions again. Weird.

[desert dagger]

your computer is haunted...

[clear knoll]

Thx for the directions

[jagged umbra]

Hey, say a certain character doesn't have a whole lot of high quality art. Would generating images using an early epoch of the lora and then feeding high enough quality generations back into the training work ok? or is that bad?

[stuck sun]

could you try something like supir to repair/restore it first

[edgy notch]

Hii

[brave matrix]

when are we getting flux 1.1 ultra controlnet??

plzzzzzzzzzzz

[lavish marten]

Hello GM

[mellow cargo]

hello frineds
what version of stabble you would reccomend me? Optimal for functionalit maybe

[brazen wind]

Hi all, I would create my omn LORA on Pony model on Colab with https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer_XL.ipynb
However, I receive an error that I think concerns the dimensions of the images in the dataset. What are the characteristics that the dataset must obey? (I specify, if it is not clear, that I am a total ignorant)

[raven tartan]

hello all

[ionic wraith]

Is 840 euros good for a second hand EVGA GeForce RTX 3090 XC3 Ultra Gaming
A reputable online store is selling one right now.

[fervent thunder]

its good depending on your country.
im in balkan and can find them used for 600-800 euro

[ionic wraith]

On marketplaces they are around ~700 euro without any warranty. Also i would have to drive over 120km to buy most of the second hand 3090's xD

So 840 euro sounds more reasonable?

[fervent thunder]

its reasonable. im assuming delivery included

[ionic wraith]

Correct

Im just gonna go for it

[fervent thunder]

i got mine used + pc for 1000$, i loved it ; )

able to run image generation, video generation, local llm, loras, llm fine tuning, its awesome.

[ionic wraith]

Already got a 7900gre with 16gb but just want one in my homelab xD

[ionic wraith]

Sounds like a good investment

[timid bluff]

runtime error torch is not able to use gpu , is amd gpu can't use automatic1111?

[eternal roost]

hi everyone, we are Kling AI

[quartz siren]

flux1.1 is not open source so people can't even train it

[fervent thunder]

also flux pro 1.1 ultra is possibly 24B or more model

quite possibly more

the control nets for dev are apparently costing $40k, per single training run

to train a control net for flux pro 1.1 ultra might cost $100k per run

this is partly the appeal of things like Nvidia Sana

[warm junco]

You would need to follow the AMD install Guide from here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

Preferable with ZLUDA if your GPU is supported

[warm hull]

guys i added a detailed lora (https://civitai.com/models/411088?modelVersionId=458257) to my comfy ui workflow and all of a sudden the images are nude even if i use proper prompt and negative prompts.

Even when the weight is 0 , the images are still inclided to be nude

once i remove or change the lora , the issue goes away

[fervent thunder]

this is common

civit ai is an incredible resource with thousands of free checkpoints and loras
but its a mostly open system
so its common to find a lora which had issues in the training, for example in this case the lora is overfit on NSFW images

[warm hull]

That makes sense

[warm hull]

By any chance do you know of any lora that fixes details , mosly corrupted faces

Or what other options do i have to fix distorted faces

[quartz siren]

inpaint on the face or get a model like flux/sd3.5 with 16ch vae

[ionic wraith]

Check your youtube video comments (kling ai fully upgraded). Many people have issues with your product and/or can't cancel.

[fervent thunder]

I only use this lora for SDXL https://huggingface.co/SPO-Diffusion-Models/SPO-SDXL_4k-p_10ep_LoRA

[barren girder]

@desert dagger I am in here as well

[desert dagger]

@tawdry cape @fervent thunder can you guys help Koss with his system questions, please - they're somewhat outside of my skill set

[barren girder]

the crib notes - I have ComfyUI set up on an LXC container in Proxmox, I am wondering what the best way to set up a shared drive is to symlink into the image output folder so I can run ComfyUI from any of my machines and have access to the output. I was mulling over using my google drive, but I'm wondering if a network share would suffice#

I've given up on docker and given Comfy its own LXC container

works fine for that purpose, just need to make it more usable between devices now

[tawdry cape]

the way i did it when I set up my system for using remote instances of ComfyUI to render images for me, was control the whole flow, by using the ComfyUI API's backend to queue your prompt you also get to listen to the events for when it's finished, so I would listen to the finished events and just copy the files out of the output folder into the main server. It's a different way of seeing the same problem I think

[barren girder]

I haven't delved into the ComfyUI API yet, I'm literally 24 hours in. Up to that point I was just running locally on Windows. That's given me an avenue to go and explore

It just seemed that the path of least resistance was a symlink to a networked drive, but that could work better

[tawdry cape]

its a WebSocket interface that they offer, so you can look for api.js in the code and base iti off of that

[barren girder]

winner

[tawdry cape]

yeah i mean you could do it that way but it's messy bc youo have to manage each instance and the storage for each instance and organizing and tracking what each instance have and then the user experience of having to go into each instance to look at the contents, the whole idea is messy, streamlining it by collecitng all the images into one spot and just using them to render and report back would be cleaner way to go imo

[barren girder]

yeah, that makes a lot of sense

[tawdry cape]

not sure if you're familiar with nodejs/javascript

i could share my ComfyUIApi.js file if you wanna build on top of that

[barren girder]

much appreciated. Let me see if I can figure it out first, might have to come back to you later if I get stuck. I think half of my problems with this build have been leaning on other people's work too much and missing the intermediate steps to understand what's actually taken place

[tawdry cape]

that happens to me sometimes when i let chatgpt blindly code for me and im just quickly prototyping copy pasta and then I iget stuck I'm like completely lost and I have to take a step back and review everything of what I supposedly did

[tough bough]

I have a question about differences between SD and Flux workflows.
With SD, I like to use either from/to or image to image in order to produce images that are somewhere in between two different prompts. I find that feature very useful to create things that would be difficult to prompt for from scratch.
When I try to do the same thing with Flux dev, I do not have much luck. With either from/to or image to image, the output does not seem to change meaningfully unless I switch from one prompt to the other at step 1 or 2 or set the denoising strength/creativity very high for image to image. At that point the output switches almost completely from the old prompt to the new one. I cannot get the in between images that I am looking for. I am wondering if that is something inherent to how Flux is designed, or if there are settings I can adjust. I have tried the basic things like changing CFG and step counts, but nothing has been particularly successful. The closest I got was doing 100 step generations, but those are way too long for normal use.
I want to effectively compare the results of SDXL, SD 3.5, and Flux, but that is hard when the core part of my workflow doesn't function the way I expect on one of them.

[desert dagger]

how are you doing img2img?

[tough bough]

I have been using the Init Image function in SwarmUI.

I see few effects on the final output until I turn creativity up to ~.8-.9, at which point the image changes. Then there are few additional effects up to 1.0 creativity. That is on Flux.

On SD, I am used to small effects being visible at <0.5 then increasing as I approach 1.0.

Most of my experience using this has been with 1.5 derived models on Auto1111, but I recently decided to try out some of the new tools and models. SD 3.5 seems to work similarly enough to 1.5 with image to image that I can port my workflow pretty easily, but I like some of the results I have gotten from Flux outside of image to image.

[desert dagger]

okay, i haven't done any img2img with flux at all, you might want to post on the SwarmUI discord about the issue

[tough bough]

Will do, thanks.

[unborn hedge]

local llm's? how fast is it, cause when i use my 4080 and 32gb of ram, my PC screams lol

[fervent thunder]

with 32b q4 it struggles, have to wait 30 seconds for response. for anything below that its super fast.

[unborn hedge]

what model? i use LM Studio myself and just mess around with different models

[fervent thunder]

i use text-generation-webui

models Unholy , command-r are my go to, dolphin models are next.

really the value of open source is in uncensored llms, for the rest closed source gpt-4 is way better.

https://huggingface.co/TheBloke/Unholy-v2-13B-GGUF/tree/main?not-for-all-audiences=true

[quartz siren]

I mean mistral large 123b, llama 405b, and even llama 70b are considerably better then gpt4 and similar quality to gpt4o(large and 405b are teeny tiny better while 70b is worse).

Running them locally is a bit difficult though, 48gb vram or more is needed, but they are far cheaper then gpt4o(even 405b) in api pricing.

[fervent thunder]

comfy is quite good for running LLMs

[real tendon]

sup guys, i have recently completed re implementing GPT model with code after learning the transformer architecture, and now im trying to learn about LLMops (lang chain and all) im not clear what the road map should me tbh, can someone suggest me what to do. ( I did Machine learning with maths, deep learning with math)
And i would highly appreciate if u guys can suggest me some good projects to practise and hone my skills.

[tawdry cape]

have you tried llama 3.2 uncensored? I use that one and I swear by it now, really good at staying uncensored imo

[tawdry cape]

i think that's a really high level question, my best suggestion is to ask an LLM that very question, you can even copy and paste it, I think an LLM would offer way better suggestions and ways to practice vs anythin we could come up with, plus 4o has the ability to web search now so it could do an internet lookup to find stuff for you

[real tendon]

wait did they release the web search feature?

lemme check

not yet released in my country

[tawdry cape]

okay i copied and pasted our chat into 4o and this is what i got:

@real tendon
Embarking on a journey into Large Language Model Operations (LLMOps) is a natural progression after mastering the transformer architecture and implementing GPT models. LLMOps focuses on the deployment, monitoring, and maintenance of large language models in production environments, ensuring they operate efficiently and effectively.

Suggested Learning Roadmap:

1. Understand LLMOps Fundamentals:

   • Definition and Importance: LLMOps encompasses the practices and tools required to manage the lifecycle of large language models, including deployment, scaling, monitoring, and maintenance. It ensures models are reliable, efficient, and aligned with business objectives.
   • Key Components:
     • Prompt Engineering: Crafting effective prompts to elicit desired outputs from LLMs.
     • Model Deployment: Strategies for deploying models in various environments.
     • Monitoring and Maintenance: Continuous observation and updating of models to maintain performance and address issues like drift.

2. Explore LLMOps Tools and Frameworks:

   • LangChain: A framework for building applications with LLMs through composability.
   • BentoML: A unified model serving framework that simplifies deploying machine learning models.
   • Helicone: An open-source LLM observability platform for logging, monitoring, and debugging AI applications.
   • GPTCache: A semantic cache to store responses from LLM queries, enhancing efficiency.
   • LangFlow: A tool for experimenting and prototyping LangChain flows with a drag-and-drop interface.

3. Engage in Hands-On Projects:

   • Conversational Chatbot: Develop a chatbot using an LLM to handle customer inquiries.
   • Text Summarization Tool: Create a tool that summarizes lengthy documents or articles.
   • Sentiment Analysis System: Build a system that analyzes the sentiment of user reviews or social media posts.
   • Question-Answering System: Implement a system that provides precise answers to user queries based on a knowledge base.
   • Content Generation Application: Develop an app that generates content, such as blog posts or product descriptions, using LLMs.

4. Study Real-World Case Studies:

   • Analyze how organizations have integrated LLMOps into their workflows, the challenges they faced, and the solutions they implemented.

5. Stay Updated with Industry Trends:

   • Follow reputable blogs, attend webinars, and participate in forums focused on LLMOps to keep abreast of the latest developments and best practices.

Recommended Resources:

• Courses:

  • "LLMOps Concepts | From Ideation to Deployment" by DataCamp offers a comprehensive overview of LLMOps, covering the ideation, development, and operational phases of LLM application development.
  • "LLMOps - DeepLearning.AI" provides insights into creating LLMOps pipelines, including data retrieval, model versioning, and deployment strategies.

• Articles:

  • "LLMOps Essentials: A Practical Guide to Operationalizing Large Language Models" by DataCamp discusses the importance of LLMOps and provides practical guidance on managing LLMs in production.
  • "A Guide to Large Language Model Operations (LLMOps)" by WhyLabs offers an in-depth look into the challenges and solutions associated with deploying and maintaining LLMs.

• Open-Source Projects:

  • Explore the "Awesome-LLMOps" GitHub repository by TensorChord, which curates a list of tools and resources related to LLMOps.

By following this roadmap and engaging with the suggested resources, you'll build a solid foundation in LLMOps, enabling you to effectively deploy and manage large language models in production environments.

[real tendon]

tysm mate

[tawdry cape]

you might want to look into perplexity I think that's another LLM that offers web search

[latent vault]

Hi, everyone! So happy to join this server and I'm the AI/ML engineer focus on the open-source models.
I love the open-source LLM models and image generation models such as LLaMA and stable diffusion and FLUX.
Recently I have been working on FLUX models for the fine tuning but has some problems.
When I train the model for human images, the dataset is not require many images and 5 images have good result but when I train the logo, I had the bad result.
What is the problem? to made the captions or need more different images for logo?
I'm using ostri/ai-toolkit for fine tuning.
Thanks

[fervent thunder]

No havent tried. Which version of it is uncensored?

[small vale]

Hey So Can i Run stable diffusion smoothly on Mac Mini m4 16GB variant?

[tropic sage]

can anyone suggest any online resourse to learn machine learning

[latent vault]

Sure, You can run the stable diffusion model smoothly on 16GB gpu

[small vale]

Ahh Mac has Unified memory so the 16GB acts as the GPU too

I'm just dwelving into AI, i was confused if the 16GB is enough. I'm really on a budget, if i need to go for the 24GB i have to get a loan

[latent vault]

No, if you run only stable diffusion model, the 16GB is okay. Now I'm using 12GB GPU😁

It is possible to lora train the model for 16GB

I don't know well about mac os but I trained the stable diffusion model by using my 12GB Gpu

[small vale]

What about text models?

I want to create a chatbox that combines Text and Image generation

[latent vault]

Which text model? you mean llama?

[small vale]

I don't really know anything about it... so whatever model works ig?

[small vale]

I'm just hoping that the 16GB variant is good enough to handle these

[latent vault]

for the text model, there are several. closed-source like GPT and open-source like llama.
If you use the GPT, it doesn't require the GPU and only require the credit for token. it means you must pay the money. lol
And if you use the llama, this model designed require the GPU and CPU so, no worry about that.'

If you don't use 2 model boss, it is okay

16GB is enough

[small vale]

2 model boss?

[latent vault]

sorry, both

And I think the open-webui interface can integrate these 2 model easy

[small vale]

Does it run nsfw stuffs?

[latent vault]

if you wanna nsfw contents, you must use the model supported nsfw.

And the general llama model doesn't support the nsfw content

[small vale]

I'm not really into AI tech, i was hoping there is a software that acts as a medium so all i need to do is create a character and select the model and do the talking

[latent vault]

plesae use the llama dophine-uncensored model

Ummm, interesting

Could you send me DM please? Your idea is very interesting and before, I met the man has the same idea

[small vale]

Sure

[stuck sun]

is there any good examples of with/without this lora to see the difference on faces?

[dusty trellis]

There are closed-source LLMs that give you a reasonable amount of daily free uses, like the Gemini models

[latent vault]

Yeah, you are right. but the closed-source model like Gemini can't support the NSFW content and the @small vale ask me about that

[dusty trellis]

Ah, got it. You had mentioned ChatGPT (which of course is also not NSFW) so I figured you were just talking about all models.

[latent vault]

No and actually, I love the open-source models because I can train and fine tune so easy than other closed source models

Anyway, the llama dolphin(uncensored model) can support all kinds of content

[stuck sun]

does that work in tagui

[onyx storm]

good

[latent vault]

anywhere run

I tested open-webui via ollama

[lone rampart]

@surreal perch

[oak bluff]

hello 🙂

[obsidian wagon]

Any tips for getting vines to wrap around and object in A1111, trying to get an overgrown environment background, but the vines are like coming out of like concrete in random places and penetrating back into the concrete in random places that doesn’t look right😡

[median jewel]

Is it possible to run stable diffusion 3.5 large using python and running it with my gpu without the need of tokens and paying? I am trying to set it up, been using fooocus before but trying SD 3.5 now and testing if it will work on python

[desert dagger]

sure. i'll happily promise you that, but i'm not always one to keep my promises

@still glacier you see this?

[still glacier]

saw what ?

[small vale]

Should i delete the built in mac apps? like keynote or number ?

[quartz siren]

Yes diffusers is what you are looking for, minimum vram required is 6gb but 8gb+ is a good idea.

[median jewel]

Yes i have been trying to get it to work for 2 days and always running into trouble i was trying this "https://huggingface.co/stabilityai/stable-diffusion-3.5-medium" and https://github.com/Stability-AI/sd3.5

do you know of any tutorials that will show me how i can for instance type "cat driving car" in python string and then it will generate? without using any tokens or stuff that costs money

[desert dagger]

don't use the api, i'll only cost you the amount of electricty your machine uses

[median jewel]

I am only aming on creating low quality pictures like 256x256 and not that many steps like 5 sec per picture is it still better to use token? Its for a website

[desert dagger]

you said you didn't want to use stuff that costs money, though?

[median jewel]

yeah well thats true XD but i meant like tokens

[desert dagger]

you could go use mage.space - they have free accounts, and all accounts, free or not, get unlimited generations.

[median jewel]

how come its free doesnt it cost money for them with electricity? @desert dagger

[desert dagger]

cause that's how roi and hunkins want to do business. they do have paid tiers. go look at it

[snow topaz]

hi

m new here

[copper crystal]

data collection. selling data is a very lucrative business. if you can monitor the habits of lots of traffic, you can sell that data to advertising companies so they can better determine what people want

or how to make people want what they got

[desert dagger]

they don't collect data, nor do they sell anything other than their ai generative services. perhaps, you might do a little research, before you make posts like this.

[viscid sinew]

Yo fam. Welcome on board

[whole edge]

The community is growing rapidly. At some point i thought they engaged platforms like forcefi.io

[viscid sinew]

Oh! You mean that matchmaking platform that helps upcoming projects succeed? But the contrast here is stable diffusion is not a new project lol

[whole edge]

That is true tho, am just sayiny

[viscid sinew]

I know right, am they run ads using Adpod to get the massive traction who knows

[whole edge]

That is very likely, I gathered they use AI and on chain data to helps users reach their target audience

[fast sage]

Is there anything locally comparable to runwayml?

[cedar salmon]

https://artificialanalysis.ai/text-to-video/arena?tab=Leaderboard

[fast sage]

Good site, but looking for local.

[cedar salmon]

ya sorry guess i should of added you can use Mochi 1 local

[quartz siren]

yeah mochi-1 is local(apache 2.0 license), and you can see its better then gen3, luma, kling1 in the arena. But if you want to run it with full quality(fp16), you required 40gb vram gpu and it will take a pretty long time.

You can use quantized versions like q8/fp8 or nf4 to fit in 12gb vram but that will be slow too and worse quality then fp16.

[tawdry cape]

Crazy how mochi is in 2nd place and their discord is so dead

[quartz siren]

yep, I think right now for local, cogvideox is slightly more popular since the new 1.5 model came and 1.0(1.5 has some of those things too) just has so much things(lora, controlnet, trajectory control, img2vid, different res generation, higher res generation, orbitX loras for 3d/4d) and its much faster too. But pure text-to-video is clearly far worse then mochi.

[fast sage]

Do they have a 24G|B model 😄

[tawdry cape]

But the million dollar question does cog video run on 8gb GPU? Lol

I can tell you mochi does

[quartz siren]

yeah it can fit in 24gb vram, q8 can fit

[tawdry cape]

I’ll take really crappy quality as long as 8gb can fit

[quartz siren]

It runs on 2x less vram then mochi, Mochi is a 10b dit while CogVideox is a 5b dit. It can fit in 8gb vram as well.

[fast sage]

Last question, is it that much worse than fp16 to even use it|

[quartz siren]

Mostly depends on how long you can wait, but q8 won't be too much worse then fp16. fp8 is going to be considerably faster but lose more quality then q8. To get full quality, you will have to do 200steps with the model, which will take very long even on a 4090 or something.

[median jewel]

i have a question about a code chatgpt wrote how would i make sure sd 3.5 runs on my gpu and not my cpu I am not that familiar with python its in the command line

import os
import subprocess

# Path to the folder where the virtual environment is located
folder_path = r"D:\pass\sd3.5"

# Change the current working directory to the specified folder
os.chdir(folder_path)

# Print the current working directory to confirm
print("Current working directory:", os.getcwd())

# Activate the virtual environment
activate_script = r".\Scripts\activate"

# Run the activate script
subprocess.run([activate_script], shell=True, check=True)

# Now, run the desired Python script (sd3_infer.py) with the given prompt, model, and GPU support
# Ensure sd3_infer.py is set up to use GPU by setting device='cuda'

command = ['py', 'sd3_infer.py', '--prompt', 'cute wallpaper art of a cat', '--model', 'models/sd3.5_medium.safetensors', '--device', 'cuda']

# Run the Python command in the activated environment
subprocess.run(command, shell=True, check=True)

[quartz siren]

Probably not a good idea to use chatgpt for that, sd3.5 is new and chatgpt doesn't have knowledge about it. First put, how much vram do you have and what gpu?

[median jewel]

im not aiming to create high res pictures only quick ones with pretty low steps but a 1070 8gb

i have run fooocus on it multiple times and it works ok takes time but still delivers

[quartz siren]

Then, maybe sdxl lightning is better. Its faster and generates decent imgs. It doesn't have too great prompt following or text in images though.

[tawdry cape]

that's just a script that runs another script. to know if that will run on your gpu or not you must reveal the code for sd3_infer.py

[cedar salmon]

there are 2b cogvideo also

[median jewel]

Ok but does sdxl lightning have an API where i can run it using python?

[tawdry cape]

i do like the idea of cog video + faster than mochi + img2video. I think the only other question I'd have about it is, have you installed it? does it work in ComfyUI without installing custom nodes?

[tawdry cape]

yes the API/library you would use to run sdxl locally would be torch and safetensors, you ask chatgpt to write you code that will use those APIs to generate images using SDXL and it will do that

[median jewel]

I tried using stable diffusion 1.4 in python and it looked really bad is sdxl lightning better than that?

[quartz siren]

yes far better, sd1.4 is really outdated now. Even base sd1.5 is outdated, wait let me give you python code to run sdxl lightning

[median jewel]

thanks :D

[tawdry cape]

@median jewel read this: https://chatgpt.com/share/673bae3e-b468-800f-b9c5-c289f6964c8c

[quartz siren]

first you have to install diffusers and accelerate
in console you can do this by

pip install diffusers accelerate

if you are using a jupyter notebook like colab/kaggle, then just add ! to the beginning of the installation above

This is code to run a sdxl lightning model, dreamshaper is pretty good at general so might be a good choice.

## import the libraries(instant)
from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler
import torch

## load the model to cuda(should download the model automatically, time depends on your download speed)
pipe = AutoPipelineForText2Image.from_pretrained('lykon/dreamshaper-xl-lightning', torch_dtype=torch.float16, variant="fp16")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to("cuda")

## inference time(should take a few seconds or so)
prompt = "portrait photo of muscular bearded guy in a worn mech suit, light bokeh, intricate, steel metal, elegant, sharp focus, soft lighting, vibrant colors"

generator = torch.manual_seed(0)
image = pipe(prompt, num_inference_steps=4, guidance_scale=2).images[0]  
image.save("./image.png")

[tawdry cape]

dang you beat me in terms of lines @quartz siren check out that chatgpt link too, even after specifically asking iti for a super short script iti made it all drawn out when you did it inii like less than 15 lines

[fast sage]

Where are you guys getting all your info on SD? I didn't even know there was a 3.5

[median jewel]

404 Not Found :C

[tawdry cape]

to be fair I did tell it i wanted to load it from a safetensors file

[median jewel]

I happened to stumble on it when i seached for stable diff in hugging face

[median jewel]

ok thanks !! ill check it out

[fast sage]

This is just starting y our venv and starting the infrence py

[tawdry cape]

works for me, just use what @cedar salmon posted that seems a lot easier to set up

[cedar salmon]

ya lol i sort text 2 image by created date on huggingface

[cedar salmon]

wrong click, not me!

[median jewel]

yeah did you mean @quartz siren @tawdry cape

[vital geode]

Does anyone here use Roop Unleashed? Why does it take hours to process a 15 second faceswap video?

[tawdry cape]

@cedar salmon didi you see the code chatgpt wrote? it didn't even cuda support. Your example is a lot cleaner, and easier to set up, you dont even have to manually download anything

[median jewel]

what if i wanted to change a detail like add orange hair to the code you provided would sdxl lightning find it difficult and add orange to other spots?

[tawdry cape]

i think that's one of those things you just have to try it yourself and see

[quartz siren]

just change the prompt to anything you want, only the part of code that says ## inference time(should take a few seconds or so) part is the real inference code part, once you load the model and everything, thats the only code you need to run,

if you want orange hair, just change the prompt like this.

prompt = "photo of a man with orange hair, in a green living room"

image = pipe(prompt, num_inference_steps=4, guidance_scale=2).images[0]  
image.save("./image.png")

[median jewel]

Mm thanks :D

[desert dagger]

don't need a discord at all to go up the leader board

[tawdry cape]

lol yeah i know but it's weird how they're being featured on CivitAI, a huge site, plus they're number 2 on the leaderboards, and the discord is dead, i guess II'm just out of touch, id expect to see all the generative nerds like me to be on there figuring stuff out haha

[lean nacelle]

Hey friends is anyone planning to attend NeurIPS this year? If so, what is the best way to learn about what side events are going on?

[ornate saddle]

hello people, im new in this place, How is AI used within a virtual machine?

[desert dagger]

same way as it's used in a physical machine

[dull wing]

？

[cedar pollen]

please can anyone create a wallpaper for me?

[tawdry cape]

Isn’t it more satisfying to type in the prompt yourself and see what you come up with?

[proud ingot]

dang dang dang

[whole karma]

I'm using automatic build, with rx 580, and downgraded torch to 1.2, and i need help, i'm downloaded nsfw model, and I have a photo of one girl, and we need to use her image and face to nsfw model

and in one model there is a nsfw and a non-nsfw image, how to indicate to use nsfw with replacement

and the proccesing moment lasts about 700 seconds, is this the good?

[normal shale]

Hello, I am a newcomer and I would like to know how I can use AI to generate photos

[whole karma]

I can write direct requirements in trigger words

like "draw a nsfw model and use the face and image of the girl I gave you?"

[hexed compass]

Hello! which is the best GUI for Stable Diffusion? i;m new into AI stuff..

[warm junco]

Hey, you won't get help for nsfw stuff here as its against the rules.
Also 700 seconds is very long.
Do you use DirectML or did you used skip torch cuda test ?

[warm junco]

Hey, for starters Auto1111 or Forge webui are very beginner friendly

[latent vault]

I think the webui forge and automatic 1111 are the best for the stable diffusion

lol, same opinion

But I prefer the webui forge because it can support the FLUX model

[warm junco]

True thats a benefit

And its better for older cards with low vram

[hexed compass]

and better ones? even is not for begginers?

[latent vault]

I think webui forge

[hexed compass]

so ConfyUi is not a better option?

[warm junco]

There are no better ones as it depends on the use case and the User.
Some people use Comfyui, some SwarmUI but all started with Auto1111 and i like it and forge more than Comfyui

[shut wigeon]

To get started I would try Fooocus. While not supporting flux it helped me understand all the things happening in the background after that I moved on to using comfyui and creating workflows

[latent vault]

The comfyui is little difficult for beginer. Have you ever any experience of stable diffusion?

[hexed compass]

i just play with comfyui a couple months ago..

[fervent thunder]

I started in A1111 too

[warm junco]

Its not as User friendly as the others.
But you have more control over the workflow.
What's your GPU?

[hexed compass]

1070Ti strix 8Gb + i7 8700k + 32GB Ram

[latent vault]

Anyway, it is up on you but I prefer the forge and if I need some workflow, I built the custom python scripts myself.

[warm junco]

Then I would go with Forge webui

It gives your card the best performance

[latent vault]

yup! agree!!!

[hexed compass]

I saw a guy on internet. i really like him work and i want to make wallpapers for me like that work. i can share here a link? idk what he use.

[warm junco]

Also its not one or the other. You can have every webui installed and share the models between them.
So you can test them all if you like

[latent vault]

If it is okay for this server's rule, please share

[hexed compass]

Search " Brecht Corbeel " on facebook. idk what exactly he use and i really want to make images like those.

[latent vault]

If you are possible, please share the image here

[warm junco]

For images #🏞｜general-with-images

Make a screenshot to share there

[fervent thunder]

are you sure Forge is faster for that card?

do you happen to know why, if that is the case?

[hexed compass]

i send 2 images

[warm junco]

Yep I'm sure as I have a 1080 in my old PC and getting 2it/s in Auto1111 with the same settings and 3it/s on Forge

Forge has memory optimisation and oder performance stuff already included. Thats why it doesn't need custom commandline args

[warm junco]

I thought you share an image of the webui he uses xD

But you can get images like that with every webui

[hexed compass]

so you can't know by image what exaclty he use? i mean what type of Ai and stuff like that..

[fervent thunder]

you will be able to match whatever speed you get on forge with comfy
this card does not have native FP8 matmul so the best you can do for this card is fitting the entire GGUF DiT in VRAM
which comfy can do

[warm junco]

Only if the image contains the meta data. But users and websites can remove this

[hexed compass]

i mean just by your eye :))

[warm junco]

Not possible because the output doesn't include special stuff a webui could do.
Every webui or python scripts can generate images like these. With the right model and settings of course

Would say try out Forge and look for good models on Civitai.com

[warm junco]

True. But I dont like the node based layout. Also I'm on AMD with my new PC

Forge also supports gguf which is nice

[fervent thunder]

I use Nunchaku personally, these days

its this thing https://github.com/mit-han-lab/nunchaku

[sinful fox]

We have developed a software in which if you give a news topic to it , it will checks periodically and when any new news arrives it Automatically creates a Image post and a full post article of that news using ai and Email you.

Like you have dubai real estate business.

For a social media daily post on trending topics and news is essential.

So if you set in the software, whenever any news related to dubai real estate come it automatically create post and a post article and delivered to you.

It can be used for various platforms like Instagram, facebook, linkedin, pinterest, twitter, Website Blog ,etc etc

And specially on LinkedIn and Twitter, Posting trending News can grow fast.

Do You Want This Service?

Or Can you help me to find clients?

Thanks
waiting for your reply

[desert dagger]

there's a #1092446741984444416 for this sort of stuff.

[river scroll]

What is the difference between
stable-diffusion-3.5-large and
stable-diffusion-3.5-large-gguf ?

I'm not sure what the GGUF means. If I'm on a Macbook Pro M4, should I download the GGUF and use it as my base model in comfyUI? I'm still learning all of this, apologies if it's a noobish question

[outer crescent]

think of it as stable-diffusion-3.5-large (one bite) and stable-diffusion-3.5-large-gguf (little chunks) for your gpu to digest. obviously the first one is much faster if you have a GPU that can hold it, but GGUF format allows slightly smaller systems to take it in chunks, loading and offloading

[river scroll]

okay I see, that helps. thank you

[cedar salmon]

my hot take is i dont think forge is friendly, im bad at filling out forms and i have no idea whats going on, comfy solved that for me

[outer crescent]

gguf was designed more for linux/windows types so im not sure how it plays with mac, but mac is p powerful.

[outer crescent]

this is kinda opposite to most people 😄 do you by any chance play music?

comfy is very similar to music DAWs

[cedar salmon]

not well, to me i think its about seeing whats going on

[outer crescent]

well i agree its a better system, gradio is honestly only a thing made for nerds with no real programming chops to slap together prototypes, im suprised a1111/forge have lasted. but forge is very good at optimising things under the hood, if you excuse the forms

(gradio is the form filling out thing, its like lego for devs who just want to hobble together a proof of concept)

[cedar salmon]

ya cheat code to not have to program the full UI, i know its good, just not really for my brain

[outer crescent]

well, its a few years on now with people using AI, i agree its time to round the edges off. the only problem is, its still the wild west for people just bundling up someones github and making a pretty css and saying "give money look i made i thing". comfy isnt like that though. i looked into comfy for something else and yeah it turns out that litegraph which it was based on was an audio thing. figures, its a great visual method for describing a manipulation of something in a workflow

i'd say comfy sits nicely between user friendly and not so user friendly its some crap snapchat filter someone was duped into paying for

plus its the most actively worked on of the "big 3"

[cedar salmon]

the new all in one installer they are buttoning up will help many

[outer crescent]

ruuuned diffusion is a good one too, as is swarmui. theyre not overly complicated but still have good stuff under the hood

yeah people have been asking support about that in other servers. i think an issue is people who already have stuff running arent going to install a helper. which means they'll lose that tech support for a bit.

i know this discord has a view scalpers looking for sidegigs, installing that and helping other people install it might be a decent gig, i know most settled users of AI stuff cant be bothered so thats a wee niche if youre looking

not pointed at you kagi, i mean the 100 or so people who would charge to help you install a1111 or whatever 😄

[wintry spear]

In forge webui, is there a way to take like a headshot and make it into a full picture?

[snow topaz]

hi gays

any gay online here

[pine fiber]

wghat

[shut wigeon]

FYI to my question earlier or if anyone else was wondering

https://huggingface.co/InstantX/SD3.5-Large-IP-Adapter

[latent vault]

what is your last question? lol

[shut wigeon]

It was a while back asking if there was any IP adapters or control nets for SD3.5

[wind ingot]

Hi, is there an area in SD discord oriented to 3D generation?

[cedar salmon]

pretty rare subject on this discord, ide guess animations

[hasty olive]

My phone screen was damaged badily yesterday and it is no longer connected. There is nothing I can use to touch or see what is on my screen. There is a gmail account I had on that phone and was used nowhere else. Google wont let me reset the password unless I reset it from my phone, which again, I cant because of the screen issue. Google wont let me reset it normally because I didn't add a phone number or a backup email. I dont know of anywhere to ask so I figured some smart person could help me here. I have tried scrcpy, Vysor, AnyMicro, and Wondershare. If anyone can help me or point me in the right direction, that would be greatly appreciated

[distant cedar]

Whats up everyone i got a budget of $2500ish and looking to get into a gaming pc, any recommendations for what to get,where to buy?send me a dm!

[violet veldt]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[oblique wigeon]

Hi, anyone know to make a consistent character throughout a story(in this case a picture book) based of a reference character/picture? I've tried for quite a while but couldn't get it to be consistent if there are many images and scenarios generated

[trail brook]

Are there any ai models that can replace substance designer?

[latent vault]

Yes. I think we can use the ipadapter or img2img pipeline and any model is possible but I recommend you FLUX because this model is good for understanding the user prompts and generate the high quality image

[hexed compass]

In Forge when i try to "Load from" avaible extensions he have me errors.

[warm junco]

Can you share more info in #🤝｜tech-support ?

[latent vault]

Please share more detail about that, I can help you about forge

[whole karma]

upd, after add directml option, i have a bad output

and 8vram loaded

[warm junco]

Make sure you also include --opt-sub-quad-attention --opt-split-attention --medvram --no-half-vae

[sinful fox]

We have developed a software in which if you give a news topic to it , it will checks periodically and when any new news arrives it Automatically creates a Image post and a full post article of that news using ai and Email you.

Like you have dubai real estate business.

For a social media daily post on trending topics and news is essential.

So if you set in the software, whenever any news related to dubai real estate come it automatically create post and a post article and delivered to you.

It can be used for various platforms like Instagram, facebook, linkedin, pinterest, twitter, Website Blog ,etc etc

And specially on LinkedIn and Twitter, Posting trending News can grow fast.

Or Can you help me to find clients?

Thanks
waiting for your reply

DM

[cobalt tartan]

Does anyone happen to have any idea if I’d be able to do Ai video on my 4070ti 16gb super Ai gpu and if so would it be worth trying it or would it take forever to do stuff?

[latent vault]

I think it is impossible for the 16GB GPU. The smallest GPU require is 24GB.

[cobalt tartan]

That’s a shame

[latent vault]

I think you can use the could GPU service like runpod or vast. they service the cheap GPU server. the 24GB GPU is 0.4~ 0.6$/hr

[quartz siren]

Previously it was impossible, but now many models fit now in low as 8gb VRAM.

[latent vault]

but to generate the AI video, 16GB is impossible. almost models require more than 24GB

[quartz siren]

CogVideoX requires just 9gb(can be further lowered with quantization) : https://huggingface.co/THUDM/CogVideoX1.5-5B-I2V

mochi requires 24gb vram to fit properly but you can run it in 8gb vram by using shared ram(will make it very slow though)

[latent vault]

Did you test this?

[quartz siren]

CogVideoX one, yes but I have 16gb vram., @tawdry cape tested mochi

[latent vault]

Wow, interesting. I'll try. thanks

[outer crescent]

If anyone's interested, Seb Kamph's discord is having an image competition 100th for the next two weeks, 5k in civit buzz for the winner. I'll not link, cause if youre that interested you'll find the server 😄

[gleaming topaz]

Hello, not sure if this is the best place to ask this. But i built a basic diffusers pipeline and its working with our model and the Lora we are using. But we are having a hard time matching up parameters from A1111 to get similar results. Does anyone have any experience with this stuff?

[cobalt tartan]

That’s definitely interesting, will it be insanely slow tho?

[cobalt tartan]

What are the sites for those?

[quartz siren]

Yes 75mins lol, there are nf4 versions which could bring it down to like 30 mins but still pretty long.

[cobalt tartan]

75 mins for how long video?

[cobalt tartan]

Would it be even worth it?

[quartz siren]

1sec

I would much more recommend cogvideox. There is just a new Lora that also improves quality now, and has so much more control. Pure text2vid is worse but with the control, you can get better vids then mochi. Will be far faster too, 15-20mins for a 6 sec video.

[cobalt tartan]

15-20 mins for 6 seconds is still pretty slow but way better.

[cobalt tartan]

Which spec gpu is the 15-20 mins?

[quartz siren]

An 8gb vram gpu is pretty slow, your 16gb gpu should be much faster. 5-10 mins?

[cobalt tartan]

That would be better yeah, It’s a 4070ti super 16gb Ai

[cobalt tartan]

The one in my laptop however is only 4050? 6gb

[quartz siren]

Probably should take around 5 mins, you can use lcm 8steps in just 1-2mins but that will be a bit worse quality.

[cobalt tartan]

Could be worth trying then?

I hope the egpu enclosure doesn’t slow the 4070 down too much it’s only thunderbolt 3 and for until I can manage to afford to build a pc to put it in.

[quartz siren]

Yeah it’s a good idea, I would recommend using kijai’s nodes, he makes really good comfyui nodes for models with lots of optimizations.

https://github.com/kijai/ComfyUI-CogVideoXWrapper

[cobalt tartan]

I use a1111 or whatever atm, I looked at the one you said but it looks insanely confusing and complicated

[quartz siren]

Those are just the updates, you don’t really need to care about them. It’s not as complicated as it looks. There are many tutorials in YouTube, you can search them up, just search cogvideox comfyui.

[cobalt tartan]

I meant Comfy UI in general looks confusing and insanely complex

[cobalt tartan]

A1111 or whatever I think I mostly understand that.

[quartz siren]

Yeah I understand, same for me, so I try to diffusers when I can. A111 is kinda outdated now, doesn’t even support many of the new image gen models I believe now.

For diffusers: https://huggingface.co/THUDM/CogVideoX1.5-5B-I2V

[cobalt tartan]

Damn is it, I used an easy install version, I don’t understand all those command based installs, they always seem to fail to install probably no matter what I do.

[quartz siren]

It’s actually surprisingly easy, assuming you have python installed
You can first just do

pip install diffusers accelerate

And then go to python and do

import torch
from diffusers import CogVideoXImageToVideoPipeline
from diffusers.utils import export_to_video, load_image

prompt = "A little girl is riding a bicycle at high speed. Focused, detailed, realistic."
image = load_image(image="input.jpg")
pipe = CogVideoXImageToVideoPipeline.from_pretrained(
"THUDM/CogVideoX1.5-5B-I2V",
torch_dtype=torch.bfloat16
)

pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

Inference code is

video = pipe(
prompt=prompt,
image=image,
num_videos_per_prompt=1,
num_inference_steps=50,
num_frames=81,
guidance_scale=6,
generator=torch.Generator(device="cuda").manual_seed(42),
).frames[0]

export_to_video(video, "output.mp4", fps=8)

[cobalt tartan]

I don’t know even on the one I install it says something has an update and to go into python.exe and put said command in to update it and it always says the command doesn’t exist

[cobalt tartan]

And sometimes if I install extensions it totally breaks the whole thing and I have to make a copy of my backup copy of the whole folder for it and use that instead which doesn’t make sense to me and is irritating and confusing.

[warm junco]

Mostly deleting the extension and then the venv folder can fix all issues

[cobalt tartan]

Really?

[warm junco]

Because old not updated extensions, break the venv folder. And by deleting both the venv gets recreated clean

[cobalt tartan]

Oh ok thanks I’ll try remember that next time

[median jewel]

When using Hyper-SD on fooocus i get a really good result back in 30 seconds what type are they using? Cause I have tried https://huggingface.co/ByteDance/Hyper-SD But i doesnt give me near as good result or takes forever. Does anyone know?

[fervent thunder]

its not that easy to use hyper

the range of acceptable settings for everything narrows at lower steps

[median jewel]

For me thats ok, I am not looking for perfect result. So if i were to say in prompt of fooocus and using hyper-SD "portrait photo of muscular bearded guy in a worn mech suit, light bokeh, intricate, steel metal, elegant, sharp focus, soft lighting, vibrant colors" it will give me a realkly good result from what i am looking for but using the link and the code there it either generates really fast but terrible outcome or extremely slow and lose the whole point of using hyper-sd

[fervent thunder]

I think try to use a normal model for now

later you can go back and find out what was causing the problem in the hyper workflow

in theory it is merely a case of selecting the right hyper lora for your step count, and setting the CFG that the huggingface page tells you
but if that isn't working then there is likely something else problematic in the workflow

[median jewel]

Its prob a good idea but what i have noticed is that when i am running flux1 my ram peaks and then the code stops because it runs out of ram but when i use fooocus it stays at decent level and my grahics card just works a lot. When using SD 1.5 and sing their code they have on the website it is going really fast but produces bad result, but could i theoretically tweak the settings to get a much better result, is that what you are saying?

[desert dagger]

flux is huge, but you should be loading it into your vram, not using your ram.

[fervent thunder]

is this on CPU, or GPU?

and are you sure you mean ram and not Vram

[median jewel]

should i just enable cuda in this?"

import torch
from diffusers import FluxPipeline
from huggingface_hub import hf_hub_download
base_model_id = "black-forest-labs/FLUX.1-dev"
repo_name = "ByteDance/Hyper-SD"
# Take 8-steps lora as an example
ckpt_name = "Hyper-FLUX.1-dev-8steps-lora.safetensors"
# Load model, please fill in your access tokens since FLUX.1-dev repo is a gated model.
pipe = FluxPipeline.from_pretrained(base_model_id, token="xxx")
pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name))
pipe.fuse_lora(lora_scale=0.125)
pipe.to("cuda", dtype=torch.float16)
image=pipe(prompt="a photo of a cat", num_inference_steps=8, guidance_scale=3.5).images[0]
image.save("output.png")

"

[median jewel]

yes ram

[fervent thunder]

is this windows

[median jewel]

yes

[desert dagger]

okay, no. you want your GPU to do the work, and that means using your VRAM, not your system ram. that's what your CPU uses. and you really don't want a cpu with an integrated gpu

what hardware do you have?

[fervent thunder]

I said this earlier on other channel: there is an issue that affects windows but not linux where loading a .safetensors file causes a big spike of ram usage which is around 2-3x the size of the file

[median jewel]

GTX 1070 gpu
i7 4.2 GHz cpu
48 gb ram 2133 MHz

so not up to date hardware but i can run fooocus pretty ok

[fervent thunder]

seems more likely that what it was, was that you filled your vram and then the ram spike was the data spilling over

[median jewel]

no it doesnt even effect my gpu at all if i look in task manager

my ram just goes bananas and gpu is at 10%

[fervent thunder]

what was the vram usage amount at, when the problem happened

[quartz siren]

Flux will not fit in 8gb vram normally, it requires 24gb vram at the least. You need to use quantization which can allow flux to fit in 8gb vram, let me get the code to use quantization with diffusers.

[median jewel]

th VRam is 3.6/8 GB

[quartz night]

Does anyone know of a stable diffusion model that creates isometric tilesets?

[median jewel]

sorry for being a noob at this

[quartz siren]

Also, flux will take a pretty long time on a 8gb gpu, are you sure you want to use that? You can use hyper/lightning sdxl as well, should give you nice images(considerably worse prompt following, human anatomy, and text) but still decent imgs really fast(just several seconds).

[median jewel]

Well im just doing trail and error and trying to find how fooocus can generate that good result under 20 seconds and all the time I am trying all of them and its either shit or takes forever. But would lightning SDXL be a good chioce for generating under 20 seconds with 4 iterations?

[quartz siren]

4steps or 4imgs? I don't know the exact numbers on how fast it will be on your device, but to generate a single img in 4steps, it should take a few seconds. 4imgs should be complete before 20 seconds.

Fooocus uses sdxl under the hood as well, flux will be much better but much slower as well.

[median jewel]

Ah ok so using SDXL is a better chioce on my device

the code you provided two days ago that was lignbting SDXL right?

[quartz siren]

Yeah most likely, instead of hyper SDXL models, I would recommend lightning SDXL or DMD2 SDXL. Both usually give better imgs then the hyper variant.

I think I did give some sdxl lightning code a few days ago.

with dreamshaper I believe, thats a pretty nice general model.

[median jewel]

but what i dont understand is when i use this code my gpu doesnt really work I is that common on windows?

## import the libraries(instant)
from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler
import torch

## load the model to cuda(should download the model automatically, time depends on your download speed)
pipe = AutoPipelineForText2Image.from_pretrained('lykon/dreamshaper-xl-lightning', torch_dtype=torch.float16, variant="fp16")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to("cuda")

## inference time(should take a few seconds or so)
prompt = "portrait photo of muscular bearded guy in a worn mech suit, light bokeh, intricate, steel metal, elegant, sharp focus, soft lighting, vibrant colors"

generator = torch.manual_seed(0)
image = pipe(prompt,height=768, width=768 ,num_inference_steps=3, guidance_scale=2).images[0]  
image.save("./image4.png")"```

and because of that it takes 3-4 minutes to generate a picture

[quartz siren]

Oh, yeah that uses like 12gb vram so its spilling to shared ram(slowing it heavily), this code should work on 8gb.

## import the libraries(instant)
from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler
import torch

## load the model to cuda(should download the model automatically, time depends on your download speed)
pipe = AutoPipelineForText2Image.from_pretrained('lykon/dreamshaper-xl-lightning', torch_dtype=torch.float16, variant="fp16")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

## some optimizations to lower vram usage
pipe.enable_model_cpu_offload()
pipe.enable_vae_tiling()
pipe.enable_vae_slicing()

## inference time(should take a few seconds or so)
prompt = "portrait photo of muscular bearded guy in a worn mech suit, light bokeh, intricate, steel metal, elegant, sharp focus, soft lighting, vibrant colors"

generator = torch.manual_seed(0)
image = pipe(prompt,height=768, width=768 ,num_inference_steps=3, guidance_scale=2).images[0]  
image.save("./image4.png")"

[median jewel]

i have a lot of pictures of muscular bearded men now lol

[fervent thunder]

its possible to do flux dev on 8GB but it needs careful managing
is a lot easier to fit SDXL in VRAM so that's likely a better choice for now

[median jewel]

my gpu doesnt do anythiong still is it bacuase there is a cpu_cutoff?

[quartz siren]

first delete any other pipe or anything else thats using vram

[median jewel]

yeah that was it now it generates much faster, thanks for the help

[desert dagger]

(and dont' run WoW at the same time on the same computer...)

[median jewel]

I have seen videos of people drawing on a website and an ai trying to depict whats on the screen and it updates really fast how is that possible. Is it just constanly running and doesnt need to reboot?

[quartz siren]

They usually have more powerful gpus and use either sd1.5 models or 1step models.

[fast sage]

Anyone have a workflow or suggest one where I can take a character image and recreate it / make new images with it?

[desert dagger]

it's one step technology. you can do that on the playstation too, martin does that and posts his exploations out to twitter

[fervent thunder]

consistency model is often the one they try these days
for the one step of euler dream

it used to be GANs

SD1.5/SDXL hyper and loras like that can do one step but not so well

[vestal dew]

Anyone know why thiis is happening? Every so often, an image gen will take 10+ minuyrd. I see little use of GPU while looking at task manager. Then next generation, it's back to normal speed. It often happens if I switch something in a controlnet, like choose a new image for openpose or face ID.

[desert dagger]

it's probably loading stuff into vram

[vestal dew]

Sure, but it doesn;'t only happen the first time I generate with certain things selected,....I do multi images for IP Adapter, for example. Sometimes it will gen 50 fast, and then suddenly the next one takes 10+ minutes

[vestal dew]

Just finished...

10:53 to 11:28...half hour to generate an image

next one 10s

9s next one, 10s next one...

now added a new face for FaceID...going to take looks like 10+ minutes

only 9 minutes 37 seconds

this only started happening about 3 days ago

before then, I could change the images...no l;ong generate times

[desert dagger]

what all did you do 3 days ago, and do you have any errors in your system logs

[vestal dew]

I looked for errors...don't see anything

Maybe something else updated something that affects A1111. I'll try installing it in pinnokio...supposedly that isolates everything

[warm junco]

Pinokio ist just an installer

The isolation comes from the webuis venv

[vestal dew]

I see

well, I have abnandoned A1111 for now. Installed SwarmUI...cant figure out how to get IP Adapter installed

tried it on SDNExt, but the output was horrific with ipadapter face plus/

Got it installed, but when using there's an error now:

File "C:\Swarm\SwarmUI\src\BuiltinExtensions\ComfyUIBackend\DLNodes\ComfyUI_IPAdapter_plus\utils.py", line 161, in insightface_loader
02:45:30.663 [Warning] [ComfyUI-0/STDERR] raise Exception(e)
02:45:30.664 [Warning] [ComfyUI-0/STDERR] Exception: Unable to import dependency onnxruntime.

that error is when I try FaceID PLUS V2

When I tru FaceID, another error: File "C:\Swarm\SwarmUI\src\BuiltinExtensions\ComfyUIBackend\DLNodes\ComfyUI_IPAdapter_plus\IPAdapterPlus.py", line 254, in ipadapter_execute
02:48:31.300 [Warning] [ComfyUI-0/STDERR] raise Exception("insightface model is required for FaceID models")
02:48:31.302 [Warning] [ComfyUI-0/STDERR] Exception: insightface model is required for FaceID models

[warm junco]

Yea for comfy you need to install the face id whl file manually into the venv

[vestal dew]

the what into the where?

Iam using swarmUI, not comfy

I see no venv dir in the swarm dir

[warm junco]

SwarmUI is just ComfyUi with an UI on top

[chilly flare]

Hello 👋

[vestal dew]

Ok, so where do I get all the files needed, and where do I place them?

[warm junco]

I dont have time right now to explain it sry

[vestal dew]

Why cant this be listed in the install instructions?

"Simply click this button, and accept the confirmation prompt, to install IP-Adapter to your ComfyUI backend." Seems like it's not so simple...something else has to be done

[chilly flare]

I'm looking for a tiny model like SSD-1B can be use in Google Colab (I'm an IT teacher, my school don't like install everything, it's complicate 😅 ). I'm not seeing a newest version. If you know any tiny model available on Hugging Face, it's can be nice ❤️

[warm junco]

Hey, you can use any 1.5 model that is 2-4gb or any sdxl model which is 6gb

Or do you mean a already cloud hosted one?

[chilly flare]

Oh no I would like to use the basic T4 on Google Colab (16 VRAM)

Currently I'm using this model:

from diffusers import StableDiffusionXLPipeline
import matplotlib.pyplot as plt
import torch
pipe_img = StableDiffusionXLPipeline.from_pretrained("segmind/SSD-1B", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
pipe_img.to("cuda")

It's not really recent (1 year ago)

[vestal dew]

there are no videos on it, no clear explanation in the install guide, tried what it showed in issues but still no luck.

[warm junco]

So you want and sdxl based then or can 3.5 medium work too?

[chilly flare]

I'm not sure if I can running 3.5 ... :S maybe this one: https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8

[warm junco]

And does your colab supports GGUF models?

[chilly flare]

Yes 🙂

[vestal dew]

12 hours and counting trying to get this to work

[warm junco]

Then here is 3.5 medium gguf:
https://huggingface.co/calcuis/sd3.5-medium-gguf/tree/main

But 3.5 as well as flux need additional text encoders like t5 gguf, clip l and clip g

[warm junco]

What do you want to do?

[chilly flare]

I can try it 🙂 thank you

[vestal dew]

Get IP Adapter working in some UI. I had been using it in A1111, but suddenly that was taking 30 minutes to generate images. So I tried in SDNExt, but the output was awful. So I then moved to Swarm - but can't get Face Plus V2 or FaceID to work.

[warm junco]

I can help to get it working on Auto1111

For swarm and comfyui you need to find a face id WHL file

[vestal dew]

Yes, but I don't know what those are, where to find them, where to install them

I have it working in A1111, but every few generations sudden;y an image gen takes 30 minutes.

[warm junco]

Can we move to #🤝｜tech-support
There I can help you to not get the 30mins wait time. I'm sure we can find the cause

[gray vapor]

What does Controlnet tile actually do? Should it actually be called "deblur" or "add detail" controlnet? Is the tile part just the workflow?

Does this treat each tile as a separate image, or does the controlnet actually enforce some global coherence?

[cinder ginkgo]

guys why the ticket support is asking for wallet address?

[warm junco]

Its scam

Please report the username and let us know to ban them

[cinder ginkgo]

It was a Discord invitation that appeared as soon as I asked you the gfx question, he asked to make a ticket in that discord server with a bot and I was contacted by one called "Admin (Don't DM first" and then a private individual who I reported was called like "samantha" and I don't remember the rest of the name

[warm junco]

Yea these scammers are so many its insanse

We ban them daily

[analog skiff]

https://x.com/bfl_ml/status/1859616264324284619?t=DftoDEhtAigmD4sQvsMl2w&s=19

[mellow zinc]

yo peeps, what's a good model for generating environments? nature and city sceneries and stuff like that

[pastel lynx]

Hey y'all, i remember to see some workflow where you could import a 3D model with some node and use it to influence the generation of images but i can't get my hand either on the custom nodes or workflow or examples that use it.
If any remember or know it, it would be appreciated 🙌

[remote glade]

@warm junco i got a 4070 now

which sd should i use now

😅

as in

which is the best ui

[whole edge]

Good evening everyone

[viscid sinew]

Hello

[warm junco]

Forge if you want to use flux

Or comfyui if you like node based stuff

[remote glade]

is it better than A1111

[warm junco]

It has flux and in a few weeks maybe sd3.5 support

But auto1111 is better in terms of extension compatibility

[remote glade]

ah

but 1 question like

is there only 1 a1111 or different like 1.5 or sdxl

or those are models and i can use any with a1111

[latent vault]

automatic 1111 opened all sd models

except the FLUX

[remote glade]

o

is flux the best

[latent vault]

I think so too

it is better than other closed source models

[remote glade]

does it have restrictions like filters

[latent vault]

What is the benefits of FLUX do you think?

[latent vault]

You mena NSFW filter?

[remote glade]

oh yea

that

[latent vault]

The FLUX has NSFW filter and until now, I didn't see the FLUX model without NSFW filter

I think at first, the SD had the NSFW filter and some one removed this filter and released

So I think one day, the FLUX model without NSFW filter released

[tawdry cape]

flux can do nsfw out of the box its just not very well trained on it, if you want nsfw models just look on civit there's a ton of resources and models to make it more nsfw friiendly

if anything i think sd3.5 has better nsfw training than flux out the box, but if you count it with loras flux is superior

[copper crystal]

its' not really a filter. It's just been trained to think nipples are smooth.

You throw the right lora at it though....

[latent vault]

you mean the FLUX can support the NSFW? If I u se only lora, It can?

[copper crystal]

you dont know? go to civit and look at loras for flux. not even a search term. Just filter the list to flux and the top results are all porn

[warm junco]

You need a lot of vram for flux and its slow

[remote glade]

aw.

whats the best model after flux

[warm junco]

With SDXL/Pony You'll have the most fun

[remote glade]

ah.

a111 support pony?

[copper crystal]

Cyberrealistic XL is better than pony if you're looking for particular results. IMO.

Pony is XL, just broken

It's not a base model. It's just a refine of XL that is disaligned so it doesnt' work as well with XL loras or controlnets

It got popular for the same reason Evony the mobile game got popular. Kids like cartoon boobies.

[warm junco]

Yes

[copper crystal]

pony isn't a base model. if a ui supports sdxl, pony is jsut a merge/refine of that.

using pony requires clipskip 2 and special tokens to in every prompt you've got

[keen kettle]

Hello All... Looking to join this group and learn more about creating and using AI models for social media

[median jewel]

is it possible to use image prompt in SDXL lightning using python? I am trying to insert a photo and put him in a desert

[remote glade]

what should i write in the webui.bat apart from —xformers if i have 12gb vram

[warm junco]

You can add --no-half-vae

Nothing more needed for 12gb

[remote glade]

oke

[limpid plinth]

Hello

[grand shell]

ofc, possible

[median jewel]

where can i find information about this and how to do it?

[grand shell]

DM

[elder nest]

hello

[tame cove]

Hello

[copper crystal]

hellO

[low moon]

have a tremendous friday. especially if you are in Korea (south)

if you are in north korea there it is always Monday.

[fast sage]

Anyone know if it's better to genreate depth map with the image, or after in some other application?

[balmy garnet]

Generate an IP three-view of AIDS-prevention

[grand shell]

here

[limpid plinth]

Hello

[ripe perch]

hello

[fervent thunder]

does flux fill work with loras ?

[vestal dew]

Any idea how to get 2 IP Adapters in a comfy workflow? I want to do like A1111 has, get two face ID adapters in there, each one with a different weight, to blend the faces.

[nova current]

hello

[opal socket]

does anyone know of any good stable diuffsion models to use for architecture and archviz?
something that gives good,consistent results

[fervent thunder]

Im interested in video fine tuning , idk if there a server for that

[latent vault]

you mean the video fine tuning service? what does it mean server in your question?

[vague turtle]

hey, this might not be the place to maybe ask this here but i'm in urgent need of "anything" that can make my work and i.e: i have corporate photoshots of my employees and i need to convert those normal portraits of my employees (wearing a suit// typical linkedin profile picture). into a pixar style profile picture (it's for a project). as an intern i have to do this in the next 3 hours. somebody please guide me to where i can do this. feel free to ask me this any question for more context. please give a solution that's possibly a free to use thing.

[fervent thunder]

Discord server ..

[latent vault]

How many people's image do you change?

[vague turtle]

a small team, maybe 10-ish photos.

[latent vault]

how many images do you have for each one person?

[vague turtle]

i have only one as of now, which needs to be converted into the pixar styled. but i can surely ask for more. what's in your mind ?

[latent vault]

I think you must fine tune the image generation model and I have the experience to work with similar project. if you have interesting more, please dm.
And actually, I think there is no free service to complete this work

[latent vault]

Have you ever heard the Cogvideo model for video generation? this model is the latest released t2v and i2v open-source model and possible to for fine tuning

[fervent thunder]

Yes and its ok

But i do want to have mochi or ltx finetunes instead

Willing to pay to make it possible

[latent vault]

Dm please

[median jewel]

does anyone have a code for image prompt in python using lykon/dreamshaper-xl-lightning?

[odd stirrup]

Screen Out : the computer is man's best friend, until when ?

https://www.youtube.com/watch?v=1nn9qrGGvus

[desert dagger]

explain please

[copper crystal]

you'll want to set up and familiarize yourself with the diffusers library, if you intend to do stable diffusion generations through code. Diffusers is the best solution. https://huggingface.co/docs/diffusers/using-diffusers/conditional_image_generation

[scenic shale]

Anyone know if James Cameron (StabilityAI board) has made any videos with the AI models yet? would love to watch what he comes up with 🙏 ❤️

[copper crystal]

Technically, avatar and it's sequel employ a lot of machine learned systems

but no. not "ai video" as we know it

[scenic shale]

Ok thanks, hope they are cooking up a model like Sora and get him to make some cool short films or show episodes
(since he is on board I assume he gets early previews of everything from SVD to new sound models)

[copper crystal]

He probably won't use ai video models in Avatar 3. The guy is fanatical about control over his shots.

[fast sage]

?

Hey lads, do you know a consistent way to get usable text using flux? Some of it is good then it just turns to gibberish.

[desert dagger]

how long is the text you're trying to get it to write? and how complicated are the words?

[fast sage]

Because we can't post images on this chat its in #🧣｜comfy-ui https://discordapp.com/channels/1002292111942635562/1204675216773619752/1309599833958645871

Not really that complicated, it's just quite a few words.

[desert dagger]

that's what the #🏞｜general-with-images channel is for

[desert dagger]

post the text itself, please

[fast sage]

https://tenor.com/view/eye-rolling-eyes-bye-annoyed-gif-13748332

😄

[fast sage]

I don't specifically have it, it's all being generated through another ml, but just looking if theres a way to generally get something more cohesive when you have a lot of words.

[desert dagger]

not really - it's also going to struggle with words with repeated letters such as armadillo, or complicated, fairly long words

[fast sage]

Second question then, does anyone know a good way to generate just blank spaces where text SHOULD be, so I can add it in post. I can't seem to get anythign to work myself blank space for text, no text etc.

This is all in flux.

[desert dagger]

nope. you can try breaking up the text into chunks like this: write "apples oranges bananas" at the top of the image. The background of the image is an abstract image of a banana. write "yummy fruit" at the bottom of the image. <--- sometimes that'll work, sometimes it'll ignore it and just do what it wants

[fast sage]

Yeah that's what I'm finding, thanks so much!

[desert dagger]

works exactly the same way for SD3.5 as flux and 3.5 both use the exact same encoders

[desert dagger]

if you really REALLY want control over the text, just put it on in photoshop

[fast sage]

Yeah that's what I'm planning on doing, but I'm just trying to set blank spaces where text should be with no luck, it's like I want the design withotut the content.

[desert dagger]

you can fix that in photoshop - just put a new layer on top of the image, use the fill to fill it with white, make the layer about 40% opaque, type in your text. then adjust the transparency of the white layer more if you need to, or remove sections of it that aren't under the text

[pale furnace]

https://lu.ma/i8bow7sr

[viscid sinew]

Hello

[whole edge]

Hey fam. What's good?

[viscid sinew]

All good fam. By the way do you by any chance know any platform that could help me run a successful ad? I have been able to reach my target audience with my current campaign

[whole edge]

Well, maybe you should try Adpod

[viscid sinew]

Okay. Could you tell me more about it?

[crude notch]

insane way to self promote

[desert dagger]

google is your friend

[cedar salmon]

thats what they want you to think