💬｜general-chat

[clear oyster]

oh wow 3.5 LFG!

what are the model sizes for 3.5?

[barren mountain]

FP8 checkpoint (unet+clip+VAE) is supposed to be about 15GB, based on other discussions I've seen at least

^ this might be wrong, people debating lol

[quartz siren]

Its 8b. Flux is 12b

[foggy halo]

SD3.5 was a real surprise hit

[barren mountain]

From a dude that sounds like he knows what he's talking about:

100% positive. You can always verify the contents of a safetensors file via the metadata.

sd3.5_large_fp8_scaled.safetensors contains:

• first_stage_model

• model.diffusion_model

• text_encoders

  • text_encoders.clip_g
  • text_encoders.clip_l
  • text_encoders.t5

[red sentinel]

Small and cute.

[clear oyster]

nice, should fit

[clear oyster]

pinged out of nowhere for that announcement lol

[barren mountain]

The "random" channel on the comfy discord is talking about sizes a lot

[foggy halo]

exactly same here lol 😂

[barren mountain]

They were smart to avoid the hype trap this time around

[proper geode]

Just found a Stable Diffusion 3.5 lora 👀 https://huggingface.co/Shakker-Labs/SD3.5-LoRA-Linear-Red-Light

[clear oyster]

but how tf, it just released lol

are people just waiting with compute to modify these base models lol

[rugged vale]

So any updates on licensing?

[fervent thunder]

it's like you joined a cars club but the club became about spaceships
the transition from SD1.5 to FLUX, Pony, and SD3.5 is crazy, the more advanced it gets the less people it interests
in a few years no one will be able to run any model

wtf is INTELELCT1, nevermind

[rugged vale]

Every 2-3 years computing power doubles

Although it's probably approaching 4 by now

[fervent thunder]

who can afford this

[rugged vale]

So it's either fork over a bunch of cash or be stuck with an old timer

[fervent thunder]

it's crazy

[rugged vale]

Also there is for some reason a hard limit when it comes to the performance to compute "ratio"

In other words, no other way but forwards

[fervent thunder]

fine, just wanted to rant a little

[rugged vale]

Me too, don't get me started on Nvidia prices

And market dominance

+refusing to give us more vram

[fervent thunder]

ong

[rugged vale]

Speaking of which, do we know how much vram this model ideally needs

[low moon]

well just have to wait for quantum PCs - generate images in nanoseconds

use quantum entanglement to generate images BEFORE you write the prompt

[twilit solar]

Is it only available by using comfy?

[quartz siren]

Sd3.5 large? At fp16/bf16 it should need fit in 24gb vram. With quantization, you can probably make it fit in 6gb vram gpus.

[rugged vale]

Ah ty

[quartz siren]

Diffusers also supports it but I think apart from comfyui, no other ui supports it. Not sure.

[low moon]

Comfyui master race. XD

[drowsy robin]

The FP8 variant of SD 3.5 fits on the Arc A770 with 16gb of VRAM*

Given you run the encoders on cpu

[lyric idol]

hopefully forge gets an update to run this 🤞

[low moon]

esp cince sd3 doesnt work on it anymore

[amber heath]

Hello just joined the server 👋

[desert dagger]

pretty sure they will - but how long it'll take, who knows

here's the SD3.5 SAI huggingface space https://huggingface.co/stabilityai/stable-diffusion-3.5-large

if you're using it in comfy, all you need is the sd3.5_large.safetensors file. stick it in comfyui/models/checkpoints - there's also an example workflow

[left grotto]

hello

[brittle oyster]

besides here and CivitAi, are there any other good places to browse other peoples work for inspiration? I like just copying other peoples prompts and tweaking them to something I like. Need inspiration...

[shy scarab]

hello just joined

[winter shoal]

You also could just find images you like and use vlm like paligemma or joycaption to get prompts from those images and try to use those in flux to see what comes out 🙂

[manic wraith]

We encourage the distribution and monetization of work across the entire pipeline - whether it's fine-tuning, LoRA, optimizations, applications, or artwork.
Does this mean the license changed? Can we expect something like what PonyXL family of models / derivatives did for SDXL?

(I remember the license of SD 3.0 was problematic enough that Civitai had to clarify stuff before it was even allowed on the site)

[barren mountain]

Today the Pony trainer said they don't plan to use SD3.5 and are sticking to AuraFlow

Other trainers will be diving in though I'm sure

[trim nymph]

anyone have an idea what settings are used in huggingface space/replicate for the sd 3.5 large model? which sampler?

[manic wraith]

I remember he said that much back when SD 3.0 and Flux were new, so I'm sure it also has to do with the amount of work involved

can't just switch because a model is the hot new thing, I just hope it's not a license limitation

I'm really glad if SD 3.5 is good for finetuners, Flux Dev can use some competition!

[obtuse cove]

Hello, recently the vae I was using for SD was marked as suspicious in huggingface, so I'm thinking about what other vae could I use for anime image generation... Any recommendations?

[barren mountain]

This is supposed to be a rough explanation:

UPDATE: the main issue is the enterprise license which Astra is worried about. And also a possible rug pull by stability. Which is understandable when he's sinking this much time and effort into the project

So some concern with the license I guess, still early though obviously

[manic wraith]

makes sense, if I'm not mistaken Aura Flow had a very permissive license closer to what SDXL had

would be really ideal if SD 3.5 can be licensed like SDXL but I guess they must be already worried about ways of making money with the release

(I'm not exactly sure if/how Stability made money from releasing SDXL)

[pine path]

After flux dropped it sounded like FAL wanted to stop putting money into Auraflow. We're already overdue for 0.4 judging from the time it took to go from 0.2 to 0.3, so maybe they already pulled the plug

[blissful totem]

hello

I'm trying to use Automatic1111's Web UI, I just installed today and I'm encountering some technical difficulties, is there a channel dedicated for helping technical difficulties?

nvm just found #🤝｜tech-support

[sudden grove]

What are some key differences in the datasets that SD3.5 and FLUX got trained on? Will SD3.5 rival FLUX in quality?

[nimble canyon]

The license changed and it’s better now but it is still not suitable for Pony. 1 million revenue is a large number but it’s ~80k MRR which is not that crazy for a company that does contracting and gets model licensing payouts(as per license it’s total company revenue, not specific model). The moment we cross that line SAI will tell us to delete the models. At least o have very strong reasons to believe that, you can of course call me crazy but I really don’t see a reason to gamble when multiple great alternatives exist.

TLDR: I have no I data to believe SAI will stand by this statement, and a lot of data pointing otherwise.

[nimble canyon]

Pony is 90% about data so it’s really not such a huge issue to try different model architectures. Plus we have enough funding to try different things…

[manic wraith]

makes sense, wishing the best to finetuners

[rigid harbor]

i hope y'all find girl laying in green grass in your lives..oh happy day!

as per usual i think the finetunes will be where we'll see the true beauty of this model. especially when it comes to aesthetics and i thin thats where it'll knock flux out

[desert dagger]

https://github.com/ostris/ai-toolkit/commit/3400882a8099645ce4c797f57ac258f1e1424ffd

[mild fulcrum]

Why is there a waitlist for ComfyUI?

[broken smelt]

thats pretty sus tbh

[unborn hedge]

i dont trust that link at all lol

[desert dagger]

for the desktop application - it's not out yet, they're taking early release applications

[elfin stump]

Are there any demos of SD 3.5 that have a negative prompt?

[sage reef]

what are you talking about man, the model is so good, you dont need negative prompt :3

[elfin stump]

My process requires a negative prompt to combine styles.

[elfin stump]

But if you have some other way of discovering what the model thinks Picasso without Cubism looks like (for example), please let me know

[sage reef]

hmm maybe i misunderstood the word "demo" in your question, i assumed you meant the online demos, like on huggingface, cause most dont have negatives. if you mean offline stuff, then i dont know why you cant put negative prompt yourself in your workflow, assuming you are using comfyui or idk...
you are now also using the word "process" which would suggest some offline thing anyway, cause online demos dont really allow for "workflows" or processes.

[elfin stump]

I meant my process of exploring the latent style space. And there are HF demos of SDXL that have negative prompts, and of Stable Cascade, and SD 3.0.

Which is how I knew to completely ignore 3.0. it was terrible with a huge bias to a particular style and subject matter.

[sage reef]

so are you using mainly online demos or are you using offline too?

cause if you are using online only, you might have to wait who knows how long until someone decides to put a negative

[elfin stump]

Currently mainly using online demos. Was using the free tiers of some services for SDXL, but that got nerfed.

[elfin stump]

Well, hence my query, right?

[sage reef]

yep i get it now, online only is a struggle i can imagine, i like the freedom with offline stuff :3

[unborn hedge]

man i made a LoRA of my character but the black glasses f*cked everything up with the eyes consistently, the struggle is real

[elfin stump]

I'll eventually invest in better hardware, meanwhile online is sufficient for me to explore how styles can be manipulated.

If you want an example of the kind of thing I do, try a prompt (in SDXL) of "By Alan Aldridge" then try the same seed adding a negative prompt of "Mati Klarwein" and then try again changing the negative prompt to "James Rosenquist".

[sage reef]

i dont get it... if your lora has black glasses, of course its gonna be on top of eyes, you are not gonna see the eyes anyway.. right?

but then again you didnt post a pic so no idea what it is supposed to be anyway :3

[blissful totem]

https://youtube.com/shorts/pNaMOMqK4n4?si=t2z0EW_dItn3T0ww

[elfin stump]

Additional basic examples in these threads:

https://x.com/NerdWorldOrder/status/1740177328955924781

https://x.com/NerdWorldOrder/status/1742294834558517695

[mortal sphinx]

Oooh..

.. 3.5 ':x

Is there a comparison of how much better it is to 3.0 in terms of images?

[sage reef]

3.0 is deprecated

[mortal sphinx]

:c

[sage reef]

a1111 is deprecated cough

[mortal sphinx]

What does that mean though

[sage reef]

its just a bad joke to say that a1111 is terrible compared to the other alternatives

[mortal sphinx]

Oh.

Is it really though?

[sage reef]

but im not gonna stop you if you want to keep using it

[mortal sphinx]

Just to be transparent, I'm not one for jokes or ribbings, nor understand them really. (Kind of why I deleted and asked in tech support instead once realising I'm in general.)

If there's something better than A1111, which ever is the best I'd like to use it, assuming it's free.

[sage reef]

it's usually for beginners, cause its easy for them to work with stable diffusion stuff. more advanced users go for comfyui and variants that use comfy in the background

[mortal sphinx]

Ah, I'm a beginner, yes.

[sage reef]

i mean its ok for now

but one day maybe you will switch over 🙂

[mortal sphinx]

Wait, please don't tell me.

https://learn.thinkdiffusion.com/content/images/2023/11/Thinkdiffusion-stable-diffusion-comfyUI-workflows-ControlNet_Depth_Nodes-1.webp

Is THAT ComfyUI?

[sage reef]

yes haha

but it doesnt have to look that complicated

[mortal sphinx]

Oh my god.

That reminds me of umm..

There's this tool I used to use when I tried to do databasing.

I was really confused by it.

[sage reef]

if you follow the youtube series by scott detweiler for example, he has a very nice friendly beginner tutorial on how to use comfyui from simple stuff to more advanced workflows, it feels like some udemy course for free lol

you could learn comfy with it

[mortal sphinx]

Good grief.. ':<

[sage reef]

its actually not that complicated at all

it just "looks" complicated

[hallow steppe]

I need help with something kind of embarrassing... I really want to see how stable would take my prompt, but I don't have a computer or anything powerful enough to handle it

[mortal sphinx]

.. I'll try it :U

[sage reef]

❤️

[sage reef]

you can always try online demos maybe?

[hallow steppe]

The demos freeze my phone.... I'm using an old android

[sage reef]

oh wait you said embarrasing LOL

[mortal sphinx]

I'm taking 'embarassing' as NSFW, so I doubt he could. ';x

[sage reef]

welp...

[mortal sphinx]

';d

wew

[sage reef]

i mean you could ask someone from here and they could dm you LOL

[mortal sphinx]

Yap :3

[hallow steppe]

No, it's not NSFW, I just said embarrassing cuz I'm poor yet still on this disc lmao

[mortal sphinx]

Aw :<

[sage reef]

it's ok man, dont hide it, we know what you want to generate 🙂

[hallow steppe]

Bwahahaha, I'm trying to generate a villainous character design

[mortal sphinx]

:U!

[elfin stump]

Was it some UML tool?

[rain basin]

whats the latest easiest packaged way to startup on stable diffusion, is it still automatic111?

[sage reef]

omg i remember UML stuff , ugh

[median holly]

anyone have experience with the intel ai playground?

[sage reef]

not a single soul

[median holly]

figured

i bought a a750

very pleased so far

wanted to try AI, don't know if theres compatibility issues with the normal apps

testing out GPU offload on LM studio rn

[visual pond]

its not that complicated as it may seem tbh

[sage reef]

spooky lady got spooked

[mortal sphinx]

Spooked :<

[sage reef]

you are doing fine :3

[unborn hedge]

im happy with my lora, even though it only has like 30 images in its training data and needs a decent amount work and finangling with settings to get a consistent result, the pure dopamine when i get something that resembles my character almost perfectly is like crack

[sage reef]

i know that feel

[opal hedge]

SD3.5L was a huge win, thank you SAI

Just wanted to put that out there

[fervent thunder]

https://tenor.com/view/rock-hyrox-rock-hyrox-eating-rock-hyrox-funny-animal-animal-eating-gif-15423214377703191267

[vestal dew]

Is SD3.5 like 2, crap? Or is it worthwhile?

[opal hedge]

It's lower quality than flux dev but with a lot more variability. In other words, really, really promising for finetuning

Also with surprisingly little censorship

[fervent thunder]

how do you make sd3.5 generate anime?

not having much luck with retro anime style, a woman with long flowy blonde hair, wearing bold makeup, in her casual t-shirt and denim pants, navel cutout, indoors, solid red background. dramatic perspective.

install this node https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb
switch it to token_normalization: mean and weight_interpretation: A1111
then try cranking up the weight on the anime token

[desert dagger]

you should probably go play with it and see what you think

[desert dagger]

use the word 'anime' in your prompt?

[fervent thunder]

did, but doesn't work, i think artist name ref is needed

[desert dagger]

i wouldn't use the word retro - that's likely to throw the AI into the wrong area of latent space.

[hoary pulsar]

Question: Is there a way we can use SD3.5 through a web app?

[fervent thunder]

Fal will have it soon

many will

[desert dagger]

sure. go to mage.space and make a free account then look for the General (stable diffusion 3.5) concept

unlmited gens

[surreal turret]

Is there a unet-only SD3.5 fp8 anywhere?

[hoary pulsar]

What happened to the original dream studio? I see it only has SDXL on it, and an old version of Stable Diffusion 1.6?

[sage reef]

even if there isnt, isnt there a way to "extract" a unet from a checkpoint?

[surreal turret]

Possibly, but last time I tried something like that in comfy it didn't work, and I don't know enough otherwise how to. Also downloading the text encoders for the 10th time gets annoying lol

[sage reef]

but the text encoders are the same, you dont have to download them again :3

[surreal turret]

you don't like bundles? 😛

[sage reef]

depends

[surreal turret]

I don't get even why it is bundled, it saves like 2 clicks for gigs of bandwidth and storage

[fervent thunder]

I prefer bundled for cloud

but I get why people are annoyed by full hard drive

[desert dagger]

3.5 does not use unet

[desert dagger]

no. unet is a neural network architecture. 3.5 uses a different neural network

[desert dagger]

you'd have to ask the people in charge of maintaining it

[surreal turret]

what about version 3.5? lol. Again, I don't know enough but what the weights without the VAE and CLIP is called

[desert dagger]

they're just called the model weights.

[surreal turret]

Well that's confusing

[desert dagger]

how so?

[surreal turret]

On Civitai that would range from loras, vaes, embeddings, etc..

[desert dagger]

those are all models of various types, they all ahve to be trained, just like your image model does, so they all have data they have assigned weights to as they trained.

[surreal turret]

I'm thinking the correct term is "Diffusion model" but that would be easily confused to possibly include the vae, and clip, at least in my opinion. shrug

[desert dagger]

a diffusion model is a model that uses the diffusion process to learn how to create images (simple explination) a VAE is a variable auto encoder model. CLIP is Contrastive Language-Image Pre-training and is an archetecture. when you talk about the AI doing this or that, what you're really talking about is a collection of models all doing individual jobs together to return a single result

[surreal turret]

Yes, but I just want the weights without the CLIP and VAE 😁 I'll just be patient and wait for somebody to eventually probably post it another day

[desert dagger]

the weights for the image model? and what will you do with them?

[surreal turret]

save the bits to my ssd and probably only run once before I download them again using a finetune or something lol

[desert dagger]

do you understand what the weights are?

[surreal turret]

does anybody?

[desert dagger]

yes. so when the image model trains, it looks at images and text caption pairs. as part of learning the information, it assignes 'weight' to the concepts and values and assigns vectors to them. the SD3.5 model weights are on huggingface - but if you aren't a developer you can't do anything with them. what you need is the actual 3.5 model.

from the US department of Commerce: "Model weights reflect distillations of knowledge within AI models and govern how those models behave. Using large amounts of data, machine learning algorithms train a model to recognize patterns and learn appropriate responses. As the model learns, the values of its weights adjust over time to reflect its new knowledge. Ultimately, the training process aims to arrive at a set of weights optimized to produce behavior that fits the developer’s goals. "

[surreal turret]

I'm really confused where we are disagreeing but I know all the "high level" stuff just nothing underneath if that makes sense

[desert dagger]

it makes sense. we're not disagreeing, but what you're asking is for me to give you all the molecules that make up flour sugar and salt so you can bake a cake.

but what you need is the actual flour sugar and salt

if i give you the flour sugar and salt, all the molecules will be in them, and you can actually use them

[surreal turret]

lol, All I know.. right now all I see is model bundled (with VAE and clip text enc), I just want the bundle-less. Because I already have the VAE and clip on my drive.. I'm too tired to continue this conversation and learn anything new

[fervent thunder]

save model node in comfy

load the result in "load diffusion model" node

[desert dagger]

https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main go here and just download the .safetensor's file thats on the bottom of that page

[surreal turret]

I'll try it tomorrow

[desert dagger]

you should probably also grab the example workflow on that page as well

i don't think you need anything else, but if you do, they will be in the folders on that page

[fervent thunder]

I get confused about all the repos and which is which

when I used SD 3.5 earlier I just set it to download from like 5 different repos LOL
and tried them all

[desert dagger]

https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main that's the huggingface stabilty.Ai page for devs but it also has the 3.5 large model on it

[fervent thunder]

ah okay yeah

[desert dagger]

and it's also been uploaded to civit now https://civitai.com/models/878387/stable-diffusion-35-large

[fervent thunder]

there is this experimental "scaled fp8" one that Comfy made but I haven't got it working yet

[surreal turret]

I saw that but also bundled lol

[desert dagger]

i haven't tried. i did download it but i don't need it

[fervent thunder]

I rarely have the correct CUDA or pytorch for stuff like that

[desert dagger]

i'm also not using Comfy's workflow. i'm using the workflow from SAI as i know it has all of the controls in it that it should have

[fervent thunder]

oh yeah I should switch to that, I was on the Comfy one

using the workflow of the model creators at first is good yeah

[surreal turret]

Looks like gguf formats are also available, but those take like 5 minutes to load for me for some reason (well the flux ggufs did)

[fervent thunder]

I haven't learnt how to use gguf yet

the cool part about gguf is they work here https://github.com/leejet/stable-diffusion.cpp

no python

[warped quarry]

halo！

[blissful totem]

guys which checkpoint do you guys recommend for anime art?

[steep sage]

Will some 3.5 model support 6gb vram gpus

[steep sage]

Waifu diffusion

[blissful totem]

btw, do you know a good checkpoint to modify clothes

and a checkpoint to modify facial expression?

with the inpaint

I want to be able to close eyes

or open mouth as if the character is talking

[steep sage]

Controlnet and Inpainting

https://youtu.be/AGqqj_xQ6IM

[fervent thunder]

where command Chanel

[desert dagger]

read the information in this channel #artisan-faq

[trim nymph]

where is this workflow

[desert dagger]

https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main on this page near the bottom along with the 3.5 large safetensors file

ignore all the other folders and stuff, that's for the devs that want to play with diffusers

[trim nymph]

do you have an idea how to get results like in the HF space or replicate with sd 3.5 large? @desert dagger

like which sampler they use and so on. the results im getting there are great

compared to my comfy results

[desert dagger]

you want one of my workflows?

[trim nymph]

yes! please

[desert dagger]

i'll DM it to you, can't post it in this channel and don't want to post in general with images

[trim nymph]

thanks

[silver sundial]

how long does it take to generate one image? using sd3.5 or sd3

@desert dagger How much experience do you have?

[hallow steppe]

I got added randomly from someone named Jon Snow.... Their messages seem a tad off and it says they are a mutual server member to this server...

[short tusk]

Hi ! im new in the server & to this community & its my first time using Stable Diffusion with the A1111 GUI, and my goal is to make some realistic women pictures taken from a smartphone in a room with natural lights & not too much exposed reflections on the skin, like in the real life

i already did some try, but i was never satisfied to the final results

[low moon]

Well first of all use comfyui or at least forge- then u can try some of the realism loras from civitai, there is even a VHS lora

[short tusk]

aight bet, ill start with this

thanks

its always better to use a checkpoint & a LoRa at the same time ?

[low moon]

depends, if you want something very specific then loras help

[short tusk]

aight aight

good looks bro

[desert dagger]

On my machine, it takes about 30 seconds

[blissful totem]

I kinda suck at prompts

[desert dagger]

it works very well to go to meta or claude, tell them what AI you're prompting, tell them it's specifically drawing an image, and ask them for a prompt for it

[blissful totem]

dunno whats meta or claude. Those are image to prompt AI's?

[desert dagger]

https://www.meta.ai/ <--- meta - it's an LLM

[blissful totem]

Thanks man

it's better than chat gpt I suppose

chat gpt has limited amount of image you can send per day

[desert dagger]

everything is better than chatGPT - meta is free, and if you just give it a prompt, i'tll generate images. if you ASK it about generating images, though, it'll tell you that it can't

[fading gyro]

any good guides for using automatic 1111 api?

[low moon]

https://www.youtube.com/watch?v=qDJrSK6uynQ

Cat with 4 H100s (send help) can play with this. XD

[lapis urchin]

can you run sd3.5 in a1111

and if so what model im guessing medium but im not sure

im running it on a 3080 btw

[lapis urchin]

(medium isnt out yet)

is it possible to use large on consumer hardware, guessing i should use comffyui

[warm junco]

Nope

[lapis urchin]

yeah switched to comfy

its my go to anyway

[silver sundial]

I tested it on H100 and took 9 seconds. but ti's not competitive with flux model

is there any solution to run flux model in a few seconds?

[quartz siren]

Try this: https://github.com/aredden/flux-fp8-api
Should only take 2-3ish seconds.

Also, sd3.5 large should be competitive with flux(few pros and cons for both), did you use sd3 medium?

[gleaming helm]

hello?

[floral umbra]

Is SD 3.5 more akin to flux in "structure"? Or closer to SD 1.5 and SDXL? Thinking since we haven't gotten a proper tensorRT converter for flux, but maybe for SD 3.5?
Asking due to it using the same text encoder as flux

[desert dagger]

sd3.5 and flux use teh same neural network arcitecture

[desert dagger]

https://arxiv.org/abs/2403.03206 this is the SD3 arXiv paper

[floral umbra]

Ah, shit. I can forget about finding a way to speed it up then Unless LCM loras is the rescue there for 3.5 due to not releasing only as distilled unlike flux.

[floral umbra]

Didn't fully understand what the paper meant as most of that is hieroglyphs for me as of right now :P

But got the gist that it more follows how it should do, vs the older models being less accurate

[desert dagger]

the older models use unext. sd3 and flux use a different neural network

[floral umbra]

Though, as usual, we're gonna have to wait for finetuned models, as stabilityA.I is still sloppy with details and especially certain anatomical traits

[desert dagger]

and while flux is frozen and untrainable, sd3.5 is flux, flexible, and trainable. just give the community time to write stuff

[desert dagger]

i'm going to take that as a personal insult

[floral umbra]

Yep, that's the plan, and i've waited for the proper release of the full SD3 since medium released in june. Now to wait for the popular trainers to go ham on the base model

[desert dagger]

a lot of stuff has already been implimented in the last 24 hours, and the dev behind AI toolbox has already got code in place for a 3.5 trainer. he posted that yesterday

[floral umbra]

Was more general question regarding how the structure was, and whether it'll be "soon", or quite some time until tensorRT has been written/re-written to properly convert and accelerate sd3.5.

[hallow steppe]

Guys, I'm absolutely horrified....

[desert dagger]

it's the season for that

[hallow steppe]

Yes, but I'm not in the way favorable

[desert dagger]

it's the month for that, too

[floral umbra]

Noice :) Cause i was going to try training for a 3.5 lora as soon as ai-toolkit updated for 3.5 as flux nailed my first ever successful lora immediately. But sdxl failed all 12+ times lol

[hallow steppe]

My friend used 3.5....

[desert dagger]

he posts updates to his twitter profile, if you're not following him, you might want to

[desert dagger]

did they now

[hallow steppe]

Used a generic anime model

[desert dagger]

there are no generic anime models for 3.5

so what model was it for?

[floral umbra]

I'm at the point where i forget i have a twitter. That's how little i use it. Plus pretty much stopped using it after musky elon bought it and killed the bird

[hallow steppe]

It was just a common anime esque one

I don't remember what exactly, it was just anime ish

[desert dagger]

you can't jsut use a model here and a model thre and hope they work. checkpoints and loras MUST be used with teh model they are trained for - or you either get nothing or you get a real mess

and there aren't any for 3.5

[hallow steppe]

We got a result...

[floral umbra]

Lol, found this on civitai:

SD3.5 Med/Lite (11GB) Improved Duel CLIP (Full Checkpoint) Duelling clips

[hallow steppe]

Maybe it wasn't 3.5 they said it was newer though

[desert dagger]

i'm sure you did. and i'm equally sure it was not correct. because you used a check point/lora written for something other than 3.5

[desert dagger]

there's a really cute kitten lora for 3.5 on there too

[floral umbra]

or you get a real mess
That's part of the fun! Especially when you use the wrong lora, and you'll get a deepfried meme as well. So 2 birds with 1 stone xD

[desert dagger]

3.5 doesn't need a lora or fine tuned for anime anyway. just needs the right prompt

[floral umbra]

Not lora.. VAE lol. Took me a good sec to remember xD

[hallow steppe]

The result we got wasn't exactly within full coordination with our promp....

[floral umbra]

What's the name of the model?

[hallow steppe]

It was... One might say very quesionable

[hallow steppe]

No clue

[desert dagger]

tell ya what. go back to 3.5. get rid of the lora, get rid of the checkpoint. jsut use 3.5 by itself, prompt it and see what you get

ther's a whole bunch of 3.5 stuff in the #🆕｜sd3 channel, including anime

[desert dagger]

using the wrong vae will get you interesting results for sure

[floral umbra]

Oh shit, forgot i downloaded the fp8 sd3.5, need to give that one a go

[desert dagger]

ROFL! forget you have a twitter, forget you downloaded a cool model... are you trying to do too much all at once?

you probably want to use that one with comfy's workflow

[floral umbra]

Also, do you know of a tool that can auto sort my models within a folder?

As i might have dumped some loras intended for SDXL into SD1.5 for instance, so wanna quick move them into subfolders of sd1.5 and sdxl, then cut them from those folders where they belong :P

[desert dagger]

um no? but if you ask in the #🤝｜tech-support channel, someone in there might

[floral umbra]

Popcorn brain mate. Welcome to the world of having adhd

[desert dagger]

had 2 kids and a nephew that had that, one so bad he could only make sound effects to express himself. not a fun life at times

[floral umbra]

True Though, not too sure if it fits as "support", as it's not really a "problem" per say

[desert dagger]

but that's where most of the really technical folks hang out

[floral umbra]

Aye. Nephew i have no idea if he has adhd, autism or what, but he expresses himself by overreacting quite like "that's so raven", or other "live action" disney actors when they exaggerate them expressing themselves lol.

Aaaand that's why i bought my loop switch ear buds. IRL volume knobs to shut everything up 3 decibel volumes to choose from

[desert dagger]

how old is he?

[floral umbra]

7

[desert dagger]

could just be in the first stages of puberty and out of sync with himself

[floral umbra]

And when he's playing on his tablet, he constantly narrates everything, at every second lol

[floral umbra]

Could be. I was quite the opposite. I was quiet, but my mental narration/fantasizing/activity was all over the place :P

[desert dagger]

and https://x.com/cocktailpeanut/status/1849201053440327913 you need to read that

[low moon]

sd35 refined models will mop the floor with flux.

😏

See this is the moment weve been waiting for

Definitely worried about Cat with4 gb of vram (send help). He surely shouldn't miss these historical moments.

[unborn hedge]

im not using flux or 3.5 cause my computers not good enough lol, ill stick with SDXL for now

[low moon]

Fair... Whats your specs?

the bare minimum these days is around 8 GB vram I believe.

[unborn hedge]

i have a (regular) RTX 4080 with 16GB of VRam and 32GB of regular RAM, and i tried Flux Dev, and it was slow but worked, then i tried to put a LoRA in and my computer started chugging really badly so i got scared and de-installed it, i might just wait until the far future where flux is easily available without all these shenanigans

[low moon]

I have a 3060 rtx with 12 gb of vram and 32 ram and I trained some loras for flux and they work... I mean its slower but it works - I can stack loras and it slows down but it works...

Takes a few mintues what can we do... at leats its local. after our next upgrade in afew years it will take a second probably.. Still amazing...

And I mean DEv too not Shnell...

[unborn hedge]

i think i also tried with schnell too and put a lora and it chugged like crazy, i use forge for flux and automatic1111 for sdxl and 1.5 cause thats what im used too

[low moon]

A1111... well maybe thats your problem... try comfyu or at least forge

[unborn hedge]

i used forge when trying out flux, not automatic1111

as a seperate install

[low moon]

just try comfyui - u can set it to use the A1111 folders so u don't have to re-download al the models again

comfy is the most optimised always

[unborn hedge]

idk about comfy's ui though with all the node connecting stuff, i like UI like automatic and forge

[low moon]

Yes. I was scared at first too for about 4 months but after that...

I gave in and I do not regret it.

[desert dagger]

hide the connecting wires if you don't want to see them. they're just for your information so you can see how the functions are connected

[low moon]

Comfyui is for "serious folks". A1111, Forge, Focus is for "casuals".

It's like an unspoken rule: everyone knows Comfy is in charge but no one admits it. 😉

[unborn hedge]

man this is giving me a headache, all i want to do is to run flux with at least 1 or 2 LoRA's without my computer chugging and that doesn't take forever to start then generate, and that is also better then SDXL with prompt stuff and inpainting or whatever so i dont have to spend forever inpainting something to make it look good

[low moon]

Hmmm....

Well this is an emerging tech.

I am waiting for about a year now for a decent local video gen.

Flux will take a few minutes on almost any system now especially with some loras.

I started about a years ago with all this during SD15 times. It took forever so I upgraded to handle SDXL. which I can handle now ok but now there is flux and sd35 and hunyuan and so on. things move fast now and it will be a bit more time before we can settle into a pipeline we are comfortable with and that does what we want it to do.

At least 10 years from now we will be able to say "we were there when this started" because this tech will change everything we do radically.

[unborn hedge]

yeah, i just wish it wasn't so confusing to figure it out

[low moon]

wELL OUR GRANDPARENTS TOOK A WHILE TO FIGURE OUT THE INTERNET

A SERIES OF TUBES

gRANDPA: HOW DO I DOWNLOAD THE INTERNET?

lol sorry for caps

keyboard is on pillow.

[desert dagger]

that's okay, today is capslock day, i think

[hardy sedge]

Works stable Diffusion 3.5 with a 4070 TI 12gb vram? Or i get Out Off vram? I dont want to wait longer then 20 sec for a picture.

[languid harbor]

how and from where do i dowload stable diffusion 3.5 on my pc to run locally yk

[boreal kernel]

How can I generate an image in this Server? like we used to do using Bots!

[gusty mist]

youll need quantized flux model. Should not be a problem with 4080

[jaunty osprey]

what's the deal with SD3.5? it's too early to bother using it, right? how does it do next to flux and SDXL? i've heard prompt adherence is much better?

[fervent thunder]

you could use it right away if you want although its probably not gonna seem quite as visually impressive as flux to you in its current state

I don't think it is far behind though

[solar wraith]

I was trying to run latest stable diffusion 3.5 large model on g5xlarge with 24 gb of GPU memory . But seems like it is not killing the task due to less computation.

Does anybody knows GPU requirements for running stable diffusion 3.5 large model ???

[barren mountain]

It should run easily on 24GB (I've done this myself), run reasonably on 16GB, run OK on 12GB, then I think recent optimizations make 8GB possible

[solar wraith]

I just saw my system RAM is getting killed .. it's not the GPU problem

My system RAM is 16 gb

[barren mountain]

It's a base model so I think the best thing to do right now is compare it to SDXL Base, and I'd say it's an improvement. Flux Dev is definitely better than SD 3.5 base, but the real question is if 3.5 fine-tunes will surpass Flux Dev

[barren mountain]

Ouch yeah that will limit you. Try FP8 versions whenever possible (unet & t5xxl clip)

[solar wraith]

Fp8 version meaning what settings I have to touch ?

Or directly use fp8 model u are saying

[gusty mist]

https://civitai.com/models/622579/flux1-dev-fp8

and for sd 3.5 altough i havent used it yet: https://civitai.com/models/879701/stable-diffusion-35-fp8-models-sd35

[solar wraith]

Thanks !

[unborn hedge]

yeah i heard of that, though im not sure what exactly that is, or if it'll work with loRAs or be as good as regular flux

[gusty mist]

do you use comfyui?

[unborn hedge]

i use forge for flux

[gusty mist]

Quantize = Exchange quality with file size. Same quality as regular flux? No. Quite as good? Possible.
Will loras work: For me yes they do, but it depends on what checkpoint the lora was trained with.
Ive never used forge so i cant help you out with that.

But I can tell you that im running flux on a 3060 with approx 1:20 per image with lora on 1024x1024

This is my comfy workflow if still interested: https://civitai.com/articles/7292/flux-easy-workflow-lowvram-gguf

[unborn hedge]

thanks for the info, last time i tried out flux with forge, it took me 2 minutes because it was unloading the model? then doing a bunch of things then loading the model again then generating?? when i looked in the command window, so i think ill try out comfyui for flux since i heard you can hide the node stuff which im not comfortable with

[gusty mist]

Yea newest update adds hiding nodes. Workflow is very very simple.
What you are talking about is the model offloading from gpu to ram. This is common procedure.
With more VRAM you need less offloading.

[unborn hedge]

ohhhh, it was probably doing that because i wasn't using any quantized versions, luckily i was just messing around with it and made some comic meme panels that got deleted with the install (i forgot to move it)

[gusty mist]

exactly! (smaller file size, less offloading too) you understood! 🙂

[unborn hedge]

im trying my best lol, AI stuff goes so fast nowadays

[gusty mist]

hard to keep up, for sure

[unborn hedge]

FLUX 2 is probably gonna release tomorrow and make this whole thing irrelevant knowing my luck lol

[gusty mist]

With probably 48gb of file size huh 😄

[unborn hedge]

400GB of VRAM required and tomorrow after that, its down to like 3GB

anyways, ill try to find the right version of quantizised flux thats appropiate for my 4080 and 32GB build

[low prawn]

Imagine:

Like deforum
With time stamps and images, but it makes a video, like upload an entire script from a movie and it plays out

[unborn hedge]

we'll probably have the singularity and agi by then (and hopefully they won't terminator or matrix us by then)

[gusty mist]

@unborn hedge https://civitai.com/articles/6730/flux-gguf
Checkout the Quantization Comparison Table

[steel fable]

What is NFT DUE?

[low moon]

the next image gen model should really be called FLEX

[cedar salmon]

FLUKE

[low moon]

and the small horse finetune will be FLEXXX

[alpine mirage]

hi

[blissful totem]

does anyone know an AI gif generator?

I want to make like an animation of a character that I created

like idle, closing eyes and stuff

[desert dagger]

try Glif

https://glif.app/glifs

[blissful totem]

is free? 🙂

[desert dagger]

i've never paid anything to use it.

[blissful totem]

I wonder if I can input an image

[desert dagger]

https://docs.glif.app/

docs page

you can do all sorts of stuff with it, go explore it

[blissful totem]

okay

[short tusk]

Hi !

did someone here already done a faceswapping on a existing image on Forge ?

because actually i try to see how to do it, but it doesnt seem to work for me

ahh whatever... have a great day & good luck with Stable Diffusion yall, im too autistic with these kind of stuff fr

[unborn hedge]

thank you for the help! i'll have to get more training data before i start using FLUX for my comic though, it'll take a little bit

[inland mortar]

would someone experienced with comfyUI give me proper launch args for a 6GB card specifically the 1660 Super as I am having an issue where my id 0 gen will be super slow (90-160 s/it) and in console it says that it's requested to load SDXL and then loads it before the gen. But then on my id 1 gen it will gen super fast at around 7 s/it but it says unloading models for lowvram mode and then unloaded 0 models, but it's generating super fast so I don't have to wait 20 minutes or longer for a single generation. This is in lowvram mode, but in normalram mode it's slow every gen. So how do I get comfyui to a point where I can always gen at that 7 s/it on SDXL

[placid hatch]

Anyone know if controlnet-union works with pony or illustrious based models?

[inland mortar]

@bleak matrix uhh

[main snow]

is MexxL_LCM2 some kinda super mega forbiden jutsu model? can't find it anywhere

[desert dagger]

i did a google search for MexxL_LCM2 and found several hits

[potent horizon]

Hi everyone, I'm newbie in SD and ComfyUI

[desert dagger]

welcome

[potent horizon]

I'm using Mac M3, I tried to generate image with Flux1.dev, it took super long to generate
is there any recommended lightweight model that I can use, to generate in seconds, so I can focus on testing ComfyUI functions?

[desert dagger]

you shoudl probably ask in the #🤝｜tech-support channel

[potent horizon]

aah nice, I'll ask them, thankss

[vast narwhal]

Hi all, new here

[main snow]

I only found one and it's a 3h download

@bright tapir

[desert dagger]

all i know is what i found on google.

[main snow]

guess it's gone then lol

[full zodiac]

yo! can anyone here help me to find out what A.I tool they used for a Comic? 😄

[desert dagger]

we can try. which comic?

[full zodiac]

http://kck.st/4eSDPpF

this one i want to know what A.I i been trying to find all day! @desert dagger

[unborn hedge]

im making a comic too! im gonna use FLUX because i heard its the most cutting edge generator so far and i need consistency and prompt adherence in my comics, especially with the super fine-tuned loRA im making so FLUX knows what my character is

[full zodiac]

@unborn hedge I'm trying to find how to get that rendered vibe the did with consistency ! You mean flux A.I? i will check it now

Image ratio on flux is too low! @unborn hedge

[desert dagger]

midjourney, stable diffusion, flux, and others, will all do that style

look on civitAI for comic book loras

[desert dagger]

you can't fine tune a lora for flux - and you probably don't need to. try just prompting it

[unborn hedge]

but i see loras for flux on civitai tho???

[desert dagger]

yeah, i know. but you said "super fine tune" and flux - is basically just a giant lora itself. it has a very narrow range. stick in that range, you don't need a lora. try to go outside that range, you're going to get nowhere. if you really want to create fine tuned stuff, use stable diffusion 3.5 which just released

[full zodiac]

You need to train a model first step

@desert dagger at loras i found a video on yt

https://www.youtube.com/watch?v=KJf67JiOWMY

[desert dagger]

you don't need to train squat. just use the base flux model and prompt it well

[full zodiac]

But the quality doesn't look this sharp . or the backgrounds etc. it's basic comic

do you get the same character without training?

[desert dagger]

there are good loras on civitAI for comic book art.

[full zodiac]

the only i found close to the quality aspect and the render is Leonardo A.I

[unborn hedge]

i have a specific-looking character though that im positive that flux wont get right

[desert dagger]

you're probably going to ahve to contact the creator of that video and ask him or her what they used

[desert dagger]

and i'm positive you can't train flux - wait a week till the trainers are updated and use SD 3.5

[full zodiac]

They say that they trained a model for the character from scratch and then do the whole thing. but they dont mention what A.I was used.

[desert dagger]

at a guess, they used SDXL

[full zodiac]

whats that xD

😭

[desert dagger]

Stable Diffusion XL

[full zodiac]

oh

[unborn hedge]

so if flux is a giant lora then whats all the other stuff on civitai? civitai says its a lora as well

[desert dagger]

are you familiar with what a LoRA is actually for? what it does?

[unborn hedge]

i am

[desert dagger]

then you know that the entire reason LoRA was developed was so you could update the weights of a model and not have to retrain the entire model just to update information in it, or add information to it, correct?

[full zodiac]

im trying Lora right now. but i doesnt seem like you can use this for pro stuff. The image quality aspect ratio is ridicullusly low

[desert dagger]

i think you might want to start by watching some how to train loras and checkpoints tutorials on youtube. i'ts not that easy

putting a good data set together first is a skill

[full zodiac]

@desert dagger 912x1144

:S LOL

[desert dagger]

that's a rather odd aspect ratio

[full zodiac]

1792x2688 is the Ultra quality i can get from Leonardo max out.

it's close to the comics aspect

[desert dagger]

okay, but the AIs all have aspect ratios that they do well in, and outside of those, you get issues

if you want to use an AI to do everything, you're stuck with the AR that the AI you pick will do without issues

[full zodiac]

yes but to have a crisp render or a good image you need at least 1000x2000 and above

[desert dagger]

naw, you just upscale

[full zodiac]

otherwise it's only for showcasing on the internet chats or sites

you can upascale but .... meh

[desert dagger]

nonsense. and you don't need that large for print, either

[full zodiac]

it does the trick yes upscaling you are correct

but its not the same . for the detailes

up scaling sometimes for example makes the image realistic

[desert dagger]

unless you're printing on a web press

[full zodiac]

I'm talking for the new era of comics generated with A.I

not for general creations

yes . if you want to print also you need a good source

[desert dagger]

and i'm talking about actual publishing. you don't need sizes that large. and if you do, you can upscale the image with something like topaz, which makes it larger without changing it.

[full zodiac]

oke!!!!

thanks for the help!!

[desert dagger]

but before you can create your lora, or finetune a check point, you need to learn how to make a good data set. you really should consider watching some tutorials

[full zodiac]

i will check it yes!!

the character consitency is the hardest part !! @desert dagger

[fervent thunder]

Topaz probably still beats the open source stuff on average yeah

the Enhance Everything discord server is good if you like upscaling

the majority of popular upscalers with .DAT .ATD etc were made by people there LOL

if someone manages to extract the control net from Supir so we can use it in regular Comfy workflows then that might be the best, but attempts have not succeeded so far

[desert dagger]

why do you want to do that?

[fervent thunder]

what happened was the Supir team made the best Tiled control net of all time
but they locked it up in a lot of Diffusers code so it needs extracting

[desert dagger]

which model did they make it for?

[fervent thunder]

oh its SDXL

StableSR is SD 2.1, funnily enough, and is still very good, proves that SD 2.1 is a strong model

[desert dagger]

2.1 is a very strong model, but 2.0 had ... issues and no one liked the 2.X series after that. so SDXL...

[rich swift]

is there something similar to AnimateDiff but for flux 1 dev ?!

[desert dagger]

nope

[rich swift]

feels sad

[desert dagger]

you could try using cogVideo with the images you create with flux maybe

[rich swift]

@desert daggerare you pretty good with it ?

[desert dagger]

nope, but others are using it and they seem to be liking it

[rich swift]

aight ima have a lok thanks

look

[silent narwhal]

Hi everyone, I have a question, I'm looking for a new computer with a budget of around 800💲 that I can use for modeling and rendering as well as using sd and comfyui, any recommended configurations?

[worldly socket]

[gusty mist]

For training a lora? If ur using GGUF variants of FLUX you can probably train the lora with the base checkpoint. (I think Q8 performs best with LORA and Q4 the worst. Some said below Q4 loras wont work properly).

However if you want to use for example a flux 8 step model like Hyper Flux, youll probably need to train it with that as the base model. Just keep that in mind.

[hushed kernel]

what settings would you use for generating realistic images with sd35, shift, steps, cfg, sampler and scheduler?

[gusty mist]

After finetuning a model, will licensing remain the same?
What about merging two checkpoints?

[quartz siren]

Mochi 1 is definitely the best open text to video if you want(it’s very slow locally, 1 min on fal), CogVideoX is considerably worse but supports image to video, trajectories and is faster.

[hard swift]

gm all

[obsidian wagon]

Where is the best place to follow to see the latest releases of checkpoints and Loras to use?

[low moon]

CivitAI?

[low moon]

Well you know in the good old days you had to wait for film to be developed before you could see anything, now you wait for video gens to generate... :/

my video on kling has been almost ready for nearly a day XD

this is why I preach local, anything else is worthless, better invest in a good GPU

[quartz siren]

A day? Yikes, you can use Fal too, it takes mochi like 1 min to generate there.

[low moon]

i know i tried it - mochi has great promise

omnigen + mochi = movie studio

from what i understand voice/lip syncing is also very advanced by now tho I am not sure what would do that locally

[unborn hedge]

I only ever tried Animatediff tbh, I wonder if CogVideoX is better, I like the image to video feature

[low moon]

it is better but not as good as mochi

until they tame mochi cog is our best local video gen tho

[unborn hedge]

ohh ok, I wonder how it runs on a RTX 4080, I took a glance at the GitHub and kept seeing mentions of 4090 but I don't have that GPU

[serene terrace]

Hi, is there no more bots for freem stable diffusion imge generation ? I only see the artisan .

[alpine nest]

hello, tell me please how to generate image for free?

[oblique elk]

There is a fp8 Version of mochi out, with unloading of the VAE it should work with 24gb VRAM at least with like 160 frames...

[low moon]

still out of my reach...

[desert dagger]

https://huggingface.co/models?other=base_model:adapter:stabilityai/stable-diffusion-3.5-large a lot of LoRAs for 3.5 already on huggingface

[lean terrace]

https://openai.com/index/simplifying-stabilizing-and-scaling-continuous-time-consistency-models/ hmm

[dawn mulch]

Can anyone show if sd 3.5 can do, 'A woman lying on top of a pool of marshmallows.' As far as I know it can do, a woman lying on grass'. But can it do marshmellows.

[desert dagger]

can't post images in thsi channel, so posted it in #🏞｜general-with-images

[frail tulip]

any suggestions for a model to generate artwork for a house track i want to upload to soundcloud? wouldnt mind there being text in there but np if not

[desert dagger]

with nothing more than that to go on, all the models out there can generate images that would work. more details maybe?

[frail tulip]

so I asked GPT to make me a prompt for artwork for my track i uploaded - ''Eerie, dark logo with the text 'Heavy Music' in bold, distorted font. Shadowy, misty background with a subtle, ghostly glow in shades of dark green and purple. Abstract, twisted shapes and faint, flickering lines, giving a haunting, mysterious feel. Minimalistic yet chilling, with a touch of horror and a supernatural vibe''

oh i need to post the img in the other chanel

not much like the prompt description .. but im totally new

[desert dagger]

yeah. this channel doesn't allow images so that it can be more conducive to chat :)

[desert dagger]

@fervent thunder https://x.com/_akhaliq/status/1849945005269336096

[pseudo bough]

Any open source way to do what runway act one does

[low moon]

cogvideoX (not very good yet) mochi 1 (very good but not yet adapted to most consumer hardware)

[cedar salmon]

@warm hull just in case i wouldn't go there, that's some random server

[warm hull]

Wait i'm confused, what happened?

[frank trench]

hello, i have rtx 3060 and i can't use 3.5 large error of memory

i do'nt understand

i try all driver and also cuda 118 and other

im bored

bored

[warm hull]

the support ticket is a scam ?

[frank trench]

Nvidia bored me

if somebody can help me

[desert dagger]

you could just go to meta.ai and tell it to review what it knows of SD3 and then ask it to create a prompt

[desert dagger]

support ticket where?

[warm hull]

I swear there was a guy who claimed to me an admin and directed me to create a support ticket. but the messages are deleted now

[warm hull]

Is meta.ai any better than chatgpt? I'll try it

[desert dagger]

if it was here, the support tickets are legit. if it was somewhere else, i ahve no idea

[desert dagger]

not only is it better, it's free

[cedar salmon]

@warm hull went afk but ya i assume it was some scam, why i said something, certainly nothing to do with SAI

[cedar salmon]

the beware of scammers message was an invite link to a discord before it was edited

[warm hull]

Ahhh ok

[warm hull]

I'll try it out for sure and see how it goes. Thanks

[warm hull]

I probably fell for that ngl i'm still confused. Cause there was some support ticket thing and i clicked on it but @desert dagger said it's legit

[cedar salmon]

in this case crystal is wrong, it was an invite to a discord server named Support Ticket, that scream scam

they didnt see it before it was edited, so confusion

[warm hull]

Damn i fell for that

It looked soo legit

[cedar salmon]

dude was quick lol

[warm hull]

soo quick man

i thought it was a bot

[cedar salmon]

yo that was literally their first message in this server

oh well, live and learn

[warm hull]

Yessir

Thanks for letting me know man

[desert dagger]

i said if it was for a ticket here - it was legit. if it's a discord server invite, it's not here

https://sandner.art/stable-diffusion-35-large-what-you-need-to-know/ review post

[warm hull]

mb , just some misunderstanding

[warm hull]

Is this your blog post ?

[desert dagger]

nope.

[desert dagger]

https://x.com/poe_platform/status/1849831370866164136 Stable Diffusion 3.5 Large is now on Poe

[broken field]

Sup everyone?

[fervent thunder]

thanks this is great

yeah I wanna put vision models throughout my comfy workflows to have them judge how the image is doing LOL

[warm hull]

Guys

I want to deploy a custom SD model(Juggernaut) on runpod. Is there a template i can start with ?

[fervent thunder]

I don't know runpod but on vast the default comfy ones work well

[fervent thunder]

you also save at least 30% by using vast

[warm hull]

I already loaded up credits on runpod. You think it's worth it to make the change to vast ?

[fervent thunder]

yeah

[warm hull]

Any reason why you recommended comfy over automatic111 @fervent thunder

[young bronze]

comfy is more complex but it is more powerful, and while it might be hard to start from scratch comfy supports workflow

[vale steeple]

I bet there is no Forge template in Runpod so yeah if you want to be straightfoward just choose A1111

[warm hull]

For now i want something more straightforward so i might so for A1111

[fervent thunder]

A1111 has maybe 0.01% of the features of Comfy

[warm hull]

On vast can i build custom models(juggernaut) images or i have to build it locally and push to vast ?

[warm hull]

I already have a live website with users so i'm looking for something that is quick and easy to setup , but i'll definately keep comfy in mind

[vale steeple]

I am pretty certain Vast have a Stable Diffusion template. I am not sure how you supposed to push the model onto the SD webui itself but well you gonna have to do that.

Hugging Face may be useful

[fervent thunder]

also you barely need a template

there will be swift install instructions for any of these UIs on an empty server

[warm hull]

Thanks. Have you ever tried pushing a custom sd model into a SD webui ?

[vale steeple]

I never use a rental GPU before

so I have no idea

[warm hull]

I'm kinda new to all this so i dont understand what you mean here

[warm hull]

Got it. Thanks anyways

[fervent thunder]

you can install on empty server

or on a basic template that only gives pytorch

A1111 and Comfy UI are just software like Google Chrome or Microsoft Excel

you don't need special setup

[gilded sand]

hello

[hard swift]

gm

[stuck gorge]

hello

[blissful aspen]

Hey guys, sorry for the stupid question. But I don't get the SD updates (2, 2.1, 3, 3.5..) I was using 1.5 a year ago with better and faster results, than 3.5 is doing right now. Why would I use a bigger/slower model than? I mainly use it for pictures of woman (nude and non-nude)

[desert dagger]

they're different models, different training. what i suggest is that you run the same prompt in each of them and compare your results

[oblique elk]

Well i would say that you don't need any face or hand restore, you can add text etc. so the base result follow the prompt better, ... Still it is possible to achive great images (lora, ipadapter, fine tuned models,...) with sd 1.5

[blissful aspen]

Thanks, I will try that! But I should use a different sampler for the different models, right?

[desert dagger]

not if you want to see what each does and compare results. you want everything as identical as possible.

[blissful aspen]

What about the seed?

[fervent thunder]

keeping seed same would be best yeah

I think most people don't, when comparing models, but it would actually help

[desert dagger]

you're using different models. they ahve different latent spaces. seed wont' have any effect

what you want to see is render quality, prompt comprehension, etc

[fervent thunder]

the seed has an effect its just really small

with default ksamplers

if you used Sharksampler with fractal noise and alpha = 3 then the effect is big

[desert dagger]

totally different latent spaces, and really what she needs to see are things that the seed won't do much even in the same latent space

[fervent thunder]

they are different latent spaces yeah so they will interpret the seed differently
but what you can find is that a certain seed will tend to force there to be an object in a certain location, if the noise has a feature there
or force a certain colour in a certain area, if the noise has a colour bias there
this effect is small, but still present for the default high frequency gaussian noise that is the default in ksamplers
if another type of noise is used, that is more low frequency (more lumpy) then the effect can be large

[desert dagger]

yes, but for what she's doing, trying to see the major differences between each model and figure out which works best for what she generates, that's a lot of work that's probalby not necessary

[fervent thunder]

yeah that's true its probably overkill
most people don't fix seed when doing comparisons like this

[desert dagger]

@fervent thunder @shell tendon so this happened https://x.com/abursuc/status/1849945966536950080

[fervent thunder]

CLIP, ViT, DINOv2, SAM and Depth-Anything in one model wow

that's really cool

[desert dagger]

i thought so :)

[fervent thunder]

interesting that the highest performing combo didn't need the depth sensing

makes sense I guess cos you can get depth info out of the ViT

[desert dagger]

you gonna play around with it?

[fervent thunder]

yeah I wanna start doing more vision stuff

at the moment I mostly just use clipseg, DINO+SAM and then my yolo fine tunes

oh florence2+SAM rather than DINO+SAM these days

[desert dagger]

:) too many toys

[fervent thunder]

literally yeah

[leaden coral]

Hello

[alpine mantle]

so im a complete noob when it comes to SD i just downloaded it and im trying to figure out where to put model files i downloaded, its not a .ckpt and nothing in the file name says where to put it. anyone know?

[ivory sonnet]

@alpine mantle Kindly contact the support, they should be able to help you from there

[desert dagger]

best to ask technical questions in #🤝｜tech-support

[alpine mantle]

alright thanks

[near silo]

Genuine question... Do people actually use Artisan?

[desert dagger]

sure

people were generating in the channels here all day today

[fading parcel]

I just started a cyberpunk 2077 fan art community on Reddit https://www.reddit.com/r/CyberpunkFanArt/ It's also for the cyberpunk genre in general. It's new so nothing is posted yet, but since there are few places to share AI art where it isn't reviled and derided, I thought I'd promote it here.

[oblique elk]

https://openart.ai/workflows/all?keyword=pulid

[barren nymph]

Hello everyone, does one of you have an idea to transform the drawings into photos while keeping all the details of the image? We find a lot of things to go from photography to drawing but not from drawing to photography. Especially keeping the details

[oblique elk]

Depends on what you would like to achive. Image to image works, you can use controlnet to keep depth or lines accurate, you could use ipadapter style transfer with a realistic image as source for the style...

[tropic stone]

hey i'm new to this server - Anybody here from india? let's connect

[torn hedge]

can someone please explain to me what the hell this is? The upload of recordings is not working on Stable audio! I keep writing to support to get them to do something about it but I guess they don't give a shit if they have our money! Why the fuck isn't it working!!!!!!!!

[unborn hedge]

i forgot to resize my training data so its all at 512 x 512...dangit, i needed it to be 1024 x 1024 because i heard that captures the details better..

[brave matrix]

Is flux 1.1 inpaiting availble anywhere?

[barren nymph]

Thanks for your answer, I had started to move in this direction but I didn't get there, the tools are not so easy for a beginner, I will insist;)

[desert dagger]

you can't just resize the images - you'll wind up with the same data just stretched over a larger area

[desert dagger]

how about some more details of what sort of recordings you're trying to upload, what errors you are getting, and so on

[wicked monolith]

Anyone using the Krita plugin?

[signal dove]

Ok

[azure drum]

I have question. Am I violating copyright if i use AI to create image of character from a movie or game without commercial use and without publication anywhere?

[unborn hedge]

i couldn't understand comfyUI so im going back to forge, this time im playing around with the quantized version of flux which should be fun and shouldn't blow up my computer this time

[desert dagger]

you'd have to ask the company that owns the intellectual property of that character - the issue isn't the tool you used to create it with, the issue is the character itself.

[unborn hedge]

oh hey i was able to load a lora this time without my machine exploding lol, looks like the quantized version worked!

[knotty stream]

guys help

1 picture (inpainting with "segment anything") is taking forever. is 1.5mb/s normal for something that needs 3GB?

diffusion_pytorch_model.safetensors: 10%|███▉ | 336M/3.46G [02:03<19:30, 2.67MB/s]
diffusion_pytorch_model.safetensors: 39%|███████████████▍ | 1.34G/3.46G [12:22<09:01, 3.92MB/s]

[desert dagger]

i would just let it finish

[knotty stream]

yea but if i want to make more than 1 picture every 6 hours?

[broken smelt]

pretty sure youre downloading a model there

you'd only download that once

[somber orchid]

can somone help my I cant seem to find other ways to run stable diffusion 3.5 only through comfy?

[desert dagger]

it's only been out a few days. give the other interfaces a chance to code in support

[somber orchid]

ah thanks

[pulsar marsh]

How viable is local generation with an AMD GPU?

Anyone got one and had success with it?

[main zenith]

boys

yall think is feasible to use SD 3.5 comfortably with a RTX 3060 on comfy?

or better to upgrade to a 3090

[desert dagger]

go into the #🤝｜tech-support channel and read the AMD guides pinned in it. follow them, and you'll have quite a bit of success

[desert dagger]

depends on if you want to use the quant versions or not

[pulsar marsh]

tyty

[main zenith]

Something with better quality and results than sdxl

[keen birch]

after formatting my pc and reinstalling windows im currently backing up all my models and stuff from my previous server installs and planning to do a clean install of a server, has anyone used the ux ui forge server

https://github.com/anapnoe/stable-diffusion-webui-ux-forge
this one rather

or is auto111 recommended for stability and compatibility

[keen birch]

kinda just want the most basic one for making cute pictures

[kindred lark]

eeellooo

[tulip beacon]

Hey Guys, what is currently the best local ChatGPT like model i can host locally with Ollama? To do the stuff like ChatGPT can, without generating Pictures and T2S or S2T

[full lark]

It depends mostly on how much vram do you have

For example I have 12gb of vram and the best model I can run is gemma2 9b or mistral nemo 12b

[warm junco]

Auto1111 would be the basic one.
But if you need flux or sd3.5 support you need forge

[tulip beacon]

I want to run it on a Dell Server, sadly no graphics card

[full lark]

You can run it on CPU, it will be way slower but it will run, how much RAM does it have?

[warm junco]

Llama 3.1 is really good or gemma or mistral nemo/orca

[tulip beacon]

gave the vm 8 haha but i have like 128gb

[full lark]

Nice!

[tulip beacon]

i have a intel xeon e5-2670 v3 2.30 ghz. I want it as a business solution for our company so the people can test it out, maybe work more efficient. Cause we blocked chatgpt natively cause of data protection reasons. maybe that helps you for the recommendations. 🙂

[full lark]

It would probably run gemma 2 9b at a decent speed, you should test it first in my opinion, then you could try with the 27b version

Sadly I haven´t tested bigger models, llama 3.1 8b should work fine too

[tulip beacon]

mhh okay is there a big difference between them in case of features? they should work with internet search too later on. Idk what it takes to make that happen i'm pretty new but yeah that should work

And it would be nice if i could restrict the model, cause its beeing used in business environment

[warm junco]

idk about the internet part but with ollama you cant connect it to the internet search.
but checkout this, maybe they have more features:
https://anythingllm.com/

they have options to run stuff localy or cloud and you can change between different services

and pdf support is nice

[tulip beacon]

oh so you can upload pdf so it can work with it? nice

[oblique elk]

https://github.com/open-webui/open-webui
together with ollama and a good vector db/framework works also with PDF, internet search and even images if you host a comfyui anywhere. You could also use a ChatGPT entpoint if you can make sure to get one executed within a save azure environment for example.

[tulip beacon]

it seems that the models seem to be the problem in case of internet not ollama itself

but i think i will try llama first to get in touch with local ai

thx guys

[warm junco]

The models are not the problem as they are local files that cant do anything when not used with a tool like ollama or Anythingllm

Its the implementation

[tulip beacon]

i read there are plugins that make llama in ollama internet capable, that true?

[lusty shore]

Hi

[tulip beacon]

"Ollama now supports tool calling with popular models such as Llama 3.1. This enables a model to answer a given prompt using tool(s) it knows about, making it possible for models to perform more complex tasks or interact with the outside world.

Example tools include:

Functions and APIs
Web browsing
Code interpreter
much more!"

this may be what i need

[warm roost]

halo

能讲中文吗

你说的是sd吗

[fervent thunder]

Hi everyone! I have a passion for reading and love diving into all sorts of topics—I'm an explorer at heart! Looking forward to connecting, learning, and sharing ideas with all of you.

[desert dagger]

if you mean generate here on this discord, you do that in the artisan channels. start by reading the information in this channel #artisan-faq

[bleak matrix]

Hello everyone! How are you all doing today?

[brave matrix]

@warm junco Do you know if flux 1.1 is available anywhere with inpainting?

[warm junco]

Dont know sry

[hollow gorge]

Im making a discord bot rn that can do that, I havent seen it anywhere online though

[fervent thunder]

Anyone using stable diffusion for architectural design? How to go about “sketch to render”?

[main zenith]

boys can the rtx 3060 run SD 3.5

does it run better than flux dev GGUF Q8

[desert dagger]

how much VRAM?

[main zenith]

12

[desert dagger]

better is a compareson word and doesn't mean anything.

[desert dagger]

it'll be slow. you might want to use https://civitai.com/models/879701/stable-diffusion-35-fp8-models-sd35

[main zenith]

does it require better hardware than flux

[desert dagger]

which will take less vram

the GGUFs are here https://huggingface.co/city96/stable-diffusion-3.5-large-gguf/tree/main

[main zenith]

But does it run faster than flux dev

Both quantised versions in comparison

What’s the difference in performance

[fervent thunder]

depends on if you are generating a negative with dev

if you are generating a negative with dev then SD 3.5L is faster, otherwise flux is faster

[desert dagger]

depends on your hardware too. but flux is frozen and has a very narrow range of what it can do. stick to that range, you get very nice images. try to go outside that range, it won't let you

[halcyon quiver]

hello everyone. Been a while since ive touched base here. I have a question and i hope it makes sense. has anything cropped up making it easier to run a locally hosted stable diffusion install? and im not talking about on my PC. I want to be able to install it on linux server on my network and access it from my pc, maybe witha web interface.

[fervent thunder]

civitai is lagging?

[desert dagger]

have you trued using comfyUI?

or swarm?

[lilac torrent]

家人们，我找不到SD的bot了，有朋友知道在哪吗呜呜

怎么把SD添加到自己的服务器呀

[halcyon quiver]

no i havent i will take a look at that, thank you so much

[desert dagger]

运行或开发所需的一切都在这里： https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main

或下载模型 https://civitai.com/models/878387/stable-diffusion-35-large

[void fossil]

nvida and mit release this new diffusion model, github coming soon but you guys can try it out 🙂
https://hanlab.mit.edu/projects/sana

[tulip tangle]

What is the recommended ram and vram for sd 3.5 large

[desert dagger]

It needs about 16gig vram

Whats your prompt

[tulip tangle]

For FP 16 ? And also what is the loss going to fp 16 from fp 32

[desert dagger]

sure

[desert dagger]

where did you find an fp32?

[tulip tangle]

So there is no fp 32 version of sd 3.5 ... okay ... then what is the quality loss between fp 16 and fp 8 ? Can 4060 ti 16 gb run sd 3.5 fp 16 effectively ?

[desert dagger]

that's what i have and i have no issues running the large model exactly as SAI relesaed it

runs fast, on my machine

unless i stupidly kick off 10 image batches without remembering i had it set for 10

[tulip tangle]

Ooo okay got it ... u cleared a lot of my inquiries

[dusty trellis]

Why do some LoRAs require trigger words and others don't? Is it just up to the creator or is there a useful purpose?

[fervent thunder]

lora training is still the wild west
there are lots of different techniques and it is unclear which is best

[kindred isle]

What is the maximum number of tokens that can be used to write a prompt with Stable Diffusion 3.5 ?

[lone hawk]

512

[vapid quartz]

Hello!

[kindred isle]

@lone hawk how did you determine this information ? thanks

[tulip beacon]

Hey guys. So there seems to be a problem with Ollama, VM and Cores, right? I have an VM in Proxmox with Ollama, WebUi and Llama3.2 3b. I gave it 64gb RAM and 20 cpu cores. but it does not take more than 50% and its slow af.

[tulip tangle]

When's sd 3.5 medium going to be released ? It's 29th october

[fervent valve]

lets wait for a little longer

[sweet token]

Mods. Remove that guy's balls

[fervent thunder]

yokoso watashi no soul society

[dapper bridge]

SD 3.5M is out!
https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/

[mystic siren]

setttings to use ultimate upscale with sd3.5 medium?

[pine fiber]

dead on release model

[austere sky]

"9.9 ONLY"

[young bronze]

9.9 GB? aww man I'm 8GB

[nocturne gull]

wow, that's a whole new level of "out of touch with reality"

[rapid tiger]

Did you really just @ everyone?

[compact swan]

I only have 8 GB

JuggernautXL it is for me...

[woeful knoll]

2024
wanting to use SD on 8gb vram
why are people like this

[warm junco]

GGuf models can be smaller for the 8gb cards

[torn wadi]

Do we have access to the api to use the new medium ?

[somber orchid]

why do people put out things and its not even accessable to eveyone?

[kindred jetty]

Yeah same
"Most consumer GPUs"
Lemme doubt that lol

[worthy tree]

How r v supposed to catch it if it runs out of the box?

[gloomy cedar]

cant even access the model

[fervent thunder]

I want to point out the graph of which GPU is compatible is misleading, mostly because of NVIDIA

there are two 4070's with different VRAM capacities

[somber orchid]

will this even work with amd?

[peak jungle]

1.5 still best for 8gb VRAM?

[teal wasp]

I had the exact same though. They should've said "many" instead to be more accurate.

I'm also restricted to 8GBs, but I'm sure there's optimizer peeps out there that'll squeeze it down further.

[manic wraith]

I'm sorry but, isn't the graphic misleading? Flux can certainly be ran on 24GB cards

[manic wraith]

probably, but see if you can run SDXL/Pony, I think you can

(worthy of note is that, despite the name, Pony can do anime and even realism with checkpoints like "pony realism" or "damn! pony realistic")

[somber orchid]

you need to edit it to run 8 gigs of vram

[manic wraith]

you mean get a pruned model or something?

dunno if quantizing is a thing but it maybe is

with 24GB I just run 6 images in parallel so I assume you can do 1 with 12GB

[eager shard]

9.9vram is insanity considering my consumer grade 3080 has 0.1 more than that

Is sdxl still the best or nah haven't been active here much the last year

[manic wraith]

RTX 3080 has 12GB

[eager shard]

No, I literally have one

[warm junco]

Yes

[gloomy cedar]

depends on which one it is

[eager shard]

10gb

Ah there's multiple

[somber orchid]

thanks

[keen badge]

anyone know offhand how much vram sd 3.5 large requires?

[manic wraith]

is that a mobile variant?

[chrome herald]

6GB even was always totally fine for XL models already in Comfy, if the card is Nvidia and of Turing or later architecture

[eager shard]

No it's just a normal 3080

Before EVGA stopped making gpus

[gloomy cedar]

Such a painful loss to the PC world

[chrome herald]

3.5 Medium should have zero issues on 8GB VRAM + 16GB system ram if you run FP8 T5 or even better the Q8_0 T5 GGUF

[eager shard]

Is 10 vram enough for normal 3.5?

[gloomy cedar]

no because 3.5 medium is normal 3.5

and that uses 9.9GB vram

[eager shard]

Then what's the difference I thought medium was for less vram

[chrome herald]

You can use GGUF quants of the actual image model too though for Large

[manic wraith]

10GB 3080 is such a dick move by nvidia, even the default should have been 16GB not 12GB

[eager shard]

I didn't even know they had a 12gb 3080 till right now LMAO

[gloomy cedar]

the 12gb is the standard 3080

[manic wraith]

I hope Intel and AMD get off their asses and stop dreaming they can simply charge 10% less than nvidia

[gloomy cedar]

whats wrong with amd? i heard their latest GPUs are great

[eager shard]

Is it an issue with AMD or just that no one is making software for AI with amd?

[manic wraith]

AMD charges a little less, for a product with a little crappier software stack (a lot of AI stuff and other productive apps expect CUDA)

[fervent thunder]

its CUDA

is the issue

[chrome herald]

It is, just don't use the big chungus full FP16 T5 encoder, use FP8 or Q8 like I said

[eager shard]

Ah so it's just software people not making amd options

[manic wraith]

Geohot (the tinycorp guy that previously jailbroke the iPhone and PS3) offered to fix ROCm if AMD open sourced it, AMD drops the ball very often

AMD steals defeat from the jaws of victory (only on the GPU front, the CPU front is doing great)

[warm junco]

ROCm is open source

[sonic wave]

btw what do all those abbreviations stand for?
FP - floating point
BF - alternative floating point format with more bits allocated to exponent rather than mantissa
that is what I know due to limited programming experience I've had
but Q?
is it just Quant? what format is this?

[eager shard]

I thought AMD was the good option for low end and Nvidia was best high end

[chrome herald]

refusing to acknowledge the existence of quants for both the image models themselves and the text encoder
why are people like this

[manic wraith]

the driver side isn't I think?

[manic wraith]

if buying new, yes! but you can get used Nvidia stuff for better prices because the market is so full of it

[woeful knoll]

womp womp

[warm junco]

Should be all open source

Only Adrenalin for Windows not

[vernal ore]

Honestly they should improve their autoencoder

[eager shard]

I don't think I'd want anything pre3000 generation vs a new amd card 😭

[vernal ore]

Yknow how much faster stable diffusion could be if they just adapted an autoencoder that could compress more

[fervent thunder]

yeah this idea is called Sana

its coming soon

quality is a bit hmm

[vernal ore]

I'd send an example right now but I can't

[manic wraith]

looked it up and it seems they actually open-sourced and fixed the thing geohot was complaining about, Lisa Su even responded to him, nice!

[vernal ore]

There's this open source autoencoder that can compress 64 fold and has a better image FID than Stabke diffusion's 8x vae

[fervent thunder]

I don't doubt that the autoencoder does that well
but I doubt the diffusion model trained using that autoencoder would do well

[manic wraith]

no, I mean you can get an used RTX 3090 for the price of a new 7900 XTX, in that case the Nvidia one makes more sense

[vernal ore]

#🏞｜general-with-images message

[fervent thunder]

oh yeah this is Sana

[vernal ore]

They also made a diffusion model

[vernal ore]

It is Sana?

[fervent thunder]

it doesn't look great though I've made over 100 Sana images

yeah this is Sana

[vernal ore]

Can you send me a link?

[eager shard]

Oh really? Albeit I haven't looked at gpus in 4 months last I checked a 3090 used was $1500

[gloomy cedar]

they dropped to the 1000s

[vernal ore]

Nvidia works with ai more reliably

So I'd rather pick that

Unless you're just gonna generate images using stable diffusion and game

[fervent thunder]

https://github.com/NVlabs/Sana

demo here https://sana-gen.mit.edu/

[eager shard]

Yeah just checked 3090 is $1200s now, not good but cheaper than a couple months ago while 7900 XTX is $800s

[vernal ore]

Holdon did they make or use DC ae?

[fervent thunder]

yeah

its confusing cos Sana debuted multiple things at once

but Sana is also the DC-AE paper

[eager shard]

7900 XTX for $849 is an amazing deal holy

[vernal ore]

Confusing much

But yeah

The models are only 0.6b and 1.6b

[fervent thunder]

other models could be trained with that VAE though

[vernal ore]

This

[vernal ore]

And sdxl wasn't really that amazing

[hoary gull]

TIL my 3090 is as valuable as it was 2 years ago when I bought it new, nice to know

[ivory rapids]

Yeah, there seems to be massive shortage in some places. I got mine used for 700e.

[eager shard]

yeah still 1k+ holding its retail value highly

[fervent thunder]

a 12B with DC-AE would be cool

[eager shard]

EU?

[ivory rapids]

Yeah.

[eager shard]

Knew it, all my EU friends got their gpus so cheap

[gloomy cedar]

What are yalls thoughts on 4070 supers btw

[eager shard]

strange it has the super title but i dont see why someone would buy that instead of a 4080

just questioning why that card even exists

[vernal ore]

It's be able to be ran on more hardware and faster

[eager shard]

bro got muted for spamming 😭

[gloomy cedar]

idk maybe because 4080s cost over 1000

whereas 4070 supers dont

[eager shard]

really? i saw a bunch going in the 800s on ebay

like 3 weeks ago

$599 i guess that makes sense

200-300 jump for a used 4080

[vernal ore]

I wanna compress Minecraft images effectively

More than 32 or 64x

Using DC ae, but idk how I'd train that

[eager shard]

efficient and minecraft in the same sentence?! insanity

apparently a new shaders mod came out for sodium that has the most efficient ray tracing for minecraft as of yet

Photonics

[gloomy cedar]

ray tracing in minecraft yuck

[eager shard]

i love rethinking voxels

[vernal ore]

I'm not bothered about shaders

Just wanna generate Minecraft video

[eager shard]

Apparently it was tested on exclusively amd gpus

https://modrinth.com/mod/photonics

[keen badge]

dont mind me, i still love my refurbished 1080 ti 11 gb xD

[vernal ore]

I wanna get a new phone

[hoary gull]

Has anyone managed to generate over one megapixel images with 3.5 large? For me everything above that turns to mush. Can't wait to test 3.5 medium later today as it should support up to 2MP

[gloomy cedar]

what are you generating on?

[eager shard]

i got a 1080 in my backup pc 😭 poor thing only being used for my media server

[vernal ore]

Ah yes
Stable diffusion 🍷

Or maybe something about whatever you did is somehow wrong

That's just how ai works

[hoary gull]

Not sure what you're asking, rtx 3090, ComfyUI

[gloomy cedar]

i meant ui

thanks for answering

[hoary gull]

It's very highly I'm doing something wrong, just recently moved from A1111, though I've read 3.5 large might only support up to 1MP

[cedar salmon]

that would be correct

[eager shard]

i like swarmui!

basically comfyui but with an option to use simple stuff similar to forge/A1111 for when you cant be bothered

[hoary gull]

I had my own UI before all the ControlNet stuff etc. became mainstream. Didn't bother to study how those work in code so just moved to A1111 and now to Comfy, though I'm going to go full circle and make a new UI that uses Comfy API 😅

[eager shard]

i did the same exact switch

except after comfy i switched to forge and now on swarmui

[vapid tangle]

Where can we try stable diffusion medium 3.5 online for free?

[sand flax]

Does Stable Diffusion Medium 3.5 have photoshopping functionalities like DALL E 3?

[rare holly]

Hey there having problems with the Web based stable diffusion all iam getting in grey images what am I doing wrong

[manic wraith]

a used 3090 costs less than a 7900 xtx if you shop around, but at the very least the same

if they cost the same it's a no-brainer, if they cost less it's even moreso

[eager shard]

really? if they're both the same why would i get a 3090 instead of a 7900 xtx

[manic wraith]

yeah, you might wanna check where is the place to buy used stuff in your country, dunno about the US, I think facebook marketplace?? but of course double check

[eager shard]

ebay primarily, facebook is better for lower end atleast in my area

[manic wraith]

seems ebay prices are ~$690 (nice)

[eager shard]

Yeah it seems a 3090 is even worse than a normal xt much less a xtx

damn 4090 is amazing though

[manic wraith]

depends on the task, if ROCm is worse than CUDA for that task, it's 24GB

[cedar salmon]

what i was saying

[manic wraith]

I think AMD caught up in SD and LLM stuff but I'm not sure

[eager shard]

yeah im looking at 3090/3090ti/4090/7900xtx/7900xt comparisons

[manic wraith]

either way it sucks balls that a lot of projects are hard-wired to CUDA and as an end-user you can't do much about it

(aside from going into the code and making changes yourself... not fun)

[eager shard]

i wonder if a xtx is either a 4080 or a 4080 super

xt seems to essentially be a 3090 ti

[manic wraith]

depending on what your task is, 7900 XTX ~= RTX 3090

but look for people posting metrics that are relevant to what you want

like tokens per second, iterations per second, etc

[eager shard]

ok yeah an xtx is essentially a 4080 super

im looking at overall not anything in particular

majority of situations

[warm junco]

I have a 7900xtx, great card for Gaming.
For SD on windows it works good with ZLUDA and hip.
For Linux you can use ROCm

[wraith rover]

/credits

How do you check credits

[vernal ore]

Years ago i trained an ai to generate Minecraft like images

I wish i could do video