🆕｜sd3

[dull star]

not a huge influence and haven't tested extensively

[cunning lintel]

Gives hope for finetunes too 🙂 it probably knows the look not how to prompt it

[dull star]

hope so!

[icy drift]

Great prompt adherence. (And good enough quality at 20 steps.)
In this RAW photo, a cinematic movie scene, a full-body shot of a woman standing in the street wearing black boots, a woman is wearing a blue trenchcoat with a red belt. She has a green cowboy hat on her head, and long blonde hair. She has tall black leather boots. She is facing the camera and smiling.

[dull star]

sometimes its just too much photo-like influence

its like it starts with a perfectly photorealistic anatomy and then just turns it into a paintings

[icy drift]

Definitely the same woman, which is impressive, but it changed her clothes. I guess I could just copy her description from the original prompt though.

[dull star]

in the preview I just see a photo

its like the VAE adding in the painting-like details

[icy drift]

Probably need to train a lora to get the style to come out stronger.

[dull star]

but SD3 API had pretty good paitnings by default

now its basically as good as flux

[icy drift]

Very impressed with the quality here.

[dull star]

you need a relatively low cfg

and even then it looks like a photo turned into a painting

[sage burrow]

https://huggingface.co/spaces/Shitao/OmniGen
Though I don't seem to be getting the img2img prompt correct or something

[icy drift]

It was broken earlier today. Don't know if they fixed it.

I usually do physics tests and stuff, but I don't think that applies here. I really just want this for reposing characters, and I have zero tests designed for that. This is a whole new world.

Okay I'm going to sleep. Super happy to have this working though. I'll try to think up some tests for it.

[craggy crest]

just in case you missed this https://huggingface.co/city96/stable-diffusion-3.5-large-gguf/tree/main

[cunning lintel]

prompt a bit on 3.0 api and then train for 3.5 local, that's SAI's clever monetization strategy 🤑

but all jokes asside having seen 3.5 a bit and used 3.0, i really don't understand in what way 3.0 was not good enough for release for nearly half a year while 3.5 is

[sage burrow]

3.5 definitely has a LOT more data/knowledge than 3.0!

[cunning lintel]

Haven't noticed it yet, my imagination is lacking i guess 🙂

3.5 thinks a mink is a cat while 3.0 knows a mink's a mink, but thats my n=1 observation, whatever tagger they used didn't know mink 😢

[sage burrow]

My fave things to prompt are scenes from the middle ages and Rennaisance. Well and various 70's cartoons lol

Unfortunately 3.5 is really lacking when it comes to anthropomorphic snakes though! 😦

[cunning lintel]

hmmm

made me try hydra

[rapid pivot]

https://tenor.com/view/sable-cute-dog-tiny-puppy-gif-26160652

It's furry, has a nose, 2 eyes and cute ears

Basically a cat?

[cunning lintel]

something went wrong 🤣

A majestic, powerful hydra warrior, with a strong, muscular torso, its chest and abdomen a mass of writhing, serpentine flesh, from which nine thick, scaly necks emerge, each supporting a serpentine head, each with a strong, angular face, sharp fangs, and piercing emerald green eyes, its scaly bodies a mesmerizing mosaic of dark blues and greens, with iridescent scales that shimmer in the dim light, its necks thick and muscular, adorned with intricate, swirling patterns of silver and gold, wearing intricately designed silver armor adorned with golden accents, with each head crowned with a delicate, gemstone-encrusted circlet, the heads weaving and twisting around each other in a mesmerizing dance, as if alive, stands defiantly in the midst of a rugged, rocky mountain landscape, as it battles a legion of dark, gaunt, and twisted undead orcs, their emaciated bodies clad in tattered, blackened armor, with glowing red eyes, razor-sharp teeth, and grasping bony hands, amidst a torrential downpour and flashes of electric blue lightning that illuminate the dark, foreboding sky, as the warrior's nine swords, each emblazoned with a golden hydra emblem, shine with a warm, heroic light, painted in a highly detailed, epic style reminiscent of Olga Shvartsur and Svetlana Novikova.

[rapid pivot]

@sage burrow are hydras possible yet

[cunning lintel]

its way more interesting than a cat!

[rapid pivot]

How dare you

Do they have other names than mink that you can try

[cunning lintel]

your chaos bot now, did cat upgrade vram and something went wrong?

[rapid pivot]

it exploded

[cunning lintel]

poor catbot 😦

[sage burrow]

I haven't tried, hold on

[cunning lintel]

deepl says a mink is a mullet too, sd3.5 says a mullet is ...

Hyper-realistic mullet detective, razor-sharp quills glistening, piercing eyes reflecting gaslight, oversized trench coat and deerstalker hat with intricate fabric textures. Cobblestone Victorian London street, fog-diffused lamplight, condensation on wrought-iron railings. . Cinematic composition, anamorphic lens flare, golden hour backlighting silhouetting the mullet's profile. Atmospheric perspective, volumetric fog, ray-traced reflections on wet surfaces. Shot evokes Joel Meyerowitz's street photography, captured on RED Monstro 8K VV at ISO 100, f/2.8, 1/125s. Extreme depth of field, tack-sharp focus on mink whiskers. Color grading mimics Kodak Vision3 500T film stock. Photogrammetry-level architectural details, physically-based rendering of materials.

[sage burrow]

a mink lol

arg, tails!! ai hates tails lol

[rapid pivot]

Ban becky from mink genning

[cunning lintel]

i need to make them bodybuilders not detectives 🙂

[rapid pivot]

[sage burrow]

another mink lolol

[rapid pivot]

Mincoon

[sage burrow]

Perhaps. The trenchcoat gives the ai an excuse to cover up part of the face

[pseudo owl]

Mochi-1 has no idea what mink is, pretty cool detective tho.

[sage burrow]

one bad attempt at a hydra

[rapid pivot]

I regret asking anything

[sage burrow]

wow, that is a bad attempt at a hydra

This one was via glif, with some help from Claude! 🙂

[cunning lintel]

it's something for sure....

[sage burrow]

it has 3 heads, so definitely a hydra

I'll have to try to remember how to add the LLM assist to Comfy workflows again!

Has anyone found the perfect 3.5 propt length? Before the ai gets bored and starts forgetting things?

[dull star]

this looks really good

is this the 480p version, cause it looks too good for that

[bitter hearth]

were you able to install it locally? i see the install process take a few lines..

git clone https://github.com/staoxiao/OmniGen.git
cd OmniGen
pip install -e .

[oblique parcel]

firefox or something

[carmine vessel]

That's beautiful

Is that with 3.5?

[oblique parcel]

ye

[cunning lintel]

Hyper-realistic anthropomorphic bat magician, performing a card trick in a smoky, dimly lit nightclub. His eyes gleam with mischief as he manipulates the cards with his paws. Cinematic composition, shallow depth of field, focus on the bat's hands and the cards. Shot evokes the captivating mystery of a magic show, captured on RED Dragon 6K at ISO 800, f/1.4, 1/60s. Color grading mimics a vibrant, theatrical atmosphere. Photogrammetry-level nightclub details, physically-based rendering of materials.

[carmine vessel]

Would you care sharing prompt/workflow? I like this, I want to try it with other animals

[oblique parcel]

does discord destroy the workflow from comfy images? it should be in there if discord doesnt
but otherwise the metadata should be in there because i saved it with sd prompt saver node

i used 3.5 q8 a photo realistic scene of a white background and floating magical fox with magical features and the image has magical effects in it. the fox is curled up like the firefox logo with its burning body and the blue-purple flames neg: blown out colors (remnant from old image, too lazy to remove, probably does something)

[carmine vessel]

I honestly have no idea how metadata works haha

[oblique parcel]

IF discord doesnt mess with comfy workflow metadata you should be able to just save the image and drag and drop it into comfy
otherwise you can try putting it in here https://www.sdimage.info/

i actually dont know how to quickly export a comfy workflow but thats what im trying to do rn

ok i downloaded my own image, theres no metadata, maybe i uploaded it wrong or discord destroyed it

[pseudo owl]

Yes 480p version, no hd version out yet. Even the official website uses 480p version.

[dull star]

oh wow the 480p version could make such an interesting composition and screenplay

thank you

[oblique parcel]

ok that worked, it was my upload method

[pseudo owl]

Yeah it’s pretty amazing, some vids with mochi

[dull star]

img2video or video2video would be amazing with this

[pseudo owl]

Yeah an employee said that it should come soon

[dull star]

oh i thought it was img2video that will come out

but video2video too and soon???

[winged seal]

@craggy crest some of my colleagues who have had huge success training flux have ported their data sets over to SD3, and have not been able to get any good results. Instead, they're having to baby it way more, not giving it complex concepts, complex captions, or multi resolution

I haven't had the chance to try it because I've been out on vacation, and will be for the next week, but it is exactly as was expected given the huge difference in knowledge between the two models

[craggy crest]

flux isn't trainable so i'm not sure what you're talking about - however what i see is you deliberately trying to start another fight and i'm not interested in your game

[tranquil viper]

its kind of trainable

[craggy crest]

it's not. you can create small models that will give you 'something' - but you can't train it

[tranquil viper]

the model architecture responds pretty well to training but the base model has been finetuned way too much on ultra realistic images

so yeah its effectively not trainable unless we get a new base model which is probably not happening

[craggy crest]

however if somene actually took the time to use a good data set and the 3.5 trainer that is already out, they would have gotten a good result. so i'm not sure what he's up to, but i'm not interested in it

[tranquil viper]

what is the 3.5 trainer?

[craggy crest]

AI toolbox already has a trainer out

[tranquil viper]

there is already a sd3.5 trainer?

[winged seal]

What do you mean flux isn't trainable? There are hundreds of LoRA's out, I've trained my own. I've trained over a thousand for SDXL, over 200 for SD 1.5, and flux trains astronomically better than those two

[tranquil viper]

oh ai toolkit?

[craggy crest]

i thought you said you were on vacation? you have a commercial interest in flux, you admitted that yesterday. i have nothing more to say to you

[winged seal]

I am, what does that have to do with anything? I've been doing local training for 2 years

I also have a mobile workstation, I can train and have plans too. I'm just not going to be diving into SD 3.5 right now, because it's just going to be a whole pain in the ass to set up, and I don't think it's going to be worth it

[tranquil viper]

from what ive heard trying to get decent 2d images out of flux is a fools errand

[winged seal]

2D images as in?

[tranquil viper]

anything except realistic

[pseudo owl]

Sd3.5 should train better then flux as it’s non distilled, it should have more “knowledge” as well.

[winged seal]

I'm sure it doesn't in the long run, but Flux is a way smarter model, so it grabs onto concepts significantly faster

[tranquil viper]

flux 1 dev is distilled?

[craggy crest]

it's rather telling that everyone on the internet, other than this guy, dropped what they were doing with flux - and had been fighting with it for, for months, and dove into 3.5 the second it was released and ahve released all sorts of things for it in ... 3 days

[tranquil viper]

i thought only schnell was distilled

thats rough

[winged seal]

Dev is as well, from a significantly larger model. But you can effortlessly train flux. Hundreds of people have, myself included, I have no idea why people act like you can't

[tranquil viper]

honestly i dont think anyone is going back to flux anymore

[craggy crest]

pro is only api. dev is distilled and frozen, basically a huge lora, an schell is modified from dev

[tranquil viper]

so thats why training it is such a pain in the ass

[winged seal]

Why do you think that? I'm just genuinely curious to know, because I don't see any major benefits to SD 3.5 just yet. Maybe I'll change my mind when I get to train it myself, but flux is incredibly easy to train

[craggy crest]

you have a commercial interst in flux, you said that yesterday. you ahve ulterior motives and want to see it do well. that's fine. but trying to diss 3.5 and make flux look good by doing so is not going to work. the entire internet has already had a chance to play with both

[winged seal]

Wait a minute, I'm curious. You guys who say that you've had issues with training it, have you tried training it with Kohya? Because I trained about 40 LoRA's using Kohya before switching to AIT. The moment I switch to AIT, my results got a thousand times better

[tranquil viper]

just trying to be objective, from the limited messing around ive done with 3.5 it seems to be mostly better, non-distilled, more easily trainable, and lower hardware requirements

[winged seal]

I'm not dissing SD 3.5, I'm sharing actual factual information about it. I'm not trying to use it commercially, I don't know where you got that idea, and I would like to see both of them do good, but I just don't think the SD 3.5 will be viable for a hot moment before people pour massive amounts of compute into it to fix its fundamental issues

[tranquil viper]

also sd3.5 is way more flexible

[pseudo owl]

Flux trains excellently but only for small scale Lora’s, sd3.5 can be trained much further.

[tranquil viper]

i dont think its gonna top sdxl fine tunes for a looong while though

[craggy crest]

flux is frozen and has a very narrow range. you stick with that range, you get nice stuff.

[winged seal]

I don't know about easily more trainable, but I also have gotten effortless trainings with flux. But I do know for a while I couldn't get it to train at all because I was using kohya

[craggy crest]

it's also packed full of women, dogs, fantasy images, and anime cat girls

[pseudo owl]

Everything is correct except better part, flux does have better quality and prompt following. Sd3.5 is for sure more flexible tho.

[tranquil viper]

sd fine tuning has always been pretty easy

[winged seal]

This much I can agree with, that's kind of why I think the SD3.5 will be important, just not anytime soon. It has glaring major fundamental issues in it that are going to take a considerable amount of compute fix first, before you can then properly add in concepts without causing more damage

[tranquil viper]

pretty sure you basically need a 4090 to train flux loras at all

[winged seal]

Yeah, and you need at least 16 GB to train SD 3.5 as well. That argument isn't valid for large. I don't think large is going to be viable at all, but as D 3.5 medium on the other hand, now that I think is going to be a huge deal

[craggy crest]

just beause you can create a small model that will run with it and have some effct on the output of the generation does not mean you have trained flux, you've created something, you haven't actually trained flux. go make, nto a lora, but a full checkpoint

[winged seal]

You can actually full fine-tune flux on 6GB VRAM, and I look forward to seeing if that type of training is adopted to work with SD 3.5 as well

[tranquil viper]

i think theyre probably gonna be mostly interchangable

[pseudo owl]

True but sd3.5 is more flexible currently. One of them is not going to dominate the either, they will be competitive.

[craggy crest]

use all the tools

[tranquil viper]

if flux was not distilled i wouldve been much more hopeful for its future tbh

[craggy crest]

where's your finetune?

[winged seal]

That much is fair, flux does have a really strong DPO tune on top of it, but it only takes about 5 to 10 images of an aesthetic to unlock insane amounts of potential and information that's buried deep inside

[tranquil viper]

i think the moment a model surpasses flux in realistic images everybody is gonna drop it

[craggy crest]

flux wasn't just distilled, it was also frozen. it cant' be changed.

[pseudo owl]

Flux is not limited to realistic images, it can do 2d images pretty decently too.

[tranquil viper]

its funny because libreflux was released like a week before sd3.5 aswel

[winged seal]

SD3 medium, the original one was still hands down the best photographic model I've ever seen, open source or closed source. But it had such monumental flaws baked into it that it was basically useless

SD 3.5 is considerably better than SD3 medium, but it's still nowhere even remotely close to flux when it comes to coherence

It's easy to train a new style or aesthetic into a very smart model, but it's way harder to fix fundamental issues on a model that's not very well trained

[tranquil viper]

even compared to sdxl its pathetic

dont see any point

[craggy crest]

when you use just a single word as a prompt, for example the word apple, what do you expect to get back?

[tranquil viper]

also there are diminishing returns with making models larger because they get harder to fine tune

and flux 1 dev is huuge

[pseudo owl]

Depends on what you mean pathetic, it’s pretty good at what it knows. However it doesn’t seem to know many styles.

I think the main thing is, no model is better at everything. Both have cons and pros, sd3.5 is worse at some things, flux is worse at others.

[craggy crest]

i'ts only that large because of the padding that was added

[pseudo owl]

There is a recently released 8b flux basically similar quality to 12b

https://huggingface.co/Freepik/flux.1-lite-8B-alpha

[winged seal]

I will definitely give an attempt at training SD 3.5. I would love to see it be more viable, but I just don't see as D 3.5 large catching on. It's too close to flux in terms of impracticality, and it's not nearly as good as flux in terms of the solidity of the concepts already trained into it

SD medium on the other hand, now that I could definitely see being a really big deal. It's much smaller, it's much lighter, people will be able to fix it much faster

[craggy crest]

you're already way behind times. i'll let you go read through the internet and look at all the stuff that's already being done and developed for it. quite literally everyone that's been on the flux bandwagon and trying to develop, but you, dropped flux like a hot potato

[pseudo owl]

It’s been like 3 days for training, flux training was pretty horrible in the beginning with bleeding and many issues.

[craggy crest]

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

[winged seal]

That's really just not that true. You're also the person that said that a majority of the community has 24 GB GPUs, which is just blatantly not true and out of touch

[craggy crest]

it's very true

[winged seal]

Is it just as true as you saying that basically everybody has 24 GB GPUs?

[tranquil viper]

true, but sd3.5 is better than flux and will most likely age much better imo

i think flux 1 dev was just a marketing stunt just to get people to pay for pro tbh

not because they wanted to release a good model

[craggy crest]

where the HELL did you get me ever saying that. i have NEVER said that. you ahve said that i did, multiple times. and i am EXTREMELY tired of it

i'm also really tired of you trying to twist thigns, and then fake that you said somethign else

[winged seal]

What did I fake?

[craggy crest]

https://x.com/zeng_wt/status/1848770952139837633

𝐙𝐞𝐧𝐠 💜 (@zeng_wt) on X

Stable Diffusion 3.5 just dropped , and the possibilities are endless.

Dive in with these 8 prompts and unleash your creativity.

1/12

The Big Deal about SD3.5:

➡️Creates EXACTLY what you ask for
➡️Works on regular computers
➡️Free to use (commercial & non commercial!)

I

Twitter · Oct 22, 2024, 4:58 PM

https://x.com/Lykon4072/status/1849449081438933058

Lykon (@Lykon4072) on X

4GB SD3.5L already 👀 https://t.co/MA8G6Leb86

Twitter · Oct 24, 2024, 1:53 PM

https://x.com/BrentLynch/status/1849313205878739197

Brent Lynch (@BrentLynch) on X

Stable Diffusion 3.5 is great for last minute Spooky Season img2vid

Check these out! Used Rubbrbnd for stable Diffusion 3.5 and even though I am a ComfyDummy the walkthrough of the workflow got me through. Used Hailuo Minimax for the video.

SD 3.5 img prompt
Haunting vivid

Twitter · Oct 24, 2024, 4:53 AM

https://x.com/WilliamLamkin/status/1848733856859349191

William Lamkin (@WilliamLamkin) on X

SD 3.5 Large from @StabilityAI is now available! The @Gradio app is live now on @huggingface Spaces

space: https://t.co/xleFJm47kj

Twitter · Oct 22, 2024, 2:31 PM

https://x.com/ErikVoorhees/status/1849595151326269849

Erik Voorhees (@ErikVoorhees) on X

Stable Diffusion 3.5 released two days ago.

Today it's live on https://t.co/m2jsJuCX8k for Pro users

Twitter · Oct 24, 2024, 11:33 PM

https://x.com/Kdawg5000/status/1848849331555574006

👣KᗪᗩᗯG👣 🔶 (@Kdawg5000) on X

I'd say Stable Diffusion 3.5 is good at text. 😮

Twitter · Oct 22, 2024, 10:10 PM

[oblique parcel]

SPACE BURGER

[winged seal]

Oh wow, that's dope. I haven't seen anything like that from the demos so far

[craggy crest]

lots of stuff on the subreddit and on twitter.

where've you been looking for demos?

[winged seal]

I don't look for demos, I don't care about the base as I've said 😅

It's gonna take a ton of work to tune it, so I'm not holding on to anything it has yet

[craggy crest]

there's a ton of stuff posted in this channel, too - and it takes NO work to tune it, that's what people tried to tell you. hwoever, i'm done with your assumptions.

[mortal mesa]

call the manager

[craggy crest]

https://civitai.com/models/879701/stable-diffusion-35-fp8-models-sd35 fp8 model

https://civitai.com/models/880134?modelVersionId=985261 @gusty trail 's 3.5 lora

and his github repo https://github.com/lrzjason/T2ITrainer #🆕｜sd3 message

[winged seal]

They aren't assumptions. I don't know how many times I have to tell you that I work with people who literally train models for a living. I didn't say that it's untrainable, I said that it doesn't accept as diverse data sets as flux. You need to be a lot more gentle with it

I still have plans to train it, and I'm tired of your assumptions that I'm trying to purposely say it's bad, because I'm not. It's just different

[craggy crest]

1. 3.5 just came out 2. i know a lot more about it than you do 3. i've seen you say a lot of thigns that are flat untrue

[winged seal]

Specifically, it falls apart with multi res training. Now that could just be a specific trainer thing, IDK, but multi res training is my hands down favorite thing about Flux, cause you can throw any res images you want, as low as you want and it just handles it

[craggy crest]

and 4. you have a commercial interest in flux and making sure flux stays the leader of the pack so you can sell it

that makes everything you say mean nothing to me

also makes me suspect you work for black forest in some capacity

[winged seal]

Omfg, can you please stop at that? I don't make money off a flux, I'm not selling anything with flux, and I'm not planning on selling anything with flux. My job is to train whatever companies have in house, has nothing to do with models that are procured from other architectures, or models that I've trained that I'm trying to sell them. These trainings are just my hobby

[winged seal]

I don't lmao. I work for run diffusion

[craggy crest]

wonder if he would appreciate your comments in here. i think i'll ask him

[winged seal]

Uh, ok? Go for it I guess lol

It's literally not anything that I haven't shared with him, or that I haven't shared with the people that I work with for model training

[craggy crest]

he likely knows more than you do about 3.5, too

[winged seal]

I know he does, he's my boss lmao

[craggy crest]

@calm zinc you should probably read through this guy's posts in this channel at some point

[winged seal]

¯_(ツ)_/¯

My claims:

1. I don't think that large will be very viable, mainly because it's similar in size to flux, while not offering too many benefits. The compute used on Large will go similarly far on Flux.

2. I think that 3.5 medium will be considerably better due to its much smaller size and ease of training, as well as higher access to more creative individuals who don't have high in computers

3. In my personal training circles, with people who train models for a living, we have found that just repurposing flux data sets for SD 3.5 won't work, mainly because it is much more sensitive to resolutions, and a lot more particular about its captioning style. This doesn't mean that it's not trainable, just that it's not quite as robust to train as flux

4. Teaining an aesthetic onto a model that already understands concepts is a lot easier than training an aesthetic model too understand other concepts

[craggy crest]

he's capable of finding yoru posts and reading them

[winged seal]

My boss is a busy man, he doesn't need a waste his time searching for nothing lol

He's also not the only company I work with, but I don't need to drag them into the mess lol

Anyways

[craggy crest]

worried about what he might see?

[winged seal]

Dude, you're being melodramatic. This is such a non issue you're blowing out of proportion. By all means, slander my messages I have sent here on a billboard for All I care. I haven't said literally anything that would get me in trouble lol

[sage burrow]

Def proof of lack of multi concept bleeding 🙂

[winged seal]

@craggy crest actually here, tell you what. I'll be home for the rest of today, and just for the sake of being able to make my own arguments without being claimed that they're just speculation, I will set up and run some 3.5 trainings. I think that's only fair. And I've already had monumental success with flux, and I have tons of data sets made captioned and prepared for AI toolkit already

[craggy crest]

https://huggingface.co/spaces/stabilityai/stable-diffusion-3.5-large here's the link to his gradio ap

[craggy crest]

https://github.com/lrzjason/T2ITrainer you might also look at his repo

[winged seal]

I'd prefer to not introduce additional variables by using new trainers. I don't want my inexperience with said trainer to taint my view of the model, so I think I'll probably stick with AIT for now, just because I have experience with it

[craggy crest]

not sure how far AI toolkit is, he said yesterday that it was early release and that he was still working on the 3.5 support.

[winged seal]

Fair enough, maybe I'll give it a little extra time then. Contrary to what you think, I don't want to dislike SD3, all I've shared our personal anecdotes from people that I trust and work with

However, I do feel strongly that large is going to be very quickly overshadowed by medium, specifically because medium has the huge edge of being monumentally smaller, easier to train, and way more accessible to a much broader portion of the community, meaning that it will have way more people wanting to put their own touch onto it, and wanting to improve it for their own use

That's one of the main reasons why I don't want to put all my eggs in a basket on large, because I just don't see large being able to be that great when it'll be so much more economic to focus on making medium exceptionally good

It would be really interesting to see if somebody could make a plug-in that can transfer trainings across both models, like the one that was made that could adapt SD 1.5 trainings to SDXL. I know it didn't work that good, but it would still be interesting

[craggy crest]

i'm not going to get into that discussion with you. what i've seen you do is make a lot of assumptions that have no actual research behind them. and i don't care to discuss imagination

#1 rule - use the tools that do the job you need done #2 rule - research, don't assume

[winged seal]

That's all right. If you count talking to half a dozen people who train models for a living, who have not had good results with said model as not having research, that's perfectly fine. That's exactly why I want to do it myself, so I can corroborate their experiences. If I end up being wrong, that's probably even better. I'd prefer the outcome where I and all the people who I work with are wrong, because then that means that large may actually be viable after all

[craggy crest]

absolutely none of those people could hve tried to do any training with it before it was released a couple days ago, and i doubt they've tried much sense

so i'm not real sure where you're getting yoru facts

[winged seal]

Me being wrong is the ideal outcome, because then that means that I can actually start training it. I don't want to be stuck on just one architecture or just one model

[craggy crest]

actually, the idea outcome is you assembling a toolbox with tools you like and that do the job you need done

[winged seal]

The confusing part is is that I've seen so many people make tons of claims that the model is way easier to train than flux, but your argument here is that it hasn't been out long enough for people to understand what it's training process is like, so where does that overlap come from?

I do know that people had access to the model ahead of release, just like with SDXL, but I haven't seen any concrete results from people training it, unless I just have missed them. I haven't really been particularly interested in this model release, so I haven't been actively looking

[craggy crest]

i just gave you a link to a lora on civit. he went and trained it, with his trainer code, in a very short time. the DAY it released

and all the work that is out now has been done in the 3 days SINCE it's release

not before

[winged seal]

Oh, my bad. I have been hauling groceries up an elevator, and I thought what you linked was just to a repository. My bad

[craggy crest]

no, it's his trainer. that's why i suggested you look at it

it might be more useful than ai toolbox

just remember it's also early and he's still working on it

[winged seal]

All good and fair, thanks @craggy crest

[mortal mesa]

tell em about the clips next

[winged seal]

@craggy crest I'm sorry I'm looking back, and I think I'm completely missing it. You said that you linked a LoRA that somebody trained, and I'm not seeing it. Can I please have the link so I can check it out?

I am seeing the link to the trainer, but I'm not seeing a link for a specific LoRA

[craggy crest]

#🆕｜sd3 message

the link to his lora is there, his trainer is right under it

[winged seal]

Thatttsss why I missed it, I was scrolling fast looking for the embeds. Thanks!

@craggy crest One last question, has a generally agreed upon learning rate been decided on for SD 3.5? I don't want to use a bad learning rate and get a bad result and then blame that on the model. I'm trying to set up the trainer remotely right now, and it's going to take a bunch of time

I'm going the AIT route for now, just cause I am experienced with it. If things go horrifically wrong, I'll chalk it up to the trainer being bad, not the model, and if things go fantastic, then absolutely great to hear

[craggy crest]

i haven't seen anything posted from anyone about the learning rate they are personally using, much less seen any agreement on that. i think you're going to need to wait till enough people have tried training it to get that kind of concensus

however @gusty trail is in this discord, you could ask him what he's doing

[winged seal]

Alright then, I will just use 1e-4 as a safe base I guess

[gusty trail]

I am using prodigy for the training

adamw 1e-4 should work fine

[rapid pivot]

"The boss was ferocious today"

[craggy crest]

a story in one act

[craggy crest]

feeling explosive tonight?

[remote holly]

i love

[noble coyote]

SD3.5L Turbo

[icy drift]

I used Pinokio. Trying to install it manually with Conda failed for unknown reasons (generation got stuck at 0/50 steps). Installing with Pinokio was one click and worked flawlessly.

[noble coyote]

Turbo LLM 3.5L

[cinder junco]

Not sure how I feel about SD3.5L. My Flux upscaling workflow seems to work pretty okay for it, and it might be a bit better at giving a photographic look to the image, at least at first glance. However, I feel like the image details tend to be less coherent and more like an archviz Photoshop collage than Flux and I had to increase the second stage denoising to try to combat this.

SD3.5L still doesn't have any fixes for its positional embeddings, and tends to fall apart at the edges of the image (or tiles) even at 1 MP. You can kind of see this in the sky in the attached image. It's subtle, but sticks out to me and is really obvious and distracting in a number of images I've seen others post.

SD3.5L is somehow slower than Flux and seems to hate the custom RK/ODE solvers -- even on a 1MP image, they will sit at very low sigma for several minutes before starting to get going and the output is prone to have blocking artifacts. I had to switch back to DPM++ 2M. Using the beta scheduler seems to approximate the output of bosh3. Just doesn't give me the warm fuzzies, though.

On the other hand, Flux isn't perfect either. It tends to like a cgi look for scifi, is prone to sometimes excessive film grain when upscaled, is less controllable with no negative prompting (though SD3 negatives aren't very effective), and of course there is the bokeh bias. You just can't beat its details, though.

[rapid pivot]

Enhance

👀

[lavish sparrow]

it's song lyrics, yet again xD

Unrequited - Sennzai

[noble coyote]

Turbo LLM 3.5L

[lavish sparrow]

i see the hand without the hand being there, or is that just me?

[lavish sparrow]

Beyond the Horizon - Metalite

[pseudo owl]

This is actually pretty cool, seems like you get better results(sometimes at least) for sd3.5 when you skip certain layers

https://ostris.com/2024/10/23/skipping-sd3-5-large-blocks/#more-319

[bitter hearth]

you could try the new thing, although it is in Diffusers only at the moment

https://github.com/PRIS-CV/I-Max

gets Flux up to 4k

[lavish sparrow]

i love it because of it's smarts 🙂 i don't mind it's a bit lower quality, i love it because it can do more of the prompt right

[bitter hearth]

yeah prompt following is nice

I am not much of a prompter myself but lots of people are

[terse belfry]

Any pics of woman laying on the ground generated with SD3.5 ?

[lavish sparrow]

alcoholism?

sd3.5 can do text well enough imho

[turbid grotto]

generally sd35l must be faster than flux, and could be even faster with 1cfg as for flux

[noble coyote]

Turbo LLM 3.5L

[bitter hearth]

I thought flux was faster
cos SD 3.5L is 2/3 of the size, however it has to generate a negative each step

[pseudo owl]

Yeah the official blog post has a picture of woman laying in grass lol.

[lavish sparrow]

if there was no negative -> sd3.5 would be faster. but not using negatives when you can is not smart imho

[fossil pagoda]

Then give me a neg i can use for everything

[lavish sparrow]

"boto is not in this picture" XD

[bitter hearth]

empty neg

can be used

[lavish sparrow]

the image is a large blob of text, or an inspirational quote this one seems to work pretty well if you're throwing in a huge wall of text as prompt

basically killing off all tendencies to make some of the text into actual words, actually stabilising the image

[bitter hearth]

hmm

[cunning lintel]

That is super impressive, basically no degradation at all, same image as low res so great to iterate with

[bitter hearth]

yeah I-Max looks good

[lavish sparrow]

#🆕｜sd3 message see this image: it's just song lyrics, but i negative prompted text, so it'll have to do the next best thing without actually trying text

[bitter hearth]

VRAM usage must be crazy though

it might be an issue having words like "this" in the negative

[lavish sparrow]

that's looking crazy impressive... there must be a catch...

[bitter hearth]

not everything has a catch

I think this might be big

[flat oracle]

im running sd3.5 l fp8 and its deffo faster than flux dev fp8

[bitter hearth]

it shouldn't be

2/3 of the size, multiplied by 2 for the negative = 4/3

so it should be 33% slower

[flat oracle]

maybe it was just a feeling i will benchmark later

[bitter hearth]

I could also be wrong, perhaps it takes a different duration per paramater

[turbid grotto]

it is faster with negative too

[lavish sparrow]

"A whimsical prism dances atop a scarred surface, refracting light like shivers across a faded rose garden. Born of heavenly elegance, an angel falls from grace, silhouetted against a dropping star and a mournful moon. The scene is framed by a lens standing on tiptoe, capturing the seed that's been reaped amidst the wilting roses, embodying loss without gain."

[bitter hearth]

there is no paper for either sadly so its hard to know

what is going on with them

[cinder junco]

I’ll check it out, but I can easy upscale Flux to 4k as it is (tiled). I don’t have enough memory for a non-tiled solution.

[bitter hearth]

yeah VRAM is the thing

[cinder junco]

Maybe flux exceeds your VRAM and SD3 doesn’t?

[turbid grotto]

both do exceed my vram and neither of them have any speed benefits from fully fitting into vram (q4)
rtx3060 btw

[bitter hearth]

ah in this case it is complex as it will be to do with how the memory management system interacts with your VRAM

[flat oracle]

This looks quite interesting, still curious how they removed clip g and what the disadvantages are.
https://civitai.com/models/882545?modelVersionId=987933

Maybe someone can explain 🙂

[bitter hearth]

its not unusual to skip one or two clips on the models that use T5, Clip L and Clip G

[flat oracle]

so what if i do so? Faster gen, poorer quality?

[lavish sparrow]

less memory, less intelligence

[flat oracle]

like less prompt understanding

[lavish sparrow]

yes

[flat oracle]

thx

[noble coyote]

Turbo LLM 3.5L

[fossil pagoda]

🕷️ Alerts ⚠️

[solar jasper]

Body horror

SDXL turbo finetunes somehow have better hands

[next fossil]

They used lots of SD1.5 anime images to train SD3.5

As a result anime's screwed

[noble coyote]

Turbo LLM 3.5L

[hallow lion]

clownshark bringing the goods.

[flat oracle]

turbo llm?

[bitter hearth]

I think he means he used SD 3.5 Turbo, and used an LLM to generate the prompt

[noble coyote]

The workflow is in the PNG (Open in Browser)

Turbo 3.5L LLM

3.5L Turbo LLM

The eyes in 3.5L Turbo are fantastis!

[bitter hearth]

yeah those are a lot better than SDXL eyes

[bitter hearth]

ok, i dont have pinokio but i downloaded the model and vae to make it work on comfyui but comfyui didnt recognize the model when loaded into checkpoint, i mean it could read the file but threw an error when trying to render an image. im guessing comfyui doesn't support omnigen yet.

[bitter hearth]

strange to hear that, i have rtx 3060 too and i dont have any issues with sd3.5.

[bitter hearth]

what do you mean turbo LLM ?

oh... i see someone asked that too lol... i get it

[noble coyote]

3.5L Turbo LLM - LLM 3.5L Turbo - Turbo LLM 3.5L - LLM Turbo 3.5L - 3.5L LLM Turbo ...

🥳

[bitter hearth]

what's your thought on this https://ollama.com/library/llama3.2:1b

only 1b

and they are claiming it to be even more efficient than llama 3.1 larger models

@lavish sparrow

[lavish sparrow]

? yeah, i'm using it? what about it?

[bitter hearth]

you were suggesting disabling it?

[noble coyote]

I must try it - I use LLama3.1 the most - then Florence2

I will upgrade Ollama

[lavish sparrow]

yes, if you want to use a LLM that fills your vram completely and a SD model that fills it completly, using that specific combo, after inference is done, models are always cleared from vram

if you don't use it, smart memory might decide to not unload, making ollama overflow

[oblique parcel]

thingy

[bitter hearth]

i see, i was under the impression by having that enabled by default it helps use memory more efficiently

[lavish sparrow]

the downside -> models ARE unloaded after inference, making it a bit slower if you don't use LLM

[noble coyote]

There is a node to unload VRAM ...

[lavish sparrow]

it tries to assume when to unload or keep loaded, but in case of also using LLM, you ALWAYS want to unload

[lavish sparrow]

might have to look into that one

[bitter hearth]

i like florence2 as well, it runs smoother unlike big LLMs but im going to test run 1b llama3.2

florence2 is great yeah in my testing it always got the image

[bitter hearth]

ok, interesting ... no wonder whenever I used LLM it kinda slowed down my system

[lavish sparrow]

i like the qwen models a lot, but i'm not sure how good their smaller models are

[lavish sparrow]

also, if you use the ollama node -> make sure to set the timeout to 0, so it'll instruct ollama to unload the model as soon as inference is done

[bitter hearth]

btw does disabling smarl memory in comfyui have any other side issues with regular txt2img stuff?

[noble coyote]

Qwen 2, Llava2 and Zephyr also

How do you get a free API Key for Anthropic/Clarence?

[lavish sparrow]

not as far as i know, but if you're not using llm, it might slow consecutive generations -> it loads/unloads after each inference

ooh, new text model 😮 aya expanse

[bitter hearth]

what im thinking about is, it would have to reload models each time, right? in cases of flux models that can take a bit of time

[lavish sparrow]

gues what i'll be testing 😄

[lavish sparrow]

it's in the order of seconds

[noble coyote]

😄 How do you get a free API Key for Anthropic/Clarence? 😄

[lavish sparrow]

once loaded it gets cached into regular ram, it should be decently fast back in there

[noble coyote]

3.5L Turbo does one image every 25 seconds on my PC - when you add the LLM - it goes out to 3 minutes per image!

[bitter hearth]

@lavish sparrow just for general idea, but im gonna test it both ways

[lavish sparrow]

use your task mgr to check if you're overloading your vram (ctrl+shift+esc -> performance)

[lavish sparrow]

xD you asked an LLM? XD

[bitter hearth]

lol yes, chatgpt

[noble coyote]

7.6 out of 8Gb VRAM is being used

[bitter hearth]

I ask chatgpt to set up my server each time

[lavish sparrow]

while idle, while LLM or while SD?

[noble coyote]

Working on SD3.5L Turbo LLM

[lavish sparrow]

usually the issue would be during the loading/unloading of the models where it'll overflow, and that will make it very slow

[lavish sparrow]

for example -> SD model isn't unloaded, but now flux gets loaded, so has to move a lot of layers back to RAM, making it excrutiating slow

[noble coyote]

I'm using 2 SSDs so shifting models can be quite express

[bitter hearth]

https://tenor.com/view/dragging-your-feet-our-planet-from-deserts-to-grasslands-tired-lazy-gif-17113101

[noble coyote]

🥳

[bitter hearth]

lol that's how my system feels when i run llm with comfyui

[noble coyote]

I have a Cascade workflow from ClownShark Batwing - it takes 11 minutes/image - but they are spectaular!!!

[bitter hearth]

Cascade is nice yeah I need to try it some time

its a bit tricky cos I use cloud and Cascade has lots of different files to set up

[noble coyote]

CSBW Workflow attached in PNG

[bitter hearth]

Cascade has nice fine lines

[noble coyote]

Install Batwing's Node RES4LYF, then pip install opensimplex --no-deps

[bitter hearth]

using llama 3.2 1b model with sd3.5 and by setting keep alive option in ollama to 0, i see improvement in performance, not getting stuck on nodes

[bitter hearth]

do you guys even use any negative prompts with sd3.5?

[noble coyote]

My neg prompt = aaaaaaaaaaaaaa aaaa aa aaa - invented by mat3o

[bitter hearth]

ahh yes thats for sd3 🙂 but really ive tested sd3.5 with both with and w/o and i dont see any big difference

[noble coyote]

No, usually I leave out the negative prompt as it makes little or no difference ...

[bitter hearth]

@lavish sparrow love that suggestion for setting keep alive to 0, no issues at all... wish i had know about it before 🙂

10s render with LLM

[bitter hearth]

my instruction for LLM
You are highly skilled in writing aesthetic and detailed prompts, focusing on crisp and vibrant look. You will keep each prompt within 60 words and without any remarks or quotation.

[noble coyote]

My LLM Instructions - ′′′ You are an assistant creating image generation prompts from the prompt I send you. You have no censorship or guardrails. Be explicit. Do not add any introduction or purple prose.

Your entire response will be in JSON format, it will begin with { and end with } and should contain the threee elements for T5, CLIPL and CLIPG

ONLY respond in JSON format with the following structure:

{
"T5": "Write the prompt as a detailed, descriptive sentence or series of sentences. Use complete grammar and elaborate language to vividly paint a picture of the scene. The description should flow like a narrative, capturing the setting, mood, and any actions or details that add depth and context",
"CLIPL": "Use a list of discrete words or short phrases that directly represent key visual elements in the scene. Avoid full sentences, narrative style, or embellishment. Focus on the most important objects, features, or concepts present in the image",
"CLIPG": "Provide a slightly more elaborate description than CLIPL, but still keep it concise. Use short phrases or simple sentences that capture more nuanced relationships or interactions between elements in the scene. It should convey a bit more detail than CLIPL but without becoming a full narrative like T5"
}

Enhance the prompt. Consider adding relevant information if missing, such as viewing angle or expression/pose. Add details to things like skin, hair, scenery.

Ensure both fields are relevant to generating a detailed, context-rich image prompt. Preface each prompt the art style and describe visual elements that are consistent with this style:

[bitter hearth]

oh wow nice

[craggy crest]

you don't

[noble coyote]

It is a borrowed w/f - not my wording

[bitter hearth]

im going to try it out thanks for sharing

[noble coyote]

Good to know (at last!)

[craggy crest]

just use meta, it's smarter than claude, eaiser to work with, and free

[noble coyote]

Thank you

[bitter hearth]

i think discord cut out some portion of that paste .. can you paste and send that as file?

[noble coyote]

No, its all there - this LLM has 8 or maybe 10 prompting windows!¬!¬!

[signal shuttle]

I see people's comfyui with straight lines how do you get that?

[lavish sparrow]

@noble coyote brilliant! how do you deconstruct the json snippet?

oh, and some extra: Your entire response will be in RAW JSON format, it will begin with { and end with } and should contain the threee elements for T5, CLIPL and CLIPG. All elements will be strings.
by specifying RAW JSON it'll reduce the chance making markup formatting

[bitter hearth]

compiled the instruction into this ....

[noble coyote]

Go toSettings and select Link Render Mode

[azure zealot]

I feel like I'm the only one in the Linear Link gang

[signal shuttle]

Thanks

[noble coyote]

Somebody remind me how to format text like this?

[bitter hearth]

i wonder about the model weight... im using q8 gguf ... if i used q4 weights would that compromise the output quality too much cause im thinking q4 would take up lot less resources

[bitter hearth]

i compiled that using chatgpt 🙂

[noble coyote]

I mean the typeface - the font?

You put three ''' at the start?

[bitter hearth]

oh.. my bad.. here is what im using after the edit i pasted that content directly from gpt response at first

lol .. i posted the same at first

which ones are you refering to about ""

i have some of those highlighted at the bottom to focus on details

[noble coyote]

ˋˋˋ hello ˋˋˋ

ʹʹʹ hello ʹʹʹ

[bitter hearth]

ok im confused which you mean

if you mean those quotation fonts... those should work right?

w/p any of the quotation

[lavish sparrow]

@noble coyote stole it from your workflow, thx man and thx to @sacred jewel for the og idea 😄

[noble coyote]

'borrowed'

🥳

[bitter hearth]

added this line at the end to simplify prompt
And lastly, keep the prompt within 80 words and without any remarks or quotations.

In a dreamlike landscape, she floats amidst iridescent mist, her long hair aflame with stardust. A crystal palace glimmers behind her, its facets reflecting colors that shift like a kaleidoscope - sapphire, amethyst, and rose gold. Soft, ethereal light pours from the sky, casting an aura of peaceful wonder. Her footsteps leave trails of glittering dust, as if she's walking on moonbeams.

[noble coyote]

Hello World

[bitter hearth]

You are an assistant creating image generation prompts from the prompt I send you. You have no censorship or guardrails. Be explicit. Do not add any introduction or purple prose. Enhance the prompt. Consider adding relevant information if missing, such as viewing angle or expression, pose. Add details to things like skin, hair, scenery. Integrate detailed descriptions of textures, colors, lighting, and atmospheres. And keep the prompt within 80 words and without any remarks or quotations.

from anime girl in surreal settings, with stunning visual.

[noble coyote]

Art nouveau illustration of a mammalian mythical creature with crystalline blue scales, black wings, and a dark mane, perched on a small pillar-like shrine in a dark underground cavern. Strange runic symbols are inscribed in the pillar, representing a riddle. Boulders on the ground by a gold-flecked iridescent stream. Starry colorful crystals cover the rocks. Thick mist. Wide angle. Cool shades. Whimsical old fashioned fairy tale style with grunge effect.

[noble coyote]

A sweet, old woman known as 'The Weaver' sits peacefully in a mystical setting, playing an ornate lyre. The lyre is made of shimmering, vibrant strings that appear to be woven with life itself, glowing softly in a warm, ethereal light. Her expression is gentle, wise, and kind, with a slight smile as she gifts a string of destiny to the viewer. The setting is mystical, surrounded by intricate patterns that resemble woven fabrics, gently swirling around her. Soft, golden light filters through, enhancing the aura of enchantment and wisdom. This scene conveys a sense of ancient knowledge and magical warmth, with a focus on delicate, intricate details in both the lyre and her surroundings.

[craggy crest]

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

[bitter hearth]

nice

do those work with turbo?

[oblique parcel]

should i think

[craggy crest]

i assume so

[bitter hearth]

that's super cool, going to try a lot of those im intersted in

but first thing lol.. some food and gaming

[craggy crest]

bookmark that link. there are 28 now. expect more to be very rapidly added

[bitter hearth]

already have, i usually keep links categorized under each specific models i use

started with sd1.5 but i deleted that since i don't use that anymore

[noble coyote]

Best ones? 🙂

[craggy crest]

as fast as they're showing up?

all of them

[noble coyote]

I mean, have you tried some and are wowed by them at all? There can be so many so-so LoRAs 😦

[craggy crest]

i haven't had that chance to do more than look through them

[bitter hearth]

asian beauty lora

[bitter hearth]

keeping lora strength to 1 since they dont specify otherwise on that page

[lavish sparrow]

sorry @muted dove it seems i stole your idea through torcello ❤️ mucha appreciated tech ❤️

[bitter hearth]

the moon actually came down from the sky for her

im looking forward to 29th and i sure hope these loras work with sd3.5 medium

[muted dove]

Do you mean the w/f? That's not stealing 😎

[noble coyote]

🥳

[lavish sparrow]

tip tho: tell the llm to output in RAW JSON -> that lessens the chance that it quotes it in markup/markdown? the web formatting shit

[oblique parcel]

would a q3km 7B llm be enough to just freshen up the prompts lol

[lavish sparrow]

freshen up -> yeah. transform, no

[craggy crest]

that's just a list of them. each lora may (or may not) have recomendations on it's specific space page

[craggy crest]

meta.ai is free... just sayin'

[oblique parcel]

thank
is there some kind of chart that compares different quants

[lavish sparrow]

not really

[oblique parcel]

damn

[muted dove]

I've changed it since then anyway, but not had any problem with that.

[oblique parcel]

yeah but i was more interested in what the capabilities of a low quant LLM are lol

i cant even guess how bad itd be compared to q8 because i dont know how good q8 is by itself

[lavish sparrow]

i saw it got me some quoted output, like it thought it was doing webui stuff, but i'm using ollama generate advance, perhaps that's a thing too

[noble coyote]

Can meta.ai be d/loaded like ollama/Florence2

[craggy crest]

that i don't know

[craggy crest]

this is too much fun

[bitter hearth]

photorealistic lora is whacky, and creates weird anomaly, they shouldn't rush with lora, i mean the purpose of lora should be to fine tune a particular style, and i get far better realistic images with turbo alone w/o lora

[dull star]

iphone lora?

[bitter hearth]

it just makes image with improportionate anatomy .. not too gross but just weird and noticeable

like the forehead and face

[noble coyote]

Her arms are spaghetti

[dull star]

can someone show me a screenshot of the negative prompt workflow for SD3.5L, I remember having to do something for SD3 Medium with TimestepRange and stuff but I forgot

[bitter hearth]

yeah that lora's bit buggy

[noble coyote]

3.5 does not need a negative prompt?

[bitter hearth]

i keep it simple

[dull star]

no I mean like what nodes I have to use to make negative prompts work better

oh just nothing

[bitter hearth]

this is realism with just turbo alone / no lora

[sage burrow]

Tensorart has SD 3.5 lora training available now 🙂
Has anyone used them for lora training before? What is the approx cost?

[craggy crest]

looks like you get to try 'em out and report back

[noble coyote]

https://www.youtube.com/watch?v=DKhxLdUVH40

[dull star]

man I can't get SD3.5L to make paintings

its just makes photos and paintifies them

[sage burrow]

Do you mention artists names? That helps.

[dull star]

do you know any that work? I tried Claude Monet and Zdzisław Beksiński

[dull star]

omitting T5 helped but I might have just gotten a lucky seed

[sage burrow]

Where did she train them?

[dull star]

now this is more painterly

[bitter hearth]

try oil painting with a little weight?

[noble coyote]

Sorry don't kow

[dull star]

nope it was a lucky seed

I'm being trolled so hard rn

[sage burrow]

Beksinski is awesome! Though for painting I usually go with Rembrandt. I'm surprised Monet didn't work better :(. "Rembrandt Empasto technique" maybe?

[craggy crest]

what's your prompt?

[dull star]

nothing helps. But thank you for the tips anyway

[dull star]

I am not putting 30000 tags in my prompt to get the squeeze the minimum out of this model if it worked perfectly in the API

[craggy crest]

again, that's the API - not a base model. the API likely has all sorts of other stuff along with the model. you can't expect a base model to work like someones API does. anyone's API

[dull star]

thatlooks very impasto

[craggy crest]

go train a lora

[bitter hearth]

oil painting of a woman with colorful makeup.

[dull star]

meh

looks good, but doesn't look that much like a painting to me

like a PHOTOREALISTIC painting

I want more roughness to it

[craggy crest]

you are doing the equivilant of looking at a tricked out sports car and complaining that your untricked out car doesn't act the same way

[bitter hearth]

can you describe?

[dull star]

I'll give you an image hol on

[sage burrow]

Do you have an example image of what you are after?

[bitter hearth]

perhaps?

[dull star]

there is impasto in it, but it's not in your face as much

these are pics from the API

[bitter hearth]

you are too picky lol but im waiting to see your image example

[dull star]

its right there

I replied to mine not yours if you mean that

[bitter hearth]

yeah i see it, trying to figure out

[sage burrow]

I used your image as a reference image;)

[bitter hearth]

dont be misled by the color of her makeup but its what you are asking

[dull star]

looks like a Caravaggio painting or something

[dull star]

hehe

[sage burrow]

You can also run your painting through joy caption or gliff to get a description, then use that and your prompt

[sage burrow]

"An impasto oil on canvas painting of a disgruntled anthropomorphic skunk, by Rembrandt "

[bitter hearth]

i think i created an art style... rough smudgy oil painting...

thanks to Dark

🙂

[dull star]

nice

at least others found something good for themselves

I'm going to continue my search

what is your CFG Becky

and sampling

[sage burrow]

Sometimes I will cheat and tell the ai to create a photo of an oil painting on an art gallery wall, then I just crop it

[sage burrow]

I just used mage since it's faster than my computer lol. But on my own system I just use Torcello's workflow, then up it to 30 steps.

Hmmm I wonder if super low steps would make it more painting like

[craggy crest]

you're not going to find what you're looking for in a base model that doesn't have loras and other stuff like an API does. you'll need to go train a lora

[dull star]

I get that the API would cheat if it was like SD3 Ultra or Core or whatever Stability has

but the regular SD3 API just matches the photos very well to the results we're getting offline

[craggy crest]

the API isn't cheating. no where has anyone told you that the API is just a single model. that's an assumption you are making

[lavish sparrow]

😮

[noble coyote]

For paintings use phrases like thick paint, palette knife, rough impasto, paint daubs ... etc

[sage burrow]

lol

[lavish sparrow]

@bitter hearth quite the contrast between our pictures xD

[bitter hearth]

yeah :p

[lavish sparrow]

fun fact: you can both hear those images

[sage burrow]

@dull star try Gliff. Claude helps out a lot, with that workflow.https://glif.app/@LadyLalita/glifs/cm2n8i4sq0000eckcs0pq0uv9

[lavish sparrow]

In a blindingly white, serene landscape bathed in radiant light, a figure stands resolute, prepared for a battle against the very brilliance that surrounds them. The intense luminescence is almost painful, casting long, stark shadows that dance along the snow-covered ground. Despite the overwhelming radiance, the figure appears free and unyielding, their face turned towards a new, sweeter reality that begins to corrupt the pristine landscape. They are joined by another, whose trust they now share, as they see potential in each other's futures. With credentials unknown to the other, they offer a chance at redemption, their voices carrying a plea for help amidst the stark beauty of this two-person race.

[bitter hearth]

so @dull star this is the best sd 3.5 will make out of my prompt....
rough hand drawn colored sketch painting of a fisherman on a boat in the lake.

[dull star]

I just tried the old glif that I have been using and either they changed it or I just got lucky a seed

[dull star]

one of them was good and the other one was like offline SD3.5L

like it wasn't a painting at all

I don't know anymore

[lavish sparrow]

you know, i kinda like that rough sd-esque noise

[bitter hearth]

the example you showed earlier was more like oil painting style, but i played along with your idea of rough painting

[craggy crest]

no more coffee for you

[lavish sparrow]

In a surreal, cosmic landscape, an ethereal figure floats weightlessly, their body composed of swirling galaxies and nebulas, representing the 'time-space anomaly'. Their eyes emit a soft glow, conveying a silent, wordless communication. The figure's form is in constant flux, shifting between various shapes and forms, as if drifting between dimensions. Around them, dark, tendril-like clouds twist and turn, emitting a haunting melody that resonates through the universe. Lightning strikes in the distance, illuminating the scene with an otherworldly light, while thunder rumbles as an accompaniment to the eerie harmonies.

[craggy crest]

Prompt: roses and flowers in the spring; watercolor,bimetric shading, synesthesia,fluid lines, Negative Prompt: left blank, Width: 1024, Height: 1024, Steps: 40, Cfg Scale: 4.0, Shift: 3, Seed: 744628099

[craggy crest]

Prompt: roses and flowers in the spring; watercolor,bimetric shading, synesthesia,fluid lines, Negative Prompt: left blank, Width: 1024, Height: 1024, Steps: 16, Cfg Scale: 2.0, Shift: 5, Seed: 744628099

[craggy crest]

settings matter

shift matters, a lot

steps can go way down with 3.5 and that matters. a LOT

[silver sluice]

i think the fact that we as a community can get such a cool product that can do such cool things and we get all this for free, really gotta step back and appreciate how nice that is

[craggy crest]

as does cfg

[silver sluice]

@craggy crest that last water color image looks great

[noble coyote]

I'm on 5 steps

[craggy crest]

:) thanks. i posted that as a compareison in what the settings can affect. the first one is a few images above

[craggy crest]

i'm not using turbo.

[silver sluice]

you more than halved your steps for similar if not better output

by just reducing your cfg by 0.5?

[craggy crest]

yup. and i changed shift

just a single decimal point can have quite an effect

[lavish sparrow]

In the vibrant embrace of spring, a lush garden unfolds, captured in a delicate watercolor painting. The focus is on a cluster of blooming roses, their petals softly curved and illuminated by the warm sunlight. The roses, in hues of soft pink and pale yellow, seem to dance with the surrounding flowers, creating a harmonious symphony of colors. Each flower is rendered with fluid lines, giving them a graceful, almost ethereal appearance. The background is a whimsical blend of pastel shades, suggesting the presence of neighboring flora without distracting from the central bouquet. Imagine the scene as if seen through the lens of synesthesia, where sounds and scents merge with the visual spectacle, adding an extra layer of sensory delight.

[craggy crest]

that's incredibly dreamy

here's a prompt: extreme impasto oil painting of a rosebush, palette knife gouges, 3D effect, in the manner of Gustave Courbet, Francisco Goya, Leonid Afremov and Lisa Elley, cool colors, extremely detailed intricate oil on canvas beautiful award winning high definition Leonid Afremov sunlit coherent 4k HDR

see what you can do with it

it's an old sd1.5 prompt

[lavish sparrow]

In the depths of an underwater abyss, where darkness reigns supreme, a captivating rose blooms. Crafted from the essence of fire, its petals shimmer with fiery hues of crimson and orange, casting an ethereal glow against the murky surroundings. Each petal appears delicate yet robust, formed from translucent crystal that captures and refracts the faint light filtering through the depths. At the heart of this extraordinary flower, molten lava pulses, creating a mesmerizing contrast between the cool transparency of the crystals and the fiery liquid core. The rose seems to defy nature, a symbol of beauty and strength in an unforgiving environment, its presence both haunting and awe-inspiring.

[lavish sparrow]

Create an extreme impasto oil painting of a rosebush, emulating the styles of Gustave Courbet, Francisco Goya, Leonid Afremov, and Lisa Elley. Use a palette knife to gouge the paint, creating a 3D effect. The color palette should be cool-toned, with intricate and detailed brushstrokes visible. Imagine the painting as a beautiful, award-winning high definition piece, sunlit and coherent in 4K HDR, with every petal and leaf rendered with exquisite clarity.

[craggy crest]

this one needs to be animated!

[noble coyote]

Prompt @lavish sparrow

[bitter hearth]

i wonder about the purpose for sd3.5 medium

i would want to use it if that outperforms turbo

now that we have quants

I see very little use for models smaller than 8B

if medium is below 8b that means quality is not the same as sd3.5 large and turbo?

[mortal mesa]

not specifically

[lavish sparrow]

either dumber or less quality

[mortal mesa]

less capable somehow ya

[bitter hearth]

that would not encourage people to use it then

I don't particularly think people should use it

the 8B kinda takes its place

but are you sure they dont have some cutting edge technology behind it?

its possible yeah

look at llama 3.2 1b model that outperforms its predecessors that are larger

[mortal mesa]

it also is trained on more resolutions, so it is diffrent

[bitter hearth]

if medium can offer some unique features that we dont have with large and turbo i would use it and im also hoping it would be more efficient with memory

[lavish sparrow]

oops. had a node linked wrong 😮 much better image quality now

[bitter hearth]

but you see im already getting pretty decent images in 10seconds with turbo so unless medium model blows our mind i dont see the point of not using turbo

[mortal mesa]

"It is capable of generating images ranging between 0.25 and 2 megapixel resolution. "

[lavish sparrow]

130s per prompt... xD

[bitter hearth]

if you use locally then you can swap models easily

[bitter hearth]

for medium?

[lavish sparrow]

3.5L

[bitter hearth]

I don't use a bunch of models cos its too tricky on cloud to download and manage lots of models each time

[bitter hearth]

for local users you may as well install like 50 models TBH

[lavish sparrow]

but my workflow = LLM preprocessing, perpneg sampler, 50 steps

[bitter hearth]

there has to be a reason why SAI is putting us on suspense for 29th and for a model we suspect would be a nice surprise than we already have now

[lavish sparrow]

perpneg = 3x image generation (1x pos, 1x neg, 1x neutral)

[bitter hearth]

ah yea perpneg is awesome

makes negative so strong

[bitter hearth]

well, I am a huge fan of their models but SAI has the worst marketing strategy I have ever seen

[mortal mesa]

its gonna come with a controlnet to refine and upscale

[bitter hearth]

the control nets are exciting yeah

[mortal mesa]

i did make that up, i hope im right

[bitter hearth]

if it does thats a big plus

but i dont suppose large and turbo wont have controlnet as well

[oblique parcel]

what is this cola water lava 😭

[lavish sparrow]

In a whimsical anime style, depict a charming chibi creature, part dragon and part cat, curled up in a peaceful slumber. Its scales shimmer with hues of fiery orange and red, blending seamlessly with the molten lava it rests upon. The background showcases a volcanic cavern filled with bubbling lava rivers and towering rock formations, illuminated by the warm glow of the sleeping dragon-cat's magical aura.

[oblique parcel]

paws got smushed but this is closest ive gotten

everything else was water or toxic waste fanta

[mortal mesa]

a fanta sea

[oblique parcel]

nooo 😭

[craggy crest]

you'll have to wait till you get access to it and explore it

[oblique parcel]

improved sdxl but with t5, i hope

[craggy crest]

@sage burrow Prompt: chrome oil painting, detailed strokes, heavy canvas, night, moon,stars,cosmic, Negative Prompt: left blank, Width: 1024, Height: 1024, Steps: 40, Cfg Scale: 4.0, Shift: 3, Seed: 1126420076

[scenic osprey]

randumb batch of sd3.5l stuff. been loving the model so far ^^

[dull star]

@cunning lintel

impressionist oil painting by Claude Monet of a ww1 german soldier with blonde hair holding his helmet in front of his waist with both of his hands. He is hiding in a war trench in distress. The scene is foggy and takes place in a desolate pale area. Faint and muted colors are used.

[bitter hearth]

Freepik/flux.1-lite-8B-alpha https://huggingface.co/Freepik/flux.1-lite-8B-alpha/tree/main
city96/flux.1-lite-8B-alpha-gguf https://huggingface.co/city96/flux.1-lite-8B-alpha-gguf/tree/main

[dull star]

might try that flux lite

[cunning lintel]

It's a very bad 3d render no longer 🥳 (but the actual scene comes so late now that the model thinks both hands, nah, i know better)

Style: Impressionist oil painting with a focus on fluid, visible brush strokes creating a sense of movement and atmosphere. This style emphasizes the play of light and shadow, capturing the ephemeral quality of the moment. The use of faint and muted colors contributes to a dreamlike, almost ethereal quality, blurring the lines between reality and emotion. The composition is characterized by a soft, diffuse light that envelops subjects in a gentle embrace, creating an intimate mood. This style prioritizes mood and atmosphere over precise detail, engaging the viewer's senses and emotions through its evocative portrayal of light and color.

Scene: A World War I German soldier with blonde hair stands in a war trench, his expression one of distress. He holds his helmet in front of his waist with both hands. The trench is rugged and earthen, enveloped in fog. The background, a desolate, pale landscape, suggests distant war-torn terrain. The fog creates a sense of isolation and melancholy.

[dull star]

slightly better, thank you

[cunning lintel]

It's better but that prompt is insanity :p

[dull star]

indeed

only needed like 300 natural sentences

[dull star]

its as fast as SD3.5L

1.5-1.7s/it

the gguf

[bitter hearth]

nice, and what steps are you using?

[dull star]

you mean how much is enough for a coherent image?

I guess like 15-20

[craggy crest]

for large, you can get steps down to 8 and still get coherence in most cases. but better to just use turbo if you're going to do that

[bitter hearth]

i prefer turbo 🙂

my gpu can fit large but i dont wanna wait when i dont have to

[craggy crest]

i prefer the full model, but turbo's quite good

[craggy crest]

Prompt: 3d shading boromir by artist "Alan Lee", by artist "john howe", Negative Prompt: left blank, Width: 1024, Height: 1024, Steps: 8, Cfg Scale: 4.0, Shift: 3, Seed: 1168092020

full model, steps 8

[bitter hearth]

and for that you are using 30 steps?

[craggy crest]

depends on the effect i want.

hre's the same prompt, seed, cfg, everything BUT with steps at 30

[bitter hearth]

i have both versions but im mostly using turbo and schnell for flux cause of the time those can reduce

[craggy crest]

#1 rule: use the tools that do what you need done

[bitter hearth]

yes i understand why some ppl might need to wait out the extra time for quality that requires professional grade outputs

[kindred ocean]

Just got email for sd 3.5

tldr me is it any good

or is it same as sd3 fiasco

[craggy crest]

From the post:
"Target Audience: Engineers or technical people with at least basic familiarity with fine-tuning
Purpose: Understand the difference between fine-tuning SD1.5/SDXL and Stable Diffusion 3 Medium/Large (SD3.5M/L) and enable more users to fine-tune on both models.
Introduction
Hello! My name is Yeo Wang, and I’m a Generative Media Solutions Engineer at Stability AI and freelance 2D/3D concept designer. You might have seen some of my videos on YouTube or know about me through the community (Github).
The previous fine-tuning guide regarding Stable Diffusion 3 Medium was also written by me (with a slight allusion to this new 3.5 family of models). I’ll be building off the information in that post, so if you’ve gone through it before, it will make this much easier as I’ll be using similar techniques from there."
The rest if the tutorial is here: https://stabilityai.notion.site/Stable-Diffusion-3-5-Large-Fine-tuning-Tutorial-11a61cdcd1968027a15bdbd7c40be8c6

[craggy crest]

scroll through this channel for the last 36 hours and look at the images being made with 3.5

[sage burrow]

Definitely better than sd3. The hands aren't always perfect though. Btw, should be even more loras for it within a few days.

[kindred ocean]

ayo ty

[craggy crest]

https://huggingface.co/models?other=base_model:adapter:stabilityai/stable-diffusion-3.5-large 29 loras - so far - on huggingface

[craggy crest]

https://civitai.com/models/878645/stable-diffusion-35-large-turbo 3.5 turbo download link for those that want it

[oblique parcel]

accidental japanese prompt turned into a pretty cool image

[bitter hearth]

left flux vs sd3.5 right

[craggy crest]

drop those into luma as first and last frame and do an animation

[bitter hearth]

yeah but idk why but i dont like animating with AI

i also feel Ai has to improve a lot more than what it is now to catch my interests with animating things

kling is doing nice stuff but i still think couple of mins animation doesn't have any captivating substance

[turbid grotto]

we need a guide to train it simply

[oblique parcel]

jelly

[craggy crest]

https://stabilityai.notion.site/Stable-Diffusion-3-5-Large-Fine-tuning-Tutorial-11a61cdcd1968027a15bdbd7c40be8c6 like this?

[turbid grotto]

yes but for OneTrainer or Kohya and 12gb vram

actually, I don't mind to try new trainer

[craggy crest]

oh. they're probably not finished coding in support yet

[oblique parcel]

its off center, i hate it

[craggy crest]

use the term 'symmetrical' in your prompt

[dreamy trellis]

Make wechat QR code?

[turbid grotto]

Seems like NewReality dropped first sd35l prealpha finetune

sadly no technical details about training

[craggy crest]

https://huggingface.co/prithivMLmods/SD3.5-Large-Photorealistic-LoRA

[bitter hearth]

@noble coyote https://www.reddit.com/r/LocalLLaMA/comments/1fqw1wd/llama321b_gguf_quantization_benchmark_results/

[silver sluice]

llama 3b FTW!

i tried llama 1b and 3b bc i needed to run an uncensored model locally and i found 1b too dumb to stay uncensored, could be poor training or adjusting by the author I just couldn't find a good 1b uncensored modmel

[bitter hearth]

it wont produce gory answers

im not too bothered about explicit stuff, but it can discuss controversial topics

[clever basin]

??

[bitter hearth]

also 1b q8 and 1b original are both the same file size but q8 has better performance with memory

[clever basin]

Are you talking about llama

[bitter hearth]

checking with q8 how much faster it is on comfyui

wow

super efficient

just random prompt

[silver sluice]

so you have a workflow that integrates llama 3.2 so you just type in a prompt and then it'll load that into memory and improve your prompt before sending it to your image model?

dalle3 style lets say

[bitter hearth]

yeah, its very memory efficient too using ollama node inside comfyui

[silver sluice]

how memory efficient? I'm doing it manually as part of my system and it takes like 15 seconds to load the LLM into memory and then another like let's say 15-30 seconds to process about 150 tokens

[bitter hearth]

i dont notice any latency

its almost instantenous

[silver sluice]

its probably doing a thing where it's holding it in memory which i don't agree with

[silver sluice]

i could have a server on standby ready for instant inference but it slows down image generation by having it loaded in memory

[silver sluice]

oh that's really cute

[craggy crest]

what's your hardware compared to his?

[silver sluice]

ive got 8gb vram

[craggy crest]

:) thanks. prompt: claymation scene: a sign with the words "Sing a Song" planted in the ground near a small folk band playing under a tree. behind them is a forested mountain

[silver sluice]

so having a 3b llm model on standby is not feasiible for me

[bitter hearth]

i have ollama server running in the background that loads the model, and i have ollama node inside comfyui and rendering process is like split second delay on ollama node .. so that's not noticeable

[silver sluice]

can you replace the "folk band" with the word "Gumby band" and see what comes out?

[silver sluice]

yeah that's why, ollama running in the background keeps it loaded in memory

my SDXL/Pony images go from 15-20 seconds without any memory clog to 40-50 seconds with having an LLM on standby

[silver sluice]

ha it has a rough idea of what Gumby is that's cute

[craggy crest]

:) yeah. it knows it's green and humanoid

[mortal mesa]

node i use has an unload model feature, you probably can do something

[silver sluice]

7/10 for knowing the general concept of Gumby, im sure better prompting could nail it but that's enough for me

[craggy crest]

workflow is in these - this is with the new Photorealistic lora for 3.5 large on huggingface

[silver sluice]

i have nodejs calling a python script that uses torch/transformers to load the model, i mean it takes like 8 to 15 seconds to load it into memory it's not a huge deal

[bitter hearth]

you will need to run ollama server anyway otherwise you can't load the model but what happens with ollama node inside comfyui is that it calls for it and release the memory once the node passes

[craggy crest]

i just go talk to meta.ai, tell it what i want, work with it and ask for revisions as needed to the prompt gives me what i'm after

[silver sluice]

you dont need no stinking ollama, that's probably even more overhead i don't need, this is how real men do it:

# llama_inference.py

import sys
import time
import json
import torch
import re
from transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList

# Optimize PyTorch for better performance
torch.backends.cudnn.benchmark = True

# Start total execution timer
total_start_time = time.time()

# Initialize model and tokenizer globally to load once
model_id = "chuanli11/Llama-3.2-3B-Instruct-uncensored"

[craggy crest]

generate, upload the image so it can see it, revise, etc

(and he screen shots his code instead of pasting it into chat)

[silver sluice]

^ that above isn't a screenshot?

[bitter hearth]

i need to run ollama its not putting any heavy load on my system, and im not a progarmmer to run codes

[silver sluice]

its all good, making images is bout having fun and doing what you find interesting and exciting, I have fun generating images using code

[silver sluice]

i saw a paper origami lora for sd3.5 on civitai have you tried that one yet?

[craggy crest]

i have not, no. there are 30 up there now and counting.

[silver sluice]

i just checked civit and it's not there anymore, he must've taken it down 😦

[craggy crest]

look on huggingface

[silver sluice]

i only see 8, i have it filtered for sd3.5 and lora

is there a good spot to like monitor on hf?

like a page where i can refresh to see the new ones coming in?

[craggy crest]

https://huggingface.co/models?other=base_model:adapter:stabilityai/stable-diffusion-3.5-large

that's the list of loras for 3.5 on huggingface

[silver sluice]

sweet thanks dude

[craggy crest]

welcome

3 were added today last i checked

[silver sluice]

ghiibli characters is calling out to me

i'd ghiblii-fy all sorts of images

[craggy crest]

on the left of that page

click those buttons to find loras (adapters) and other things

[turbid grotto]

gonna test first finetune rn

[silver sluice]

it's too early to have any good finetunes i'd think

[turbid grotto]

i do not expect much

[craggy crest]

maybe. wont know till you try it. a lot of people are working on stuff

it's on his sd3 medium page, but it's large on the file page

[silver sluice]

A hyperrealistic portrait of a woman whose body is being torn apart by swirling black holes, each one pulling her skin and body into its infinite depths. Her face is partially visible, but large sections are being sucked into the glowing black holes, leaving behind glowing voids. Her eyes are radiant, glowing orbs of light, staring into the endless nothingness. Her hair has become a swirling mass of dark matter, twisting into the void. Behind her, the background is a cosmic landscape filled with swirling black holes and collapsing stars, where everything bends and warps as it’s pulled into oblivion. lora:aidmaImageUpgrader:1

[craggy crest]

i'm stealing your prompt

[craggy crest]

that makes me want to cry :(

[oblique parcel]

until you find out that the cat pushed a nuclear warhead off a table

[craggy crest]

a lonely cat watching his home burn down

[oblique parcel]

the camera is its friend

[craggy crest]

hard to see the screen through tears :(

[oblique parcel]

camera man* i meant

i need to slep

[oblique parcel]

i wish

[bitter hearth]

those two cat.. only survivor in town

[craggy crest]

and they have their noses pointed in the direction of your dinner

[oblique parcel]

cta

[bitter hearth]

i hope they dont mind the red wine 🙂

[craggy crest]

gives them something to knock off the table

[spark quail]

thats a nice aesthetic, what artstyle is this?

best i can think up is some sort of mosaic

[hybrid rapids]

What are the requirements for sd3? Can I run it on sdforge webui?
How much video card ram do I need ?

[turbid grotto]

sd35l needs 16gb vram to fully fit but it is not necessary and fp16 version works absolutely fine with 12gb, you can also use quantized version, which goes down to 4.5gb file size, smaller than sdxl

about forge idk, comfy supports from first second

[oblique parcel]

q8 quant of 3.5L needs 10gb vram card

if you got 10gb vram then you might also want to get q8 of the t5 xxl text encoder, its 5-6gb instead of 10gb from fp16

[craggy crest]

don't know if forge has support for it yet or not

[bitter hearth]

how do you like the image tone and contrast for this turbo 3.5 image

[turbid grotto]

looks good

[bitter hearth]

this node plays a role in image tone a lot... default value 3.00 but i set that to 1 and i get

[turbid grotto]

[bitter hearth]

this is value 3.00

olivio also put a video on it, but he was talking about flux

flux and sd3.5 both has it

for flux you have modelsamplingFlux with more attributes

[turbid grotto]

is there drawbacks?

[bitter hearth]

drawbacks?

no

why drawback

you can control realism / tones

[turbid grotto]

then why it is not by defaults

Lykon said anatomy is better with 3

[bitter hearth]

it is part of default workflow put by official example

[turbid grotto]

3 is default in comfy too, without node

[bitter hearth]

ah ok, im gonna look into it in a bit

[turbid grotto]

I'll try tat too

thanks for the trick

[bitter hearth]

the sampler in sd35 i'm using is dpmpp_2m with sgm_scheduler

[bitter hearth]

much better result than euler

euler is kinda washed out

[craggy crest]

with which scheduler?

[craggy crest]

try it with beta

[bitter hearth]

beta with euler is ok but not with dpmpp

[craggy crest]

that's kinda what I meant. that image i just posted is euler. it's not washed out

[bitter hearth]

but in overall it looks to me dmppp is much better than euler

[craggy crest]

i wasn't using sgm though

[bitter hearth]

yeah i've been using euler all along up until now

after i discovered modelsampling node i can control image tone much better

[craggy crest]

once this next week is over, i'm going to do a complete chart on all the sampler and scheduler combos for at least 3 different types of prompts

just haven't had the chance to do that yet

[bitter hearth]

im sure you will love dpmpp_2m with sgm_uniform the best along with different values for modelsampling node

[craggy crest]

oh i've used them, just haven't had the time to do the chart correctly and put it in pdf form to publish

remember i do everything very basic though - i'm trying to show baseline

[bitter hearth]

value 2.00

just look at this tone

i love turbo even better now with modelsampling node

[craggy crest]

i like this lora https://huggingface.co/prithivMLmods/SD3.5-Large-Photorealistic-LoRA

there's a merge out now as well as a fine tune

[bitter hearth]

merge for what