🆕｜sd3

[muted dove]

With a flex in their profile of "Crypto holder". 🤣

https://tenor.com/view/laughter-emoji-gif-26681851

[turbid grotto]

did sd35m out??

https://huggingface.co/stabilityai/stable-diffusion-3.5-medium

[bitter hearth]

could someone put on civit pls my shell scripts are written for that

[signal shuttle]

SD3.5M is OUT https://huggingface.co/stabilityai/stable-diffusion-3.5-medium

[finite osprey]

waiting for the first woman lying on the floor

[signal shuttle]

Made with SD3.5M

[bitter hearth]

detail levels look good

[signal shuttle]

Text ain't half bad

[turbid grotto]

yea that is definitely improvement

[cobalt moon]

This image shows a cozy, sunlit room with a relaxed, homey atmosphere, featuring Furina from Genshin Impact. She stands barefoot near a window, dressed casually in a tank top and shorts, holding a plate of food and a glass of water, looking slightly surprised or concerned. The warm sunlight streaming through the curtains softens the scene, creating a peaceful, domestic vibe. On the left side, a pair of playful, water-like creatures—one wearing a whimsical top hat—are causing mischief in the kitchen, splashing water all over the sink and bubbles floating in the air. The room is filled with small, charming details, like shelves full of books, potted plants, and scattered objects such as cans and a tipped-over cup. The balance between Furina’s calm stance and the chaos in the background gives the scene a playful, slice-of-life feel, capturing a moment of quiet absurdity in a seemingly ordinary day.

not a great prompt but hey.

[turbid grotto]

textures are really good

[noble coyote]

Flux/Florence2 + LoRAs img2img

[patent acorn]

smh base model doesnt recognize furina

[cobalt moon]

doesn't recognize Loona from Helluva Boss either

I mean the basic feature is there

[patent acorn]

cuz its not trained on dat

wait for finetunes

[turbid grotto]

am I crazy or it is not worse than large?

[patent acorn]

"sd3 medium is all you need"

[turbid grotto]

no, 8b is better but that one good too

[dull star]

waiting for gguf

[cobalt moon]

it's here already
https://huggingface.co/ND911/stable-diffusion-3.5-medium-GGUF/tree/main

[dull star]

oh okay

[turbid grotto]

why, it is small

[dull star]

I was looking at city's huggingface

oh wait yeah

I just got used to gguf lmao

[turbid grotto]

bro was camping

[dull star]

didn't even remember WHY I was using gguf

[cobalt moon]

someone like me with 2GB VRAM

[dull star]

yeah its good for that

[patent acorn]

wtf

[dull star]

thank god for these GGUFs

they have so much better quality than fp8

[turbid grotto]

oh curios how q4 gonna run

[cobalt moon]

will try that tomorrow

like with my 2GB VRAM setup

I did have my 6GB VRAM laptop tho

[dull star]

q4 is good on like 7B and up models (talking about LLMs)

but this may apply to DiTs since these are transformer based

but the point is, a large quantized model might outperform a small unquantized model

but it will be way slower for sure

hmm waiting for all the reddit posts that trash on SD3.5 Medium

its to be expected

[turbid grotto]

I hope it will be adopted, it looks so good

look at this aesthetics

[cobalt moon]

The comments saying "why use this when you have Flux" already appeared

That guy probably have dual 4090 setup

[dull star]

yeah he can fill flux and T5 at FP32 on his GPUs

[turbid grotto]

imagine opposite happens

[patent acorn]

LOL i saw that

[turbid grotto]

q4 is smaller than 1.5 but speed drops from 1.16s/it down to 1.41s/it

quality does not drops heavily

q4 vs fp16

[noble coyote]

"My Triple 5090 512Gb RAM, 3 x 10Tb SSD is gonna ... "

... need a pocket nuclear-reactor to power

[patent acorn]

wanted to try style blending between photo and anime, unfortunately it didnt do well in medium maybe i suck at prompting 😂
"a photo of a cafe at night. there is an anime girl sitting on a chair"

[turbid grotto]

i couldn't too but I don't know correct prompting for that

[patent acorn]

only seen lykon trying 3 subjects in separated style

[noble coyote]

raw photo of a cafe at night. an anime girl is sitting at a table

[patent acorn]

ok lemme try

nope

[rapid salmon]

Super nice. Curious though, how was a K quant made? When I look into the options for quantizing models (e.g. leejet's stable-diffusion-cpp), they only have non-K-quant methods.

(also I'm dying to see Forge get K-quant support, from my understanding it's still not in yet, though I could be relying on outdated info)

[lavish osprey]

this is kind of impressive, ngl

[rapid salmon]

It seems like ComfyUI is the only source (that I know of) that supports K quants but...I really prefer other tools. :x

[turbid grotto]

yea that is very good at various styles

[remote holly]

is realy good

[lavish osprey]

about to upload my upscaling workflow

[bitter hearth]

ok nice, tiled upscale is what think is the thing I might use it for

[turbid grotto]

what this node doing in "SLG" workflow?

[sacred jewel]

Where will it be uploaded to?

Thank you..

[lavish osprey]

HF

waiting for review

[lavish osprey]

this is SGL sampling

[bitter hearth]

I don't use small GPUs but the smallest quant being just 1.79 GB is kinda cool

[lavish osprey]

runs on a mobile phone lol

[bitter hearth]

oh yeah I forgot phones
that's a legit use case for smol model

[turbid grotto]

hmmm that is nearly perfect

[muted dove]

Is that just by adding the SGL node?

[sacred jewel]

This one I assume?

[bitter hearth]

if its anything like as good as PAG was for SDXL then this is a big deal

been missing PAG so badly

[turbid grotto]

there is workflow on hf page with "SGL"

[turbid grotto]

seems like it does help

not always of course

[graceful osprey]

SLG SLG SkipLayerGuidance

Yes it should help

The underlying mechanism is very similar to PAG

[lavish osprey]

@lunar canopy how comes Dango has the yellow name and not me

[graceful osprey]

Oh that's my old days dev title still there ? Lol

[lavish osprey]

chat, I need dog plushie vs dangos fight images now

[bitter hearth]

oh awesome
in my opinion PAG is the biggest quality boost we got in last year or two

in terms of just a single node giving a boost

[lavish osprey]

one thing you might have missed with Medium, is that it works at 512x512 too

unlike Large that ONLY works at 1mp

at the end of the day, they complement each other very well

I generated oner 2k images with Medium as refiner yesterday and I love it

[graceful osprey]

SLG is not as direct as PAG in terms of layer selection given how DiT model works - but play with it - it is fun

[noble coyote]

Flux/Florence2 + LoRAs img2img

[bitter hearth]

yeah that's okay, even a little bit of PAG effect would be good

[graceful osprey]

Relevant layers:
2,4,7,8,9

[bitter hearth]

Divisive Norm and Spectral Modulation from here gives a bit of PAG effect also and works on all DiTs https://github.com/Clybius/ComfyUI-Latent-Modifiers

should stack with SLG

[turbid grotto]

6s per image on rtx3060 with 512 and looks fine!

[lavish osprey]

I'm also curious to test this stuff here ^^^

[turbid grotto]

it could go down to 3s with 8step lora and maybe to 2s with tensort

[icy coral]

Wouldn't it be wise to release your base models together with an example fine tune, one maybe done in partnership with a prominent fine tuner, so the people that are less tech savvy and more focused on quality rather than trainability and versatility can have a glimpse of what the model could offer down the line?

[lavish osprey]

hey that's naked

[turbid grotto]

oops

[lavish osprey]

I'm gonna release some finetunes soon

[bitter hearth]

the model might be stronger structurally at 512, cos in my opinion flux is strongest structurally at 512 or sometimes even 384

[lavish osprey]

less long distance attention issues with a smaller image

cause... you don't have long distances

you can also attempt a "cascade-like" workflow

where you generate low res, then upscale and refine

[bitter hearth]

yeah "cascade-like" is how I do flux, its good for these

[turbid grotto]

[icy coral]

Nice, because people love focusing on defects any base model has that are fixable relatively easily through training, and maybe releasing them together would prevent that somehow

[dull star]

how do I use the skip layer guidance

[bitter hearth]

the simpletuner dev showed some screenshots of sd3.5m fine tune testing and it already looks better

[dull star]

or do I even need it

[lavish osprey]

well, my finetune currently focuses more on artistic view than on fixing issues. But we're also working on those.

[dull star]

sd3.5 medium

https://tenor.com/view/edp-i-mean-its-all-right-gif-26702511

[turbid grotto]

where can I see them?

[dull star]

idk if euler ancestral is helping but I'm using it for sure

im gonna do 40 steps with it

since this is almost 2 it/s

https://tenor.com/view/nice-good-well-done-gif-14854677

medium still has that superior VAE compared to sd3.5 large

no speckled "dust" artifact

[bitter hearth]

terminus research group discord

[dull star]

lol dem fingers

[remote holly]

the styles are amazing

[turbid grotto]

3072х512

medium

[dull star]

https://tenor.com/view/michael-rosen-like-project-noice-smile-gif-17002790

[remote holly]

this is amazing

[dull star]

thanks

I promped it

euler ancestral, cfg 4, ddim_uniform, 40 steps

sadly it doesn't like uglyness, almost like flux

[remote holly]

thanks

[dull star]

Prompt: The man is shirtless and is barefeet and covered in mud. His long pants are old and torn. He looks thin and frail.

Reality: The man is genetically perfect, has abs and a nice jawline, he is NOT covered in mud cause that would be unappetizing... Rather, we just put a brushed him with a tiny bit of dirt. He looks like someone who is well fed and his pants cannot be torn as that doesn't look aesthetically pleasing...

[analog dome]

I know a lot about stable diffusión but one thing I didn't search and I would like someone to explain it to me what does the TX5 (or whatever it's called) do to sd3.5 ? what improvement does because im using it without it and it's working fine but i want to know what I'm missing

[dull star]

T5_XXL is a large language model, but unlike most of the ones that we use today, this one's an ENCODER too.
Some models can use it as an encoder (like CLIP I think??) to improve prompt adherence and text capabilities.
If you are mostly making images without text or rather simplistic scenes then yeah... you won't ever need it.

[dull star]

ok putting "ugly" before "man" gives him a worse jawline and a receding hairline. Perfection

he still has abs though and his pants are NOT torn

[analog dome]

thank you very much, I think this sd3.5 + the announced controlnet support will do wonders

[dull star]

@cunning lintel

[turbid grotto]

is there info about controlnets?

[dull star]

BEFORE and AFTER | SkipLayerGuidanceSD3 (default settings)

[turbid grotto]

is this sd35L?

[dull star]

medium

[turbid grotto]

looks great

[lunar canopy]

because

[lavish osprey]

coming

[turbid grotto]

256x256 lol

2s per image

native fullhd

[sacred jewel]

SD3.5M ... random seeds, all other settings the same.
prompt: what?

[vital surge]

Sorry, looks squished

[sacred jewel]

Prompt: zombies running screaming with giant billboard in the background that reads "WHOA"

Not sure if my output is indicative or I am doing something wrong? Using the workflow example from the HF Repo for SD3.5M

Definitely not happy with these results

[bitter hearth]

not necessarily doing something wrong, the model isn't as strong as the big boys

[mortal mesa]

Uh where is skip layer guidance node from

[bitter hearth]

its a core comfy node

[mortal mesa]

mmm i thought i updated, ill do it again

[keen palm]

This is 0.5 more B than I need though

[bitter hearth]

might have to do it manually from github

[mortal mesa]

ya i havent actually found it yet lol just started to look

OK it is indeed part of Comfy, had to update

[sacred jewel]

a more complex prompt using the Triple CLIP Text Encoder

[bitter hearth]

it will work on flux too but it needs adapting

obviously SAI won't make the flux node so we have to

[keen palm]

Will there be a turbo version of Medium?

[bitter hearth]

not sure Clip L and Clip G would like the big paragraphs

I don't know much about prompting so I'm not sure, but it might be doing harm giving big paragraph to weaker text encoders

[lunar rivet]

I see only downvotes under the 3.5 medium announcement, is it that bad?

[keen palm]

Think it's just nightmares from the 3.0 model

[bitter hearth]

go by Arxiv reactions rather than Reddit reactions

[lunar rivet]

I mean thumbs down under announcement here but guessing reddit isn't doing much better

[bitter hearth]

oh, reddit and discord are same anyway

it takes a while but in a few months there will be Arxiv papers covering these models, there's already a few papers that talk about Flux

[mortal mesa]

Full HD

[digital thorn]

Medium test

[signal shuttle]

I wonder how fast SD3.5M going to be on my dual 3090 setup 🤔, shame I can't test it out right now

[bitter hearth]

for the most part you can just take the parameter count to be the speed multiplier

not always true but its not far off

however Flux goes double speed per parameter cos it doesn't need a negative

[sacred jewel]

Didn't do much better with the smaller prompts either 🤭 which is why I have both short and long examples. The prompt was expanded by GPT-4o and tailored to the strengths of each encoder.

What do I know 🤷‍♂️

[bitter hearth]

seems like the image quality you got in the end was similar to the others
its a limited model compared to flux and sd3.5l

[sacred jewel]

OK, now Large vs Medium using a smaller prompt for both L/G

[bitter hearth]

yeah large is just drastically better

might be the case that medium is nice for tiled upscale, not sure yet

I think Lumina is worth looking at again also, there was a fine tune of it to 2k resolution in the I-max paper, sadly they did not release it but it would be possible to replicate it

[sacred jewel]

Large vs Medium different prompt:

Using @lavish osprey 's upscaling workflow which uses both Large and Medium models...

[bitter hearth]

showing some potential

on the upscale

needs a bit more aesthetic finetuning or preference optimisation

[sacred jewel]

That's a mighty hitchhiking thumb

SD3.5L

[carmine vessel]

In general, how does this perform in comparison to 3.0?

[sacred jewel]

SD3.5L original vs. Upscaled result through Medium (Lykon's upscale workflow)

[signal shuttle]

Apparently SD3.5M loras can be trained in mere minutes https://x.com/peacej/status/1851288045712191572

Jerry Chi (ジェリー・チー) (@peacej) on X

trained a DreamBooth LoRA on my face in a few minutes using
Stable Diffusion 3.5 Medium. 数分で訓練できました。　Updated guide coming soon.
#SD35M #stablediffusion

Twitter · Oct 29, 2024, 3:40 PM

[carmine vessel]

Mostly good, but what is the versatility in unique applications such as stylized artwork?

[sacred jewel]

Not sure... I would have to put them side by side... let me see...

[pseudo owl]

It’s definitely better at human anatomy by a mile, a bit less then large, not sure about other things.

[lavish osprey]

oh a random Jerry Chi

[bitter hearth]

3.5m hands are a lot better than 3.0m hands

its a big improvement over 3.0m

its easy to forget what previous models were like, did sd 1.5 for last week and didn't see 1 correct hand in like 1000 generations lol

I really love sd 1.5's lighting and general "vibe" though sometimes, no other model like it

[pseudo owl]

Yeah I think flux spoiled us, basically perfect anatomy each gen.

I still prefer flux but sd3.5’s is much more creative than it.

[signal shuttle]

And more trainable too

In regards of full fine tunes

[lunar rivet]

huh 3.5 medium isn't that bad after all, resolutions above 1MP don't break like in 3.5 large and the quality isn't terrible for the size and speed

[bitter hearth]

flux spoiled us yeah exactly
cos what comes out now will always have flux as context

[sacred jewel]

SD3.5M vs SD3.5L vs SD3M

same settings for all three generations.

[slender fox]

Based on what I've tested so far, even though medium outputs feel messy, it creates quite nice looking skin tones.

[sacred jewel]

SD3.5L vs SD3.5M vs SD3M ... same settings for all three

[lunar rivet]

[bitter hearth]

hands are the final boss of models

[heady copper]

my last experience with diffusion models is SD2 but by god it's progressed a lot

[serene spindle]

Because SD2 sucked

[sacred jewel]

SD3M vs. SD3.5M vs SD3.5L

[signal shuttle]

Wait I just realized SD3's "woman laying in grass" was just a copy of SDXL's "woman doing yoga" prompt

[sacred jewel]

p.s. misspelling was on-purpose.

[heady copper]

you could do some interesting stuff on SD2 but compared to now especially it's pretty mid (and believe me, I loved that model for some reason)

[serene spindle]

I liked 1.5 a lot more tbh

[heady copper]

I was mostly doing embedding training for 2 which was pretty neat but unfortunately the architecture for embeds is no more

oh well, this isn't the channel for nostalgia lmao

[bitter hearth]

you can still do embeddings 🤔

[sacred jewel]

SD3.5L vs SD3.5M vs SD3M

[bitter hearth]

SD3.5L did it 🙂

[heady copper]

Actually? That's cool- thought SDXL kinda killed it with the multiple text encoders, worth looking into I suppose

[slender fox]

Well, hands are not going to happen, I tried to generate a woman showing her hands, did 100 images, only 3-5 ok.

[bitter hearth]

I think multiple text encoders is gonna go away also

they are mostly a temporary anomaly cos its expensive to train a model to fully replace clip

[sacred jewel]

Flux for comparison...

although the only one that got the BACK of the hand was SD3.5L 🤦‍♂️

With more "encouragement" it finally got the back of the hand LOL

[signal shuttle]

Bro is holding a gun to the models head with that "(back of a hand:2.5)"

[keen palm]

Multiple text encoders with sdxl?

[signal shuttle]

SDXL uses clip_l and clip_g if am not mistaken

[bitter hearth]

yeah

clip g and clip l for SDXL

[keen palm]

Okay. I haven't seen any people experimenting with dual clip loader workflows for sdxl, so I thought it was a single.

[bitter hearth]

SD 1.5 can use T5 using ELLA

as a weird bonus

[sacred jewel]

I was like " IMEAN IT, DO NOT TEST ME! HAVE :5 READY!"

I take a Mythbusters attitude. If it doesn't burn on its own, I will MAKE IT burn.

[red nacelle]

is it tune'able tho? like what are the requirements to tune 3.5L? in terms of memory only lol

not mentioning datasets because somehow people find/make those which is lowkey crazy to think about

[digital thorn]

If lora training works this time I'm sure that it will be a really good refiner model to use...at least for my graphic card. 0.25 denoise for the image to the right. Just a quick test

[muted dove]

It smoothed the skin detail away.

[cunning mesa]

Also lost a thumb.

[sacred jewel]

across the board, it looks smoother overall.

[digital thorn]

Depends on how good we can train loras for anatomy of course. If it's even trainable...I'll believe that when I see it though

[cunning mesa]

SD3.5M is probably a good test bed for training if nothing else.

[frail shoal]

did you guys have any luck inpainting with sd3.5 medium ? Results get completely squashed. worked well as a refiner but inpainting not at all. I'm wondering if i'm doing something wrong

[bitter hearth]

are you using a stochastic sampler or a deterministic one?

for inpainting stochasticity is the most important thing

[frail shoal]

using the suggested sampler dpmpp_2m with sgm uniform scheduler

[bitter hearth]

would suggest trying the node version of dpmpp_2s, the one that lets you adjust eta and s_noise
keep s_noise at 1 and set eta as high as you can without the image breaking
that can help a lot

[frail shoal]

it was supposed to be a face

[frail shoal]

will test the behaviour with large to see if it's the same

[bitter hearth]

it should give you higher image quality as well as helping inpainting

the main downside is stochastic samplers need more steps

[slender fox]

SD 3.5M testing after work, pretty much random stuff I wanted to try. I guess it could have many uses despite of several things that don't seem to be working with it.

[gusty hamlet]

SD 3.5 Medium, M2 Mac Pro

[bitter hearth]

by far the best set of images I have seen from SD 3.5M
really good job

I think with a bit more fine tune it has potential

[frail shoal]

sd3 medium hqd inpainting problems as well. But the 3.5m version seems way worse. SD3.5 large seems to be working a bit better, but not good at all. 30 steps here, 5 cfg, 0.4 denoise

[frail shoal]

can you share some settings ?

[bitter hearth]

not sure if I could convince you to try more steps
inpainting is a hard task, a lot of steps can help a lot

[frail shoal]

how much steps for inpainting ?

[bitter hearth]

I would have done 100-150, but maybe 40-60 would be ok?

[slender fox]

I don't think I had any special settings. Tried different samplers and schedulers, steps from 20 to 40. CFG from 3 to 6.

[bitter hearth]

if you want to do like 30 steps then the stochastic sampler I recommended might be worse
for low steps, Deis and UniPC can be good

[frail shoal]

well with flux, sdxl and sd 1.5 you don't need more steps. But i hope it's a settings problem. Generating images seems fine, step swap as well (refiner), but inpainting i'm getting horrible results and low quality. Wondering what might be the culprit. I will try more steps to see

[frail shoal]

ah you are not even using the new slg node

[frail shoal]

very nice results however

[bitter hearth]

I agree flux was able to inpaint with less steps
even Schnell could inpaint lol

[dim geyser]

my macbook pro did fine with SD3.5L but it really isn't liking SD3.5M. all the results are coming out messed up.

[slender fox]

I did generate 1000+ images at one go, and I may be picky about what is OK... so there is that too 😄

[frail shoal]

but in terms of image quality is very nice

[slender fox]

Yes, I did tweak the prompts for several hours, generating images doesn't take long with decent GPU.

With SD 3.5 medium.

[bitter hearth]

ah I didn't realise these were with euler
so with a stronger sampler you could get some more quality also

[frail shoal]

what stronger sampler ?

[bitter hearth]

euler is the weakest out of any of the default comfy samplers
for deterministic, DPM++ 2M, Uni PC and Deis are particularly good
for stochastic, DPM++ 2M SDE, 3M SDE and 2SA are good, although only 2SA works with Flux and SD 3.5
you can get better samplers than those default ones but they require custom nodes

[frail shoal]

what are those better samplers, i have custom nodes for samplers

[bitter hearth]

Clownshark's node pack gives lots of stochastic samplers that work with Flux and SD 3.5
https://github.com/ClownsharkBatwing/RES4LYF
Sampler RK is the latest node as far as I know, and its got a few to choose from

if you put Eta to 0 then they run in deterministic mode

[quaint flare]

What has happened to her thumb?

[digital thorn]

Yeah, bad hands in Medium. Like I said, if lora training works it might have a brighter future than 3.0 model.

[vale oyster]

I gotta ask, as I just had to hit the wrong setting in my BIOS so my PC is in no post mode atm, as it power cycles.

So I need to rely on a cloud provider -- is there any good big names as of recently for 3.5 and flux?

I was looking @ modal.. But is runpod still a actively worked on cloud gpu provider?

[violet escarp]

I have a feeling it's going to be painful to finetune sd3.5 medium and people are just going to go back to large. I think the architecture makes it so you need a high parameter count to train efficiently. Flux is a good example. It's as big as it is because that's the cheapest to train, even if it sounds counter-intuitive.

[hallow lion]

so medium looking good huh

much fastor and higher than 1 megapixel

[pseudo owl]

yeah anatomy isn't that great but nice for its size, great for people with weak gpus.

[hallow lion]

[real terrace]

2:54 generation time here for 823x1152

In original workflow clip3 was fp16

Really like the details, they need upscalling but pretty interesting to work on

[dim geyser]

SD's introducing post mentions the Stability AI API, Replicate, & DeepInfra

[patent acorn]

the post-falling down the stairs

[dense blade]

What's your performance difference between SD 3.5M and L? I use SD 3.5M (default) and SD 3.5L (Q8) with following speeds: 3.96s/it (M), 7,18s/it (L) - both use the same settings: shift 3, 40 steps, 4.5cfg, dpmpp_2m, sgm_uniform. I expected to get better performance from the M version.

My env: Linux, AMD Radeon RX 6700 XT (12GB), pytorch 2.5.0+rocm6.1

[sullen moss]

Hmm

[hallow lion]

Red Panda?

So basically we have entered singularity when a new ai model comes out every day.

[craggy crest]

vaporware

[winged seal]

I have done a few hundred images of ranking. Red panda seems to have a little more variation in composition, mixed with really bad Dall E 3/MJV6 "Aesthetislop" which people look at and go "oo pretty" cause there is an unnecessary amount of contrast/noise/"detail"

Examples of aesthetislop, where there is just noise and "detail" everywhere to try and get the same reaction as jingling keys in front of a baby's eyes lol

I get why people like it, but its just all noise/nonsensical detail over anything logical or visually pleasing in a toned down way

[hallow lion]

fast food for the eyes

[winged seal]

pretty much haha

just overwhelm the senses so people don't take a look closer and realize its all meaningless noise and nonsensical details/tones

[hallow lion]

i like the new word aesthetislop

[winged seal]

A lot of people who do the head to head comparisons on that space go for "oo pretty" within 2.5 seconds of looking instead of taking in the composition, stylistic variation, or interpretation of the prompt

[winged seal]

my friends in my research group have been using it for nearly 2 years now haha. It really does best explain the look haha

[sacred jewel]

Zombiecore LoRA - Match made in heaven 🤭

[winged seal]

goodness, more baked than snoopdog

[craggy crest]

it's also vaporware right now

[winged seal]

we shall see, we don't even know what model it is

Blueberry was around for a bit, and now you can use it

I am just curious who its made by

[craggy crest]

we also don't know if the images you're voting for were created by something new, or something old with a new name. we know nothing about it at all

[winged seal]

you mean for redpanda?

yeah, we know nothing about redpanda as of now

[craggy crest]

yeah. it's vaporware. the images could have been created with anything. once we get a demo, and know who's behind it, then it's time to worry about it

[winged seal]

I mean, you are saying pretty obvious stuff lol

but yeah, curious to see who made it

last time it was a new Black Forest Labs model, so we will see

[winged seal]

We have been due for a new Dalle4 for a long time, honestly. Dalle 3 was a prompt adherence champion, but my god did it look abysmal

[errant dust]

We'll agree to disagree.

In any case, for those awaiting it. The GGUF build of SD3.5 Medium is out

https://huggingface.co/city96/stable-diffusion-3.5-medium-gguf

[sacred jewel]

I can probably make one of those charts in Photoshop in :30 seconds... I'll call mine Hawk Tuah and give it a score of 2001, a rate of 95% and a selections # of 76871 and profit 😎

[errant dust]

What is Red Panda?

[real terrace]

Hi excuse me, how do you load those models in Comfy?

[cobalt moon]

A mysterious AI model that no one have a single clue where it came from

It just appeared in Artificial Analysis

[errant dust]

GGUF?

[real terrace]

yes

[cobalt moon]

I think it's probably a codename

[errant dust]

Workflow is there. You need to install the GGUF extensions and make sure Comfy is updated

[cobalt moon]

like Flux's codename "Blueberry"

[errant dust]

https://github.com/city96/ComfyUI-GGUF

[real terrace]

great thanks, I found the missing nodes there

[real terrace]

Also I see now that the text encoder can be quantized too

what is the SD3-5 vae? I didn't use none before I think

[craggy crest]

@dusky thistle comfyUI has a new scheduler in it

[craggy crest]

all models can be quantized. that doesn't mean they should be

[real terrace]

great, yes I was kind of dazzled about so much quantization

at this point I load some model and it worked somehow

I don't even remember where did I get it

The only thing about these models is that they take so much I don't fiddle or play too much with the parameters, I just set prompts and img size

[dusky thistle]

what scheduler

[real terrace]

your gun is backwards sir

[hallow lion]

lmao

[craggy crest]

linear_quadradic

[real terrace]

they need refinement

sh,don't shout the prompt bro

god dammit

[sterile pendant]

Afaik, it's pretty much only used with mochi or w/e that model is called. I haven't tested to see if it works at all with other models

Off the top of my head, it should likely look like something between an sgm_uniform and simple sigma curve

ahh nvm, looks like they use some wonky exponent for it. looks like it spends a lot of time shifting things around a bunch, then rapidly drops into fine details

it's definitely meant for flow models only. sdxl models usually have concave curves and it didn't respond to that when I tried to run this with an sdxl model.

either way, it's likely not all that useful for non-mochi models.

[craggy crest]

Sd3 is a flow matching model

[sterile pendant]

It is, but that kind of curve is not what the model was trained on. So sure, it will work, but it's not optimal. It will spend far less time resolving medium and fine details using that linear quadratic curve

Think of it like model shifting to an extreme value away from the norm

[dusky thistle]

i think beta's been best

it's got just enough time spent on the fine details

[craggy crest]

Shrug

[dusky thistle]

papa toilet

[severe phoenix]

please can anyone explain to me what these two clips do. is one more powerful than the other?

[sterile pendant]

Yeah beta has been my go to for a while now

[dusky thistle]

T5 and clip L, T5 will be for sentences, L for danbouru style tags, T5 is way way more influential

[severe phoenix]

ahh thanks, what kinda tags are these?

[craggy crest]

here's the architecture

SD3 uses all three encoders, flux only uses two

clip_g is the workhorse - it drives the entire thing

clip_l and t5-xxl work along with it

t5xxl gets your detail rich, narrative, natural language prompt

clip_l gets your artsy, ambient, background, fine details prompt

clip_g gets the no-nonsense, just-the-facts-mam information about what the image is

IF you do not use a node that allows you to put in a seperate prompt for each encoder, then the text you put into the positive prompt will be given to each encoder anyway, and they will fight with each other - they aren't in sync

[real terrace]

nice explanation, I would like to experiment with that concept, it seems to be really important if they work like that

for example in this img the text that appeared was just some text to reinforce the idea, not some main title to write

[craggy crest]

here's a workflow to start with. this has the SD3-2b-medium model loaded, but you can just change that to SD3.5 if you wish

[real terrace]

it would be in the clip_l I guess

[craggy crest]

the encoders alone do lousy text. you need at least two of them for it to be readable

[real terrace]

I wonder how much different could you find when trying a single prompt and these 3 clip prompts, is there something done?

[severe phoenix]

damn thanks for t5xxl level explanation 😃🙏

[real terrace]

the idea is that I didn't want that text to be readable, it was just some text to reinforce the idea of the image

[craggy crest]

this is an SD3.5 Large workflow i'm working with right now, it also has the node for all three encoders. to get the workflow: 1. click the image to open in viewer. 2. click the Open in browser words. 3. right click, save as. then drag and drop into comfyUI

[craggy crest]

so type random letters: here we see a sign with the word "swrenwtw rerwun ewwn"

the AI will try to guess what it might be

[real terrace]

there wasn't a sign mentioned, it was "advertisement" so it make a text and put something in the prompt.

[craggy crest]

just an example of what you'd do for random text

[real terrace]

so clip g would be the classic SD prompt, clip t5xx would the "chat GTP" type promtp, clip_l would be booru tags, like "trending on artstation" ?

[craggy crest]

i would advise you to never use trending on artstation unless you just want random dice rolls for noise. and most of the other tags like it. they are just noise.

here

clip_l: trees, warped, twisted, weathered, ruined, cracked, flames, fire, smoky; atmospheric lighting; peter mohrbacher, james jean, william morris, ernst haeckel, zaha hadid
clip_g: apocalyptic ruins; pink tree growing flames;
t5xxl: the scene is set in apocalyptic ruins. in the center, we see pink tree growing flames rather than leaves; All that is left of the buildings are parts of walls with crumbling bricks

=======

clip_l: thick impasto painting, heavy canvas texture, noticeable brush strokes
clip_g: Stratocumulus; Renaissance Island beach sunrise, maya beach by artist "Paul Dougherty", by artist "Nicolette Ceccoli"
t5xxl: the scene is a stunning sunrise at the beach. cumulostratus clouds cover the sky with red and gold, while transparent green water and white foam breaks on the light brown sand,

=====

there are 2 sets of prompts to play with

@dusky thistle comparisons. model: sd3.5 large sampler: euler_ancestral schedulers: on the left, linear_quadratic. on the right, beta workflow is in the images

[dusky thistle]

that looks real good on the right

what's interesting i've noticed with this flow models is it's like it just needs to set the trajectory real carefully

almost like aiming a rifle

[craggy crest]

it does. i'm working on that sampler/scheduler sheet - should be done tomorrow

[dusky thistle]

and then pow, it can take massive steps

[craggy crest]

did you grab the sd3.5 medium workflow with SLG?

[dusky thistle]

no idea, been coding all night

[craggy crest]

skipping layers, like matteo did with flux back there. skipping the blocks

[dusky thistle]

just got a brutally, just, oh-my-gawd-don't-want-to-admit-it bug fixed that's been hanging around for weeks

cool yea sounds like PAG with zeroing out the V

[craggy crest]

https://huggingface.co/stabilityai/stable-diffusion-3.5-medium

files are there when you're ready

[craggy crest]

do you remember that really long node matteo did for flux, where he set up every block so you could adjust them?

and then he talked about how certain blocks didn't do anything and we played around with not using them?

[severe phoenix]

ohhh so clipl recognizes artists and styles and t5 is just straight up natural language. woow i'll try this thanks!

[dusky thistle]

it's great having comprehensive nodes like that

i make shit like this all the time

[craggy crest]

okay, that's what SLG is. only not nearly as long and annoying a node as those

there are two workflows on that repo, one with, one with out and you should probably DM lykon about that setup you have and talk to him

[dusky thistle]

sounsd good

gonna have to go pass out soon here but will do

[craggy crest]

and i want your workflow

i can't find the one matteo released

[dusky thistle]

3rd order RES is looking pretty nice with ultracascade

which WF?

[craggy crest]

he's got teeny wings coming out of his neck

[dusky thistle]

hahah yeah

[craggy crest]

the one that screen shot was from

i'll bet those nodes aren't avaialble, are they?

[craggy crest]

add the term symmetrical to the prompt as the first term

[dusky thistle]

they should be

https://github.com/ClownsharkBatwing/UltraCascade

i think they're in there, if not i can get em to ya

[craggy crest]

i'll go poke around and let you knwo if i can find them or not

DM lykon, and go crash

[viral plaza]

@lavish osprey @graceful osprey SD 3.5 Medium modelspec resolution is 1440x1440, however the model struggles to run this resolution natively: (it's fine at 1024x1024, and I bet it can upscale to 1440 happily, but native from the ground at 1440 doesn't seem to work too well. Maybe with the skip layer guidance, longer prompt, luckier seed, and all you can push it into working? idk even then I feel like recommending 768 or 1024 with an upscale is smarter than recommend native start at 1440)

that's a photo of a cat and that staircase it's inside of, is the early steps completely corrupted, and the latter detail steps managed to refine it into something that almost looks like a real structure

[bitter hearth]

Flux doesn't do well at the advertised top resolution of 2048 either

unless you are using I-max

[turbid grotto]

i was able to generate FullHD

but not first try

[viral plaza]

To be clear, I'm not discussing the advertised top resolution, I'm discussing the defined default resolution, which inside of the released SD3.5 Model is set at its top, but should not be

that's defined in the metadata header of the model if you're unfamiliar with modelspec

which https://github.com/Stability-AI/ModelSpec?tab=readme-ov-file#inferencing-tools-and-uis Apply relevant keys where logical (eg resolution in image models should be applied as the default resolution for images made with that model) that key is intended to be the default usage resolution

[bitter hearth]

ah thanks
I knew there was metadata inside .safetensors files but I have never looked at them

okay yeah you have a good point then

[noble coyote]

Clip_l, Clip_g, Clip_round_the_ear - what does this one do?! 🥳

[muted dove]

With questions like that, you'll soon find out my lad! 😄

Even the strongest have their weak days.

[bitter hearth]

on Civit there is someone adding Flan-t5-xxl to Flux and SD3.5

[patent prism]

a cat

[muted dove]

I doubt it's a cat that's adding it.

[real terrace]

so what are the best clips (!?)

I'm trying this

[real terrace]

ty, I have to try that

[meager patio]

Hello everyone, I really need help. I was struggling all night, installing a stable diffusion, literally for one job, and right now I'm launching it for the first time and it drops what's in the picture, talking about old drivers. I can't update it because this is the limit for the video card.

I want to ask someone to do literally one job. I'll be extremely grateful if someone responds.

[real terrace]

This large model took as much as the GGUF sd.35 medium q5_0

[bitter hearth]

the big speed up is only gonna come once you have a setup that fits within your VRAM

regarding clips, some people like this for Clip-L https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14
although I am not sure about it

[sterile pendant]

Another big key difference about sd3.5 medium is that the t5 should contain no more than 256 tokens

[bitter hearth]

Flan-t5-xxl for T5 I am more sure about

[real terrace]

what what setup that might be?

[bitter hearth]

make sure the text encoders get put onto the CPU before the Ksampler starts
and try Q4_0

or just offload the text encoders entirely

after text encode

[sterile pendant]

And SLG works well with 3.5 medium. Don't know about large though since medium has the slightly different architecture with extra self attention layers or w/e it was

[bitter hearth]

SLG will work with Large and Flux I think
just needs porting and recalibrating

its a clever idea, it drops layers for the negative so the structure gets messed up

[fossil pagoda]

👁️

[bitter hearth]

since models try to do the opposite of the negative, you end up with better structure

for Flux this will require generating a negative, with the 100% speed penalty

would be good to try this method with Perp-Neg also

[muted dove]

SD3.5L using Google FLAN and Flux refiner.

[remote holly]

what are the resolutions suported by sd3.5 ?

[muted dove]

I just use the same as SDXL, which are in the "SD3 Select Latent Resolution" node

[graceful osprey]

Hmmm weird. 1920*1024 should just work out of box

ah but you might need a higher shift

like 4~5

[viral plaza]

1920x1024, Seed 1, a photo of a cat, CFG 7, Sigma Shift 3 on left and 5 on right, Sampler Euler scheduler Normal, Steps 20
in both cases while watching live previews it's clear the early steps are failing and the latter steps are trying to recover the corruption into something coherent (in this case a TV screen type of effect on the background, and a cat with a patchy outline lol)

[graceful osprey]

Steps 20 - try dpmpp_2m steps50

should work - let my try as well

[viral plaza]

here's that at the end. The two other images are intermediate previews just to show what it's doing in the early steps

for comparison at 1024 the early step previews have a relatively clean coherent image, just without details built yet

[bitter hearth]

In Flux I got better results with stochastic samplers
the built in DPM++ 2SA in Comfy has been adapted to work with ret flow

[graceful osprey]

Noisy beginning is kinda expected but should not bu this bad

[viral plaza]

and also those early previews resemble the final image moreso than the high res

[viral plaza]

yee

[bitter hearth]

I've been trying to port restart sampler to ret flow, it might help with these models
its given me the best results in SD 1.5 and SDXL

it adds the noise in a special way, its a bit tricky

[remote holly]

what are the schelduders and samplers ?

[remote holly]

does someone knows a prompt system for prompt enhancing

[noble coyote]

Flux/Florence2 + LoRAs img2img

[severe phoenix]

omg i never knew stuff like this existed! is there one for img2img

[turbid grotto]

hi, I found same problem when not using T5.
1,2 - clips only
3,4 - clips + T5

[flat oracle]

why is my workflow i used for sd3.5 l not working for sd3.5 m

[viral plaza]

oh shit, i'm so used to ignoring T5 with SD3 since it did almost nothing with the SD3.0 models, is it actually needed for SD3.5 Medium now?

[flat oracle]

you can run with 1 text encoder, 2 or all 3.

[viral plaza]

i know that lol refer to context of the conversation above, it's a much more specific thing

[turbid grotto]

i have been using with no clips too and saw no errors until increased resolution above 1024px

[flat oracle]

you asked if t5 is needed. The answer is no. lol

By any chance you know why my workflow is not working?

[viral plaza]

again, within a specific context of the conversation above

[flat oracle]

this is an open chat sry dawg. Maybe reply to the message

[viral plaza]

seed=1, steps=20,cfg=7, CLIPs + T5, Steps=20. Seems still wonked out. left is 1440x1440, right is 1920x1024

oh i left dpm++ on instead of euler, slightly different details with euler but still wonked

[turbid grotto]

"a photo of a cat"
seed 1
steps 20
1440x1440
dpmpp_2m
sgm_uniform
cfg 4
fp16 everything

[turbid grotto]

maybe too high cfg?

hmm there is stripes in background but seems to be gone if I specify

[viral plaza]

that looks very similar to my results on euler at cfg 7. It's clearly corrupted on the background, and the cat itself doesn't look great (those eyes are barely existent)

[viral plaza]

the specified background reduces the visibility of damage, but that image still looks pretty off

i think this is just a case of, the model wasn't trained enough at or above the 1440x1440 res, and the more input you give the more it can compensate and make it work, but if you give minimal input the difference shines aggressively

this is a general case with image diffusion models - they are largely self-correcting, so giving it more of anything helps cover issues. More prompt tokens = wider attention = more self-correction. More steps = more actual entire runs of the model = more self-correction

(relatedly in the other direction, generating with empty prompts is a great way to sus out implicit biases in a model, whatever it outputs without conditioning will be approximately representative of the types of content it was trained the most on and may have developed a general bias towards)

(in that test, SD3.5Medium has a very nice broad range, and seems to favor 768x768 for the most stable generations, even at 1024x1024 with empty input it displays some striping patterns)

[turbid grotto]

yea there is definitely undertraining in some cases but I do not judge it much as it is base model and will be improved if adopted by community

[viral plaza]

yee

[turbid grotto]

does that means model is unbiased or that is bad thing?

[viral plaza]

that means model lacks any visible bias, which is good

I've seen some models eg continually generate humanoid outputs on empty prompt, indicating that the model was violently overtrained on humans to win on "omg it makes pretty girl" aesthetic evals but otherwise useless as a foundation for anything

SD3.5Medium has a very broad range of random outputs from that, indicating it likely hasn't been overly tuned on anything in particular, making it very optimal for a base model for community tuning adoption

(noting however that's just a quick n dirty test, not a guarantee, so grain of salt and all that)

sd35m promptless outputs

maybe a touch of over tendency towards artsy stylings but eh

and for comparison, Flux Dev

burnt hard on high quality photoreal art

that woman included is The Flux Woman who everyone sees constantly, and the community have taken to identifying by way of her specific "butt chin" shape

[turbid grotto]

majority of finetunes
but it is probably inevitable future of sd35 to become coherent. I compared sd35 with flux on "lying on the grass" and sd35 often tried to do the hardest thing, like upside down, while flux often choose the easiest route by generating non-upside-down and non-horizontal (idk how to name it)

[turbid grotto]

yea I had similar result

so, you either have diverse model with lacking coherence or coherent with lacking divercity

[viral plaza]

yeah, i went on whole rants about that during the sd3 launch. no model is good at upside down woman, sd3 was just the one model that tried to do it without asking

[viral plaza]

i think the ideal option is the diverse capabilities, but with something to avoid it defaulting to things it's bad at. eg an LLM prompt augmenter that takes short prompts and writes long prompts for it that don't suck and give good defaults for things the image model might make silly choices for

it's funny how often people have said my discord bot running flux schnell looks better than their flux dev results (or similar comparisons)... the model's worse, but my discord bot has an LLM that extends your prompt, and that does magic

[lavish osprey]

not sure if it's just the distillation, but those all look like MJ images. Synthetic data, not just any "photoreal art". Especially the woman "sameface" with the butt chin, that's 100% MJ

[lavish osprey]

yeah, another artifact on training on real world data. If you go on photo datasets they're usually using various different angles. Models trained on synth data tend to overfit on upright position or, in general, simple poses.

[signal shuttle]

After playing a lot with flux and SD 3.5 large, i began to forget how fast 1s/it was, Man SD 3.5 medium is so fast

[noble coyote]

Budget's just finished - penny off a pint of beer!

I'm off down the pub!!! 🥳

[flat oracle]

if anyone wonders. got it fixed updating comfyui via manger.

[noble coyote]

<---------------- needs OmniGen on ComfyUI

[turbid grotto]

I probably should add some tiny llama to extend prompts

I often see people saying sd35 is not cinematic\realistic but that is sooo wrong, they just used to models defaulted to certain style. In that cause you have to prompt to get it and you will

[turbid grotto]

agree, even if I'll have to run it for a 5min, it would be interesting, and I saw it is actually great, but suspiciously reminds me flux's aesthetics, ||but I might hallucinate||

[viral plaza]

#1243166025000943746 message relevant omnigen testing in the swarm discord earlier

[noble coyote]

Its the power to meld disparate elements of different photos - seamlessly - which excites me

[viral plaza]

seems like it has cool capabilities but its general image gen quality is very "meh" tier, making it hard to justify the disgustingly long gen time

[muted dove]

The more you drink, the more you save! 🙂

[noble coyote]

My 8Gb VRAM 'excluded' me from using Omnigen on pinokio

There is an Omnigen Huggingface https://huggingface.co/spaces/Shitao/OmniGen

[signal shuttle]

(SD3.5L + SD3.5M)

[errant dust]

Maybe I am just choosing the wrong subjects, but so far I have been underwhelmed by SD 3.5 Medium. Compared to Large of course. (workflows included, but prompt is basically: "a comic with strong outlines of a tree that stands tall in the center, with cats of various sizes, colors, and expressions scattered around its trunk and branches. Some cats are nestled among the leaves, while others are perched on the tree's main branches. The tree's trunk is brown, and the leaves are green. The cats exhibit a range of facial expressions, from curious to content." First is SD 3.5M at 768x768, then 1024x1024 and finally SD 3.5 L

same seed

[untold valley]

Anyone have a quick TLDR of the state of things? How is 3.5 compared to last released model or Flux? seems there are stripping patterns is that the worst of it?

[errant dust]

I have not seen striping issues myself, but it seems to completely whig out at sizes over 1024 x 1024. Large that is.

the output is quite different and they have different strengths and weaknesses. If you want hands, well, Flux is your choice, but there is more to imagery than that. Comics are clear win (no Loras) for SD 3.5 L

[untold valley]

other than that all good?

what about text?

faces/body how maleable is it?

can it do the dreaded woman laying on grass

astronaut on the moon

[errant dust]

The thing is I don't really do stock photography with people. But that much is fine. Just try it IMHO. Nothing endemic

[rapid pivot]

Hello Chess man

[untold valley]

comfy still the best way? forge? anything developed better?

[rapid pivot]

[errant dust]

Hey @rapid pivot

[pseudo owl]

Large can do woman in grass too, but obviously worse then flux in anatomy. Flux is also better at text.

[errant dust]

It would imply I am trying a bunch of other GUIs. I tend to stick with what I have if I am satisfied and it does what I want and well

and COmfy fits that description

The trick is to explore them with the themes and types you want, and leanr which are best at what

[untold valley]

so essentially Flux is currently outperforming and we will need to wait for 3.5 models to be retrained by enthusiasts? were legalities and licensing changed to encourage more of this or we still at old restrictive ones?

[errant dust]

I can tell you my findings, but they only cover what interests me. So anime? No clue. Chicks with boobs? Check Civit. Etc. Artwork? Stock imagery. Creative content? Text? I'm your guy

I mean, in terms of text, the clear no.1, of any AI, is Ideogram 1.0 and 2.0. It isn't even close. Flux isa decent second, but distant. Logos? Flux is the king today. All the others. MJ included, are just behind

MJ6 that is. MJ 5 was strong

Satirical comics? Imagen 3 is the best now. Also not close.

[untold valley]

afaik Flux hard to train and that's why we were waiting on 3.5 or model like it. and it appears 3.5 needs more training tossed at it.

[errant dust]

I don't fanboy any model. I just tryt to leanr what the best tools are for each case

[untold valley]

looking for local

[flat oracle]

sd3.5 M works pretty good for me. However, CLIP Text encode node (positive as well as negative) takes extremely long to load, any ideas why?

[errant dust]

Have you tried the GGUF models?

[flat oracle]

Sorry, forgot to mention, thats what im trying rn

[errant dust]

I use the T5 Q8_0. WOrks fine for me. Though YMMV

[untold valley]

ah damn, do we also now have different T5 models we need to deal with?

[errant dust]

Flux has a big edge RN in that it has been out for 3 full months now. So TONS of strong LoRas. SD 3.5 hasn't had this level of development yet, so give it time

[errant dust]

GGUF means it is quantized. Designed to load and be processed faster and with less memory needs

as it is loaded in blocks

and not as a whole

[noble coyote]

Changing to 3.5 brings a tokenizer fault at the negative prompt

[untold valley]

models need to be trained, not just loras. the downside of Flux model, however we go back to how easy/ licencing for 3.5 is otherwise wont get models created

[flat oracle]

im just wondering because im even trying with dual clip loader t5 & clip l + gguf variants of sd3.5.
Shouldnt it be way faster than flux 1 q8 with both of them?

i mean the ksampling is pretty fast. no problems at all, just with text encode

[errant dust]

I use three clips, but have not bothered to time the entire lot. I should and will, but overall I have found them to be pretty close. At least not different enough to warrant Ohs and Ahs

[errant dust]

I am talking output results. For me the talk about what is the driver for an improved image, be it LoRA or checkpoint, is semantics. I don't actually care so long as I get what I want from the combination.

[flat oracle]

kk thanks!

[untold valley]

you can theoretically have an endless stream of loras, but I do not like to stay with a single acute style. Its easier when a model inherently knows what it needs to. and makes the image generating creative workflow more malleable. essentially Lora's are an overfitment run on top, sure easy fast way of getting what you want if you want specifics but as a whole a model by itself that understands you is better.

[errant dust]

Reminds me of an arument I had with a strong chess master who offered to mentor me to mastery myself. He set out this study plan that involved self studying matches as far back as the late 1800s so I could 'absorb' the evolution of the game and blablabla. I told him I would prefer a more focused and pragmatic approach that maximized results (so long as they did not impair evolution). We soon parted ways. I don't have infinite time, nor patience.

How is citing endless culture and names going to matter if I lose because I simply played worse? Same for AIs. Talking about why one way sounds nicer on paper is unimportant to me. If I get the results I seek, all the technical perfections of another setup or method won't matter one whit to my eyes.

[untold valley]

i am asking the wrong person then. was looking for more of a power user, jack of all trades approach.

[errant dust]

I wasn't aware you were asking anything at all.

You asked what is best. I told you in results. You then railed about why models and Loras should be splitting the output development differently, which is semantics. But there was no question involved.

Is it the purity of some technical aspect that matters to you? Or better images that fulfill your requests?

I made clear where I position myself

Devolving into personal attacks? Is it that hard to just have a civil discussion?

[viral plaza]

If you were expecting 3.5 Medium to be equivalent or better to 3.5 Large... well your expectations are backwards. Medium is smaller than Large. It's expected to be a bit weaker on quality, but with the perk of running fast and using less vram

[viral plaza]

SwarmUI is the best UI

[untold valley]

many thanks, will give it a try, took a hiatus from SD and trying to find where thing are at as a whole. Swarm is a gui with comfy backend right?

[viral plaza]

if you use dualclip (CLIP G + CLIP L) in non-gguf, it should be blazing fast and near-equivalent results to triple (the T5 on SD3 is relatively weak, but the most intensive to run)

[viral plaza]

yes

[errant dust]

I quite understand, but the images I showed emphasize issues in the output that make it unusable. For those topics as I explained

I used SD 3.5 L to show that it wasn't a general failing by SD 3.5.

That said, here is the type of image that Flux can only dream about doing, but that SD 3.5 can do:

[pseudo owl]

Depends on what you mean outperform, but yes imo. Sd3.5 is more creative but flux is obviously better at text, anatomy, and sometimes prompt following.

So sd3.5 is a great tool too, and it also has more knowledge but from what flux knows, it performs better.

[pseudo owl]

Prompt?

[untold valley]

thank you, appreciated

[errant dust]

in the image

all settings and meta are always in images I post here

an Impressionist Cartoon of a tree covered in whimsical cats on the base and branches all drawn in a variety of colors and facial expressions in the style of Andy Kehoe and Skottie Young. The outlines in the layered 2d art are strong and reminiscent of Keith Haring.

[dry wave]

from what I heard flux is also easier to train, despite being a distilled model

guess the way to go for now is using sd3 for creative arts and flux for more realistic/clean images

[pseudo owl]

Maybe for Lora’s but not for full finetuning I believe, and yeah i think the flux and sd3 for different tasks is a good idea.

[errant dust]

How about the 32-bit T5?

[dry wave]

who needs full fine-tuning

[errant dust]

huge file even GGUFed. (18+ GB)

[viral plaza]

blink

what

[errant dust]

I meant F32, sorry

[viral plaza]

what about it

[pseudo owl]

It’s better than Lora’s especially if you want to make the model learn more. Loras are only good for small scale learning(one or few styles)

[errant dust]

You said the T5 for SD3 is weak. I asked whether you expected tangible benefits using the F32 version

[viral plaza]

GGUF makes things smaller by reducing the bitwidth, so if you explicitly use a high bitwidth yeah it's big

[viral plaza]

no, the problem is that the model's backbone (the MMDiT part, that does actual image generation during sampling) doesn't particularly care about what the T5 has to say, because CLIP G provides much more useful information during training

the bitwidth doesn't enhance anything, it's just narrow refinements to precision - fp32 requires twice the memory of fp16 and microscopically more precise data, it's incredibly pointless to use fp32 for anything outside of training

even fp16 barely provides more precision data than fp8 does

[errant dust]

I use Q8_0 myself

a cross between them, though very very close to fp16

[viral plaza]

(this btw is the reason why Flux doesn't have CLIP G: by removing the clear G signal, they force the model to learn T5, and once it's willing to learn T5 it can achieve better results in the long run)

[errant dust]

That's quite interesting. I assume this was understood before SD 3.5 went into training. Any reason why this path was not chosen?

[signal shuttle]

(SD3.5L + SD3.5M)

[errant dust]

Hmmmm! Playing around with the samplers with SD 3.5L dpmpp_2s_ancestral just gave me my best result by far in my Keith Haring cats in trees prompt. So far my request for his style and strong simplistic art and outlines was ignored, though it did adhere to my desire for rich multicolored whimsical cats. To be explored:

[craggy crest]

you're dissing someone that is what you're looking for. this makes little sense to me

[craggy crest]

use the workflow in this then, this is 3.5 large

[untold valley]

we will do this another time, in the mean time have you found SLG workflow usefull or you running it without it?

[craggy crest]

i'm currently finishing the sampler/scheduler compare sheet. however i did work with it some yesterday, and i played around with the same idea when matteo released his block skip node for flux a couple months back. it's very useful, but you need to be extremely careful with the values. one interesting thing is that by enabling skip, you may get better text - but you will also lose fine details. image with skip turned on, on the left. image with it off, on the right

it's implimented in 3.5 medium as a way to tweak things like hands and feet if needed, or other things that are warping that you don't want to warp.

it's not intended to be a 'turn it on and use it on everything'

[untold valley]

they are completely different pictures all together, guess there is no free lunch. results for skip for text based on that sample are crazy though.

[craggy crest]

skipping layer 2 only - skip on the left, without skip on the right

you can see that the fine details are deleted

[untold valley]

wow yes, right picture looks really nice with the splattering of paint

[mortal mesa]

i dont think its barely been touched yet, people plug it in and use the defaults in it, can have lots of effects but i sure as hell aint gonna figure it out

[craggy crest]

if you're referring to the block skip for unet called free lunch, this works in the same way, but the blocks we use for SD3 are NOT a unet structure so you will ahve to carefuly play with the values and layers to get an idea of what you want to turn off or on and when

[craggy crest]

i am. that's the project i start after i finish this sampler/scheduler compare sheet

[untold valley]

there is a lot of possibilities with this, maybe there is a magic combo somewhere, will need to go see the paper to see if they talk about what layers represent what aspects

[craggy crest]

actually, you really want to go talk to matteo. or talk to @dusky thistle - they're the ones that have done the most digging into this sort of thing

[untold valley]

thank you for the resources, let me dig into and familiarize myself with it first.

[sacred jewel]

Nihilism LoRA

[untold valley]

@craggy crest top is SLG lod have mercy, really hard to find a sweet spot, finding start percentage the higher it goes the more distorted, needs low initial value.

[craggy crest]

heh. yeah, it's not an easy, point and shoot, idea. the people that have dug the most into what each of the blocks are actually doing are clownshark and matteo. and yes, you want very small values. not sure if your node will allow decimals but if so, try them. 0.001 vrs 0.00001

[viral plaza]

i assume just keeping matched to SD3
ie avoid too much arch change until there's enough to justify an "SD4"

[craggy crest]

https://replicate.com/stability-ai/stable-diffusion-3.5-medium SD 3.5 medium is up on replicate if you don't have a machine that can run it

[hallow lion]

Cats with 4GB VRAM (send help)

[craggy crest]

@bitter hearth

[craggy crest]

just type @cat and then pick from the list to tag

@dusky thistle one of these is dpm_2/linear_quadradic, the other is dpm_2/beta - 3 guesses which is which

[sacred jewel]

I will guess Quadratic is the more detailed one (second one)

[mortal mesa]

Beta/Linear_quadratic, worked very nice on this one, my opinion

[craggy crest]

ddim_uniform tends to warm images up

[bitter hearth]

linear_quadratic is a bit of a spicy choice
sprints through low sigmas at lightspeed

[craggy crest]

dont' trip at that speed...

[mortal mesa]

sprints through low sigmas at lightspeed```

[craggy crest]

@bitter hearth

[mortal mesa]

i dont really know what im doing but i got detail daemon in there too bosting early if that even makes sense

[bitter hearth]

detail daemon will help with linear_quadratic
because it will offset some of the detail loss

so yeah that makes sense 🙂

[craggy crest]

these are all dpm_fast. one is exponential, one is karras, one is normal

can you guess which is which?

[bitter hearth]

never actually checked out what exponential graph looks like TBH

I think the third one is normal

hmm not sure

[craggy crest]

;) now you have to go play

[real terrace]

👍

I still can't fully understand what to put in each clip; sometimes I feel like I have it clear, but then I don't.

If I have to repeat or reinforce the idea in all

I'm using this GGUF model, but it takes 4:40 minutes, while

this large model I think it took 4 minutes only

[signal shuttle]

GGUF models trade speed for less vram requirement, so a GGUF model will be slower then an FP8 one while also being much lighter on the vram requirements

[real terrace]

now that you say it, with this model my PC haven't go OOM or stutter, that's true, iirc with the large one there was some steps where videos would stop playing for a bit

[pseudo owl]

The red panda model which was #1 on the text-to-image leaderboard(beat flux.1.1 pro, flux.1 pro, dev, schnell, sd3.5 large and turbo): <
https://x.com/recraftai/status/1851706399631224939>

It has very bad prompt following(maybe only in this prompt idk), but great realism.
prompt: a high quality photograph of a white cat sitting on top of a blue dog on a brown couch in a living room. Behind them, is a window, with 4 cow pictures, one in each corner. Outside the window is outer space and a ufo.

Flux.1 schnell 8-step on left, Recraft v3(red panda) on the right

[dull star]

indeed its finally revealed on the image leaderboard too

idk man it looks average

especially if its closed source

[lucid swift]

bs its very good

[pseudo owl]

you will see realism and text is pretty great but prompt following is very disappointing. No reason to use it over flux or sd3.5 large imo.

[lucid swift]

recraft

[dull star]

I guess it does do "crappy style" photos out of the box, but that's required for it to be used as a service with no lora options

A close-up, realistic portrait of an elderly man dressed as a military soldier. He has deep wrinkles, white stubble, and a stern, weathered expression, symbolizing years of service. The uniform is slightly worn, with medals and insignia visible, and a green camo pattern typical of an army soldier. His eyes reflect resilience and experience, capturing the weight of his journey. The background is blurred, focusing solely on his face and upper uniform, creating a dramatic, respectful portrayal.

the texture is nice

[lucid swift]

use the hard flash style option it looks much better

[dull star]

I don't want to login

idc that it's free

wake me up when its open weight

[pseudo owl]

Yeah its nice at realism, but as I said above, prompt following isn't great. Open source models can do similar gen's too in terms of realism and have much better prompt following.

[dull star]

and its gonna be paid sooner or later

[lucid swift]

it is payed

[dull star]

how much per image/credit

[lucid swift]

but u get like 50 free images

and then 5 free every day

but after then 10 or 20 dollar a moth

if you want omore

you can also ulode images and make your own style

[pallid ginkgo]

kinda like those ai video websites

[lucid swift]

yes

[pseudo owl]

Flux schnell(considerably worse then flux dev, and sd3.5 large) on the left vs Recraft v3 on the right,

prompt: A blue block on top of a red block. Next to the red block is a green block with a candle on top of the green block.

[lucid swift]

it seems like its worse at promt understanding

i cinda like this image

reminds me of something idk

but it has less often this fake real look with that smoth skin

[dull star]

aesthetics are good though

[dull star]

oh like ipadapter for SDXL

like there you could load an image and it would make an image in that style

granted, it wasn't perfect, but it was good enough

[pseudo owl]

Yeah its considerably worse then the new open source models on prompt following(flux dev, schnell, sd3.5 large, turbo) imo.

But yes its for sure more realistic then out of the box models, but there are many ways to make models more "realistic".

[craggy crest]

t5xxl: give it natural language that is rich in details and adjectives. clip_g: just give it the basics of the image. clip_l: give it all the fine details and artsy stuff. example: t5xxl: closeup on a vase with roses in it, dew sparkles on the petals. the light shines in from the side at a slant casting shadows across the scene. clip_g: roses in a vase near a window. closeup. dynamic light and shadows. clip_l: sparkling bits of light bounching off dew on the rose petals. soft, bokeh background.

[bitter hearth]

the model does look good here
during the testing period on Artificial Analysis I mostly voted against it though

my preferences seemed to be for Flux Pro 1.1, Ideogram V2 and Midjourney V6

however these ELO tests are unfair to the open source models

because they get fine tuning and Comfy workflows

but ELO test doesn't reflect that

[lucid swift]

when i tested it was always very close most of the time promt understanidng was not the porblem

[bitter hearth]

its close yes, at the top of the leaderboard

I did come away with the impression that the gaps between models are very small

[lucid swift]

at some point you cant make it more realistic

[bitter hearth]

there are upscaled SD 1.5 images that look like photos even

the difference between workflows is like 100x larger than the difference between models

[lucid swift]

but these are very limited finetunes

[bitter hearth]

yes but I don't see the downside in checkpoints specialising

you could train a small router model to route prompts to appropriate checkpoints for example

I guess storage, and loading/unloading to VRAM is the disadvantage there

[craggy crest]

SD3.5 medium. Prompt: fantasy painting of a handsome lion Knight,long wavy hair, slight smile, piercing green eyes; emerald, symmetrical,intricate engraved armor; hyperdetailed. the words "Kings wear Crowns"

[craggy crest]

too many checkpoints after a while and you get lost in the clutter

[bitter hearth]

Civit does feel like that

interesting that even Midjourney chose to have their anime checkpoint separate though

[low stone]

yikes that recraft is rough. I just tried about 10 prompts that look amazing on flux and was really unimpressed by recraft. I'd probably pick sd 3.5m over it for looks (just don't prompt for hands).

[sullen moss]

Refract vs Ideogram

To be honest, I still don't understand why this model is in first place...

[pseudo owl]

Yeah I mean it’s great at realism but it seems even more constrained then flux and has much worse prompt following.

[low stone]

that refract would be decent on its own, but that ideogram pic is killer and makes refract look like sdxl

[craggy crest]

at a guess, it's some one's attempt at a flux finetune.

[bitter hearth]

its a big one, going by API pricing and timings

[winged seal]

@craggy crest Wanted to share that I have taking a bit of a "break" after all from bigger over arching training, in favor of my first concept training on Flux Lite, which seems to be working extremely good for my first attempt, so I am happy about that haha

[bitter hearth]

the aesthetic fine tune is a bit off, especially small details

its got signs of a strong model though in composition and blur

[craggy crest]

cool - pictures or it didn't happen

[winged seal]

Training a dappled sunlight LoRA. Cooking up some more examples, just real slow cause I am on my 3060ti lol

my dataset is pretty small and less than ideal, so I will be trying to get good enough results to supplementally train off of for better feature reach

[craggy crest]

prompt?

[winged seal]

striped dappled light on the face of a young black boy. The light on his face is in striped and he is wearing a gray tank top with very short black hair, dark background

[bitter hearth]

right hand side definitely better in both cases, nice job

[winged seal]

thanks <3

[craggy crest]

sd 3.5, no lora

[winged seal]

mixed in a smidge of my realism LoRA's to make it look a touch better as well haha

[craggy crest]

just your prompt

[winged seal]

hoenstly, not bad

[craggy crest]

modified your prompt to:HDR photography: striped dappled light on the face of a young black boy. The light on his face is in striped and he is wearing a gray tank top with very short black hair, dark background

sd3.5 large, no lora, just the prompt

[sacred jewel]

Comic Book LoRA

[winged seal]

its a decent beginning for that as a concept, nice

[craggy crest]

you really like the word 'whoa!' don't you?

[craggy crest]

and no need to sit there and battle fo rhours and hours and hours with a base model that's frozen and doesn't want to do what you want

[winged seal]

ok, this looks ASTRONOMICALLY better haha

[craggy crest]

??? what... does?

[winged seal]

Oh yeah, I talked to two people with doctorates in this scene, and they both agreed that flux is not "frozen" and that it was a completely false claim made by some dude on reddit with no real truth behind it

They spoke way higher level about it with me, much higher level than I understood, thats for sure but yeah, two different people were 100% sure that "frozen" is a BS claim from people who don't understand flux or distillation. One said friend was the creator of Libre Flux and the writer of the paper used to prune flix down from 12B to 8B, so I would wager he knows what hes talking about 😅

[winged seal]

this

[craggy crest]

i coludn't care less what they say. i know what was done, they don't

[winged seal]

casually doesn't read the part where they wrote an entire paper about flux and pruning/de-compressing it

[craggy crest]

casually reminds you that i'm a programmer and said i 'know' not 'im assuming'

[winged seal]

ok man, I will continue to believe people who are much more qualified 😅

it doesn't matter anyways, flux trains great regardless of what some people try to say about it 😅

[craggy crest]

i highly doubt they are much more qualified. just that they support your assumptions.

[winged seal]

do you have a doctorate in machine learning?

[craggy crest]

i don't need one. if that's a requirement, then robin rombach, the creator of stable diffusion, is unqualified

[winged seal]

actually no, I am not gonna get into this, it literally does not matter in the grand scheme of things

anyways

[craggy crest]

of course it doesn't. you can sit there and pull your hair out, melt your gpu, and pay outrageous electric bills to train unnecessary loras. i don't care.

[winged seal]

how are things going with medium? I heard its pretty good

better or worse than large?

[craggy crest]

you haven't played with it yet?

[winged seal]

No, I haven't

I am waiting for better training tools

[craggy crest]

ah. well - get a few moments to go play with it and see what you think - try using it as a refiner for your images generated with flux

[pseudo owl]

Well considerably worse in anatomy at least but ok in art styles. Decent as a upscaler/refiner.

[winged seal]

jesus, worse in anatomy 😭

I guess it is more of a refiner after all, so oh well there

[craggy crest]

it is a smaller model - that's why you can also use SLG with it if you need to tweak the anatomy

[mortal mesa]

how would you do that

[craggy crest]

a lot fo people have been playing with it as a refiner

[winged seal]

Yeah, that seemed to be more of the direction people were leaning in for a while

[craggy crest]

grab the exmaple workflow with the SLG node for comfyUI, update your comfyUI, and then play around with the values

[mortal mesa]

oh yopu dont have any input

[winged seal]

I'd be down to use it on my flux lite gens

[craggy crest]

you asked how you would do that. not what tests i'd done?

[pseudo owl]

Yeah that can help a bit, it’s far better then sd3 medium in anatomy but aroundish base sdxl in anatomy(worse without slg).

[winged seal]

jesus thats rough

base SDXL was a bit of a nightmare haha

[craggy crest]

there are 3 workflows in the huggingface repo for medium: SD3.5L_plus_SD3.5M_upscaling_example_workflow
SD3.5M_SLG_example_workflow
SD3.5M_example_workflow

but update your comfyUI before loading any of them so you get the new node and scheduler

[winged seal]

Is hould know... I have trained over 1000 LoRA's on base and was part of the beta program for it where I gave feedback on it before public release

[pseudo owl]

I would kinda expect it since it’s smaller then sdxl. But it’s workable at least, not as bad as sd3 medium.

[winged seal]

I mean, its basically the same size as SDXL, and its over a year newer with new tech, much more training, and a WAY more mature industry

[craggy crest]

it IS sd3 medium, finished

[mortal mesa]

was wondering how you would tweak SLG for anatomy as you said, but no info

just looking for info

[winged seal]

I still think that medium will be far more viable than large, so I wil have my eyes on that if anything

[craggy crest]

SLG is skip layers guidance. so you are turning various blocks on and off, and changing values. as an example, this is using it to tweak text. for this i had 3 layers turned off and only two of the three encoders working on the prompt. With is on the left. without is on the right

[sacred jewel]

@craggy crest better? 😛 😛 😛

[craggy crest]

those are really cool :) very DC comics style

[sacred jewel]

It's a nice LoRA indeed. Available for SD3.5 as well 😉

[craggy crest]

zombie ironman?

[sullen moss]

Very cool!

[pseudo owl]

Sd3.5 2.6b vs Allegro 2.8b(text to video model)

[bitter hearth]

work out which step the detail you want to work on appears, then apply SLG only around that step, trying each scale from 0-10 in intervals of 1
you can start with the default blocks

its important to have it off for the final smaller sigmas

cos it will reduce fine detail, just like PAG does

you also probably want it off for the first 30% of sigmas or so, because it could reduce image diversity

[craggy crest]

i think you just made kagi's eyes crossed. can you boil that down?

i think he really would like to be able to use it

[pseudo owl]

I don’t think this is Albert Einstein lol, allegro 2.8b
prompt: Albert Einstein walking around in a futuristic world, far away

[winged seal]

Thanks <3 Training flux 8b haha

[bitter hearth]

maybe this leave layers and scale at default, start 0.3 end 0.7

keeps it away from the spicy areas

[bitter hearth]

I like minimax the most out of these proprietary video ones

[sullen moss]

Flux

[limpid thunder]

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

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

Other languages
en: help
ja: help Japanese

[kind flume]

help

#artisan-1

[pseudo owl]

Allegro is open source with Apache 2.0 license. It’s 2.8b params dit.
There’s also mochi-1 which is considerably better also Apache 2.0 and 10b dit.

In the text to video leaderboard, mochi-1 is #2 behind mini max, and beats kling, luma, gen3.

[pseudo owl]

Mochi is amazing especially for a open source model. I prefer it to mini max honestly. These ones are generated by the official website which uses the open source model and a upscaler.

[craggy crest]

sound on

[sacred geode]

@craggy crest Is this correct "frozen just means no gradients apply to it"?

[craggy crest]

not what I mean, at least. flux is basically a giant lora

[sacred geode]

I was asking in general from a convo I just had. To me I have had so many failures on Flux then I take the same dataset and throw it at SD3.5 and BANG, success after success. I mentioned that SAI said theirs were open (unfrozen) weights and was told they will say anything to stay relevant. Don't know just know I have had no real success with Flux locally while I have with 3.5. If the malformed appendages could be fixed then it will be a dream to work in for me.

[bitter hearth]

there has been a couple of pieces of news about Flux that are important regarding training
the simpletuner dev managed to train flux for over 2500 H100 GPU hours without it collapsing- this is $7,500 worth of training, more than needed for the vast majority of checkpoints
secondly the realvis dev managed to train a new checkpoint on flux de-distilled and boost the aesthetics a lot
so progress can be made its just tricky

[sacred geode]

My hope is for Juggernaught since they are working with BFL directly it gives us a simpler to train on Flux.

[craggy crest]

not sure who told you that but - flux is not only distilled, it's also DPO'd. it's frozen. 3.5 has had none of that done to it.

to do anything at all to flux, you basically have to break it and then assemble the pieces into something that's not flux

[sacred geode]

I agree, and thank you for the confirmation.

[craggy crest]

there's a reason that when asked about it, black forest said they were not interested in making it trainable. also a reason they never released any information on how to train it.

[sacred geode]

BINGO

[craggy crest]

and a very specific reason why you can train SD3.5 almost instantly, and it is such a massive fight to do anything to flux

[sacred geode]

Local it sure is for everyone I know who trains

[craggy crest]

we're being 'watched'

[sacred jewel]

wristwatched to be exact 😛

[craggy crest]

there's also a reason that all the devs who have been fighting to do anything with flux since it released dropped it and grabbed 3.5 the second they could

[craggy crest]

rock-n-rolex?

[sacred geode]

Beautiful news

[craggy crest]

trainers came out, there was a lora the DAY it released and the guy that made it didn't ahve early access, etc

it's almost effortless to work with

[sacred geode]

Flux is sledgehammer training then you finally manage it and concept bleed ruins it. SD3.5 is a sponge.

[craggy crest]

flux is an uphill battle unless you stick in it's very tight range.

it's a tool. a very good tool, with a very specific purpose. use it for that and don't try to make it be something it's not

[sacred geode]

I wish we had a save to lora node, or save to checkpoint then extract a lora from it, for 3.5.

[craggy crest]

not how it works. you have to train a lora, just like you do any other model. you can't just 'save' a file

[sacred geode]

I'm not stupid I meant you have 3.5, you have your lora. you adjust the lora. perfect. time to save it out. YES, there is a save lora node in comfy just not for 3.5

the old way was save it to checkpoint and extract

[craggy crest]

well - if you save the workflow, you'll always have your settings. and you can save templates

just set the lora up, and save it as a template

[sacred geode]

of course, but this is for release. once we get the tools we had in XL then I can rock the training for 3.5

they should come fast if the community can leave flux for 3.5

[dusky thistle]

lakers win the world series!!

[craggy crest]

there's a basketball world series?