🆕｜sd3

[sacred jewel]

Mechanical Insect LoRA

[sacred jewel]

Dreamlike surreal digital style LoRA

Surrealism style LoRA

[sacred jewel]

Soviet Robot Hunters LoRA

[fleet meteor]

Hey bro how its going? Do you have a workflow? for cogvideo img2vid? Most of the workflows I´ve found are way complicated and they use LLMs for captioning the source image (I only wanna do text to video , I can caption it earlier)

[pseudo owl]

I think these workflows are generally decent: https://github.com/kijai/ComfyUI-CogVideoXWrapper/tree/main/examples

[fleet meteor]

Thank you! I´ll try it

[sacred jewel]

BALLZ

[bitter hearth]

the facial expression on the rock LOL

[sacred jewel]

rogue giant robot LoRA

[still lance]

Create an eerie, gothic world filled with whimsical, exaggerated characters who inhabit twisted, shadowy landscapes. The scene should blend dark, muted colors with vibrant accents, capturing a sense of fantasy and isolation, while evoking both beauty and unease.

[sacred jewel]

Dali Flux LoRA

[bitter hearth]

feather

[noble coyote]

i2i Florence2/GGUF_Flux

[mental galleon]

#🏞｜general-with-images

#✍🏼｜rules-and-tos

#📣｜announcements

[bitter hearth]

amazing

[sacred jewel]

Futuristic Display LoRA

[hallow lion]

Housing market in a nutshell.

[vague topaz]

果冻蛋糕

[noble coyote]

Flux1.Dev.fp8 in PortraitMaster

[hallow lion]

👍

[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

[cunning lintel]

Schnell

[bitter hearth]

[sacred jewel]

BALLZ

[fathom wharf]

diamond painting

[outer blade]

extraterestre control society

[open vigil]

#artisan-1

[cinder junco]

Any way you can share your prompts? They don't appear to be embedded in your uploads.

[sacred jewel]

Sure.

Which image(s) ?

[cinder junco]

I like a lot of your stuff, so it would be great if you could turn on metadata somehow, lol. I don't know what tool you use.

But if you can't or don't want to, you could start with just the one I linked.

[pseudo owl]

https://pyramid-flow.github.io

New CogVideoX competitor, based on sd3 and produces pretty good videos. Only 2b and produces results similar to 5b CogVideoX.

[errant dust]

https://blogs.nvidia.com/blog/ai-decoded-flux-one/

"Flux models now support the NVIDIA TensorRT software development kit, which improves their performance up to 20%. Users can try Flux and other models with TensorRT in ComfyUI."

"The Flux models’ dev and schnell variants were downloaded more than 2 million times on HuggingFace in less than three weeks since their launch."

[bitter hearth]

yeah was a couple of weeks ago

its been a decent speed up

[noble coyote]

HOW do we get away from you?

I see your motif is a monkey - is that what we'll become?

[pseudo junco]

@fossil pagoda

[low stone]

how did you covert flux to be tensorRT? Every time I've tried it, it says OOM on my 4090.

[bitter hearth]

was a datacenter card

[noble coyote]

I wonder if TensorRT for Flux will work on my 8Gb VRAM RTX 2070?

[low stone]

oh ok. have you tried taking the resulting file and using it on a local 4090 or something?

[sacred jewel]

My Rene Magritte LoRA

[cunning lintel]

(schnell+lora)

[bitter hearth]

I haven't but that should work

[sacred jewel]

Same... but I haven't tried it on the smaller quantized models (if it even works)

[bitter hearth]

someone should just do it once and put it on hugging probably

[dusky thistle]

could you upload it to HF? 🙂

[sacred jewel]

PokeZom LoRA

[bitter hearth]

ok will have a go

[low stone]

this guy needs to release a version of it without the zombies. would be great for general stuff

[low stone]

ran the same prompt against the stability sd3 ultra api. Not sure what's up with the subject doubling.

[bitter hearth]

always a risk when an edge is 1344+

[low stone]

Yeah this is just using their 16:9 option in the api.

[bitter hearth]

never quite found out what Ultra actually is

its substantially better than SD3 Large

it has more high frequency detail which implies something like noise injection, noisy stochastic sampling or an upscale followed by a downscale

[real terrace]

So any news in the Flux world? I mostly used Schnell on Comfy, as Dev takes a lot to generate. Couldn't ever run the nf4 versions

I have a RX 6700 and Ubuntu

I'm sad that I cannot really test stuff like with SD XL because the generation time, even the loading time of the model

[bitter hearth]

not sure cos AMD

[real terrace]

Yes I read some improvement for Nvidia aparently

[bitter hearth]

yeah its good, it was 2 weeks ago though

nvidia reported late

[real terrace]

Flux Dev generation time is Interstellar time generation for me, at least usually results are great but I cannot play much with it

At least Schnell sometimes delivers amazing results, but it seems like Dev always has some twist

[bitter hearth]

some people prefer Schnell designs

[real terrace]

And I kind of lost interest in generating in SD XL as I presume I would get more interesting stuff with Flux

[real terrace]

oh great to know

[bitter hearth]

it takes a long time for tooling to get made, over a year
at the moment SD 1.5 and SDXL are strongest models

cos they have the most tooling

[real terrace]

also it seems not many SD XL new models improvements, like it has reached its limit

[real terrace]

what do you mean by tooling?

At the same time, some people make great images with whatever model, so, I don't know, but I know Flux follow prompts better and that's amazing, I'm tired of just praying in SD XL

[bitter hearth]

tooling as in software stuff

[real terrace]

Really in SD XL and with ip-adapters and ControlNEt, the posibilities are endless

[bitter hearth]

yeah at the moment there is so much to explore

[real terrace]

there are tons and tons of Flux loras, at some point I wonder if they should be integrated or something

[bitter hearth]

someone on this server had a go

they make checkpoints called "mangled merge"

they did a big one for SDXL and a newer, smaller one for flux as an experiment

[real terrace]

actually I haven't tried much as usually in my xp Lora needs a lot of tries and set up until they work as intended, if they are any good in the first place... so because long generation times, I don't want just to frustrated

[real terrace]

I was wondering if something like that would just make the model a real mess, as many Loras are actually bad

[bitter hearth]

it did, apparently

the SDXL one is cool but he said a few times the latent space gets wacky

[sacred jewel]

Indeed

[sacred jewel]

My new hobby is to just download and test flux LoRA s ... 100 per hour.... Should be done in a year 🤪🤪🤪🤪

[bitter hearth]

yeah there are so many its amazing

there was a good realism one on reddit today

turns out you can outpaint a flux image using SD 1.5 and it will continue the image
even though SD 1.5 could not create that image from scratch

[low stone]

Wow that's pretty neat. I think someone brought out a new in painter/outpainter for flux too within the last day or so

[noble coyote]

Flux_GGUF

[hallow lion]

Flux is so good it influences other models to be better.

[turbid grotto]

any news about sd3?

[noble coyote]

https://discord.com/channels/851231576494702613/1294323099155693568

[noble coyote]

Cleopatra?

[real terrace]

the original vs outpaint is noticible but pretty nice. SD 1.5 kind of lacks the lighting effects/texture or whatever, that kind of mist in the original.

You can try to do it with SD XL

I think it will generate better

just some tries, not perfect but better lighting already?

[sacred jewel]

Purz Face Projection LoRA

[noble coyote]

Purz Neon Sign LoRA

[sacred jewel]

Hey, you beat me to it... I haven't gotten to that one yet 😛

Purz Dried Flowers LoRA

[sacred jewel]

Purz Neon LoRA

[bitter hearth]

thanks that's nice

this example worked a little better
this is 100% SD1.5 though

and one from SDXL

[hallow lion]

Adequate depiction of South Korea.

[bitter hearth]

yeah the problem with SD1.5 though is
there was no mention of Korea/Korean in the prompt though LOL

[sacred jewel]

Purz VHS Box LoRA

[bitter hearth]

wow it learnt the layout really accurately

[real terrace]

pretty impressive... I don't know, I have something against 1.5. It makes pretty believable profesional photography stuff, but usually two crowded like it cannot compose the whole image, so it just add stuff...

There was a user that posted SD 1.5 and the quality was impressive but the composition was like that

[bitter hearth]

it does have a higher amount of small objects than sdxl yeah

[real terrace]

yeah like these ones

I wonder if it is some kind of upscale

I guess

[real terrace]

really incredible

[bitter hearth]

I asked him once and he said model is Epic Photonism

which is a great model

and then tiled upscale

[noble coyote]

Purz Neon Sign Lora and GGUF_Flux

[noble coyote]

[Are we at 'banal' yet?!] 🥳

[sacred jewel]

Very cool

[noble coyote]

Eaullama

[sacred jewel]

DonM Illustration Styles LoRA

[pallid summit]

Imagine/ gourmet platter of hot dogs

[hallow lion]

It tastes funny. something happened to the meat when it went through.

Yes I haven't taught craziness to AI. something is always lost.

[sacred jewel]

Comic Book Vintage LoRA

[hallow lion]

Dirk Lasermaster.

[fleet meteor]

Yeah xd, I should have tried it with flux, somehow food was better with sdxl in some cases

[steel beacon]

I have slowly been working on a full Trump deck. 12 down. Many to go.

[real terrace]

let me guess isometric something

[sacred jewel]

Actually, no... I think I prompted 3D Game Objects or something like that. It is a short prompt

[sacred jewel]

The actual prompt:

toolbox 3D game technology

[real terrace]

wow just that prompt, no "trending on artstation" or anything

[noble coyote]

Werf!

[bitter hearth]

squeak noises

[autumn arrow]

Can anyone tell what model this fake restaurant is using?

https://www.instagram.com/ethos_atx?igsh=MXNreDFsNXcydWZiaw==

[bitter hearth]

flux

[pseudo owl]

Kinda amazing I can do all of this in 1-step with schnell, nothing else.

prompts: “4-image grid”, “the word “schnell” made out of cake”, “gta5”

[bitter hearth]

1 step? wow

[pseudo owl]

Quantized too lol

[bitter hearth]

I tried 2 step and it was great

need to try 1 step now

[pseudo owl]

Yeah with 1-step, however anatomy is messed up, 2 step helps very much in anatomy.

[sacred jewel]

Flux... fer sher

[sacred jewel]

Meh, I am waiting for no step 👀

[bitter hearth]

the massive blur is hard for other models

[sacred jewel]

https://tenor.com/view/just-kidding-christina-aguilera-christina-gif-5467111

[bitter hearth]

no step sounds nice, you just get the image lol

I noticed with flux pro 1.1, the blur is even more

maybe it takes a strong model to understand blur

[zenith hemlock]

my own flux lora, fluxdev

[short thicket]

Working on 2 new versions of Mangled Merge for Flux. 'Matrix' which focuses on realistic loras and 'Magic' for 2d loras. 'Matrix is almost done. This image is with an additional 228 loras on top of the original 230 from v0, 55 more loras to go for this one and then I am going to work on Magic. I started merging 4 loras at a time in sets of 3 and then smoothing with 2 versions of the Della merge method that I created a couple weeks ago.

[bitter hearth]

Mangled Merge is a really interesting project

[short thicket]

Thank you. I'm gonna try finetuning on top of it once these two versions are complete.

[bitter hearth]

do you find it very different to base dev now?

[short thicket]

I haven't tested just yet, I want to try and get all of these loras merged first. But one thing I am finding different from v0 is that cartoons or painterly images don't work anymore. I can get plastic 3d rendered looking things but not paintings or illustrations. Of course I could be wrong. that's just my initial findings based off of 1 seed.

with realism, you lose some aesthetics too.

[bitter hearth]

hmm okay

are there benefits as well as losses?

[short thicket]

yes. I'll show you a couple instances... Mangled merge on the left vs dev on the right.

[bitter hearth]

its done a good job lowering the blur yeah

[short thicket]

MM on left Dev on right. Keep in mind it's still WIP needs some more smoothing. But the prompt was "anime,cyberpunk, A young girl with large, apprehensive eyes stands amidst a cacophony of disembodied, bloodshot eyes. Render this in a gritty, expressionistic style, emphasizing jarring color contrasts and impasto brushstrokes. Employ a worm's-eye view to enhance the feeling of being watched. The girl's face and the surrounding eyes should be the focal points, illuminated by an unseen, sickly green light source. The background recedes into an indiscernible darkness."

So anime basically goes out the window in a lot of cases.

Same here MM left Dev right. Painterly images is still doable.

seems to add signatures for painterly things. I'm not using any negatives for these.

[bitter hearth]

yeah the changes are pretty big here

[short thicket]

Yeah. I'm finding the Della merge method pretty interesting too. I coded 2 different versions, 1 that follows the paper closely and another that works better with merging models patched with loras. I don't think either are the greatest at merging loras in but I'm finding they work great for smoothing overfit models.

[spring yew]

promt: make a logo

[runic tusk]

No.

[noble coyote]

JustEat_v1.safetensors

[noble coyote]

Flux with PortraitMaster

[noble coyote]

Cool!

[upbeat lynx]

Can you make a hairstyle changer?

[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

[steel beacon]

Anybody know if there's a way to have eye occlusion on Reactor?

If my model has glowing red eyes, it'd be nice to have face swapping with those glowing eyes

[noble coyote]

PortraitMaster+Flux

Prompt: "An Art Nouveau style storybook page depicting the Whispering Woods scene. The page features elegant, flowing borders with organic floral and vine patterns. At the top, the title 'The Whispering Woods' is written in a beautifully ornate script, surrounded by decorative elements. The scene shows the shadowy forest with graceful trees, fireflies lighting a winding path, and curling roots, all framed by the decorative border. The colors are soft and magical, with deep greens, silvery moonlight, and glowing gold from the fireflies. The layout feels like a page from an enchanted storybook, with a mixture of imagery and text space."

[languid prism]

if you are a branding specialist, please imagine a image branding picture for Midea Industry Group, the picture is about good lifes in the city, should contains the elements of home appliance, air conditioners, new energy cars, e-bike, energy storage products, people are enjoying happy life created by Midea, the picture should reflect the company value of nature and sustainability

[runic tusk]

No.

[winter tusk]

Generate high resolution render output from this

[muted dove]

Teachers suspected of 'fleecing' students of their buzz

[proper umbra]

I have been trying for days to achieve the same results as XY on my local installation. However, I am failing.

I know that they use Stable Diffusion XL but nothing else. I have a sample prompt here.

"Create a cartoonish illustration of a muscular, anthropomorphic rat. The rat should have exaggerated, bulging muscles and a confident expression, resembling a bodybuilder. It should be holding a red dumbbell, showcasing strength. The rat should have large ears and a cartoonish face with a smirk. Incorporate a bright and bold color scheme with black as the background to make the character stand out. At the bottom of the image, add the text "GYM RAT" in a large, bold, red font that looks hand-draw"

Attached is the comparison.

Can anyone help me how to get a similar result locally? I have tried it several times and the results were always better with stablediffusionweb.

my settings

[pseudo owl]

https://nvlabs.github.io/Sana/

yeah, I think stability is kinda cooked if above is good as shown.

Sana is 0.6b parameters, competitive to flux dev, undistilled, supports 4096 res natively. 1 sec for a single img on a decent gpu.

[noble coyote]

Looks interesting!

[rapid pivot]

@craggy crest look dm

[dull star]

uses Gemma2-2B instead of T5-XXL

I would personally not put sana 1.6B so high up on the GenEval score, but for how small it is it really makes up for it.

if its really good to train for, PurpleSmartAI/Astraliteheart might love this

Idk how far they are into training for AuraFlow

[pseudo owl]

Yeah I think flux dev is still the best but its amazing for it's size, undistilled, 4k native generation, and incredibly fast.

[low stone]

rather excited for it. pixart was great.

[cunning lintel]

those scores get weirder and weirder, how is pixartsigma slightly below sdxl and way above lumina-next while pixart-sigma and lumina are in the same ballpark and definitely progress compared to sdxl. And then we have flux-schnell above dev. Meanwhile SD3, which just falls apart on 60% of my promps is above pixar and lumina which manage sane results on most prompts

if this sana model handles a variety of styles similarly to as pixart, it could be fun to use, the demo images have this very noisy grainy look though, time will tell.

[bitter hearth]

I agree with their scores, for what it's worth. Lumina can get pretty mushy, and Sigma feels a bit below SDXL to me

SD3 is only bad if there is a person in the prompt

[cunning lintel]

Guess we value different things :p

SD3m outputs are just awful to my eye, boring full frontal photo's of whatever i prompt with no atmosphere at all

[bitter hearth]

SD3m is overtrained on photos a bit

which would artificially boost its benchmark scores a bit

this is the sort of image benchmarks are testing for: https://www.researchgate.net/publication/328362220/figure/fig4/AS:1086455329890335@1636042544303/Samples-of-LSUN-bedroom-dataset.jpg

just like generic photos of a room

so models that are overtrained on that style do really well in benchmarks even if they can't do other styles

[cunning lintel]

yeah, and geneval seems automated object detection

says nothing abput coherence, sadly also noticable in those sana images, that word "fast" on the cat's sign, the letter T is halfway out of the sign 😬

[bitter hearth]

I took a longer look through the paper

its gonna be good I think

competitive scores on all of ImageReward, FID on MJHQ-30K, DPG-Bench and GenEval

and its 100 times faster than Flux Dev

[fleet meteor]

Do anyone know how to run flux with vae and clip l included (I have to run t5 separated) in comfyui?

What I mean is , is there a way to load only the T5 model?

[mortal mesa]

there is a load clip, not the clip vision one

[bitter hearth]

can't with default comfy nodes

if you only have a full checkpoint file

[icy drift]

I would say these are unusably bad, but at those speeds and resolutions... Hmm.

[icy drift]

Just ignore the included Clip-L and use the DualCLIPLoader node. (You shouldn't be prompting clip-l anyway. Use the flux prompt node and leave clip-l blank or else you'll seriously hurt quality. Do some tests with a frozen seed if you don't believe me.)

[bitter hearth]

depends so much on the seed and prompt whether its good to have Clip-l

to a good extent I agree though

[noble coyote]

The Weight Family are nearing the end of their incarceration!

[noble coyote]

Clip-l adds very fine detail imho

[icy drift]

Well, that could be true. My tests used multiple sequential seeds and just showed that overall Clip-L hurt performance. On a seed-by-seed basis, maybe you would find exceptions.

[icy drift]

Opinion? Did you test this? (I did not.)

[noble coyote]

Yes, tested way back when Flux was released

[icy drift]

Will test.

[noble coyote]

The detail added was very very fine

[icy drift]

Not sure what that means. What am I looking for here? (The problem with prompting Clip-L is you get concept deformations. E.g. the teeth of a mimic chest bleed out onto the floor around it.)

[bitter hearth]

for the most part I just leave stuff as default for prompting (so I use both Clip-L and T5)
and then I feed Florence 2 node to it
I'm not much of a prompter

[noble coyote]

Mousy-cide

[icy drift]

My prompts are very specific, and I roll anywhere from dozens of times to hundreds to get one image that actually follows my prompt. Different strategies I guess, but you should give it a try and see if you prefer the results.

[icy drift]

Is this for Sana?

[noble coyote]

SD3 Medium

[icy drift]

Isn't SD3 worse than Flux though? Did I miss a new API version?

[noble coyote]

If you plan and plan your prompts, you can get workable images with SD3 - but it mangles human anatomy. But with Flux, it's so much easier

[bitter hearth]

yeah my prompts are super basic, stuff like this: (Photo:1.3) of a street in a city. There are taxis and lamp posts. There are bins and plants. There is a garbage can and a drain.

[icy drift]

Yeah you're not even specifying composition. I don't know if clip-l would hurt in that case.

[icy drift]

It's like asking an out-of-shape guy to take off his shirt though... Come on dude. 😨

[bitter hearth]

I looked at your tests before

did you test T5 with full text, with tags for clip?

I feel like your tags for clip might have been too long for clip to handle maybe

[icy drift]

That might be a thing too. I saw one of my favorite youtubers using just a few short tags for clip. I'm testing now to see if using the same prompt for t5 and clip adds detail, but I'm not sure that's what @noble coyote meant.

[bitter hearth]

I tried as well some of the new fancy clip or clip long fine tunes
but I never got better results from them

[icy drift]

T-5 only null-test. Prompt:
In this RAW photo, A furry anthropomorphic hamster live-action anime girl is holding a sign above her head. The sign says "CLIP-L FINE DETAIL". She is wearing a denim jacket and sneakers. She is standing in contapposto. Her silky, glossy brown and cream-colored fur is highly detailed and shining in the warm sunlight. She is smiling widely, with sparkling eyes. Her denim jacket has a rich fabric texture. Her shoe laces are untied. She has long, wavy, glossy brown hair flowing down her shoulders. She is standing in the park on a field of grass and wildflowers. Fluffy white clouds drift through the blue sky overhead. The photo is taken on an antique polaroid camera and is extremely highly detailed.

Doing clip now with same prompt, then will try with a few short tags, and just check detail changes.
(Same seeds / settings etc. Only varying prompt for test.)
(Also I'm using one of those fancy Clip-L versions.)

No significant detail gain from just pasting same prompt into clip.
Trying short tags next.

clip tags: RAW photo, extremely highly detailed fur, hair, fabric texture, eyes, grass, wildflowers, clouds, film grain

Very definite and obvious detail gain!!! 🥳

That's fantastic.

Wonder if I have time to test a difficult prompt and a portrait...

[bitter hearth]

wow nice

yeah this matches my experience

Clip-L with maybe 6-10 tags is nice

on models with Clip-G, sometimes Clip-G can be good with just 3-4 tags

[icy drift]

These textures are really something else. I love it. Out of time though. Gotta go.

[bitter hearth]

okay bye, thanks for tests

[icy drift]

Also these were all with the 8-step hyper lora.

[bitter hearth]

in my testing this one I put before maxed out T5 ```(Photo:1.3) of a street in a city. There are taxis and lamp posts. There are bins and plants. There is a garbage can and a drain.

unless you need very specific things

[icy drift]

Always. 🙂

[bitter hearth]

haha yeah

some people get really good results from prompting

[sterile pendant]

Depends on the model architecture and how well it was trained. Speaking of models, I saw that that new tiny Nvidia model uses Gemma instead of t5. Guess they found a way to make it work and apparently, it does a much better job with understanding. But that makes sense, since t5 is almost archaic by ML standards now lol

[bitter hearth]

yeah that is true different diffusion models will make better or worse use of T5
and yeah the usage of Gemma in the new Nvidia model it exciting

its the Pixart team apparently

[wary portal]

SD3 still hasn't released better weights than SD3 medium?

[muted dove]

No

They've gone silent.

[wary portal]

Well with that upcoming Sana, they are probably panicking.

[muted dove]

Not heard of that one

[cunning lintel]

Nah, they're laughing, sana looks less then stellar :/ https://nvlabs.github.io/Sana/

The more i look at those images the worse they become, if all you need is speed, maybe people will see value as some gimmick to run a model on phone or something, but it looks not so great to me

[muted dove]

Well, the SD3 release wasn't anything to celebrate either 🤷🏻‍♂️
I'd like to know if Stability still exists as a company and if they intend to continue releasing models. It's been complete silence since Flux released.

[cunning lintel]

I'm pretty sure there's still the intend to release something as it was teased in their fine-tuning guide which is kinda sort of official communication and maybe more official than lykon's hints on twitter, but how well SAI's new thing works is anyone's guess

I really had high hopes for this new pixart / sana thing 😥 but it's just "fast"

[muted dove]

It's too small to be any good...0.6B

[noble coyote]

Isn't there a 1.6B version too?

[wary portal]

yeah flux has been amazing so far.

[wary portal]

You are right, example images like this looks pretty bad.

[noble coyote]

^..^<

[zenith hemlock]

but its insanely fast

[pseudo owl]

Its much much faster then sdxl, sd3. Pixart, and 105x faster than dev at 4k.

It can natively generate 4k images natively with a speed 105x faster than flux, it’s undistilled too. It will use way less vram then flux dev.

Dev does seem clearly better but you can gen 10+ images way faster then a single flux gen and probably get a better img then dev.

I have to say though, humans might be an issue tho, they don’t show them in many poses.

[pseudo owl]

Yeah sd3.5 was supposed to be available for some people for testing but idk what happened to it now. 8b was decent, worse then flux in most things but non distilled and more pleasant aesthetics by default.

No news about 8b or 2b now from what I saw.

[dry wave]

who cares how fast it is when it cannot make any consistent image? The generations are all messed up

[pseudo owl]

They aren’t messed up from what I see. There are some artifacts but even flux has them. All diffusion models have them(like the text being weird in the bottom left of the cat img).

It is lower quality then flux dev but as I said for single gen but way faster. It won’t replace it or anything but it will be a solid alternative.

[dry wave]

all images I saw so far were full of errors. Much worse than sdxl

[young leaf]

help

[dusky thistle]

wonder if SAI is going to actually release 8B or just wait for it to become obsolete :/

[pseudo owl]

Yeah idk abt that, did you check the page? Sdxl has an incredibly hard time writing a single word, and has very bad prompt following compared to the newer models and this one.

[young blade]

has anyone heard anything new from Flux team aside from their blackforest twitter account?

[hallow lion]

Help me Clownshark. Help me make real movies.

[dusky thistle]

these are the settings

[rapid pivot]

no dude, trust

SDXL is a god at generating images all we hear from it is true

no base model compares to "sdxl" lmao

(meme)

first time I see sana stuff

looks very fast, interesting stuff

[bitter hearth]

if Sana really is 100 times faster than dev

then we could do tiled upscale up to 8k

and then downscale to 1024

in the time it takes dev to do the image normally at 1024

[stoic turtle]

so are we still waiting for sd3 something or is that whatever was delayed here now

checking in after a few months

[rapid pivot]

We are waiting for sana now

[stoic turtle]

the weights i think they were

those here yet

[mortal mesa]

nothing new yet, nothing known, please hold restructuring

[mental bison]

Any news about SD3.1?

[radiant ledge]

two weeks

[mental bison]

[dry wave]

I don't really know what you wanna say with that. Sana is fast because it is not a transformer method.

They use linear attention, which is faster than normal attention, but also leads to inferior results. To compensate for that they add convolutional nets.

yes its fast, but it suffers from all disadvantages of this architecture

it is also a bit funny because they act like "we have a DiT which is super fast and does not need position embeds - nobody achieved that yet!"

although SDXL is using exactly such an architecture: convolutional backbone, no position embeds

SDXL is using real attention however

and that's what you see in the results. Its superior to Sana

If you want super fast results use Paella. its a convolution only architecture

its bad as Sana but its super fast

[dry wave]

cause its using CLIP. I think we all learned now that you need a proper LLM as CLIP has no good text understanding

[bitter hearth]

it might be possible to bootstrap T5 to it like Ella anyway

I sort of don't like judging image models by their text generation ability

I know some people need that, but it feels like a more niche usecase

[pseudo owl]

Did you check the images in the project page or arvix? Sana base beats paella by a very large margin and even sdxl in a few cases while being much smaller.

Maybe you are comparing it to fine tuned sdxl models, base sdxl is not very great.

[pseudo owl]

Yeah true, even sdxl can get decent text rendering with enough training I believe. I think to judge a model, it’s more of a mix of prompt following, image quality, humans, and text rendering.

[dry wave]

lol, whats the difference of fine tuned sdxl models to sdxl base?

its same model architecture

[pseudo owl]

No I’m not talking about architecture, I’m talking abt image quality.

[dry wave]

yes, if I make a new model and train it on better data I can beat other models that are trained on bad data

I know

and I say: their architecture is probably not good

that's all i care about 🤷 They won't beat any other good image model with that. i would rather wait for a next PixArt then

[gusty trail]

sana is the next pixart.

[dry wave]

there is only one author shared between both?

anyways, seems a step in the wrong direction to me

[bitter hearth]

I would have much rather had the exact opposite yeah
a model that is 100x slower than flux but with better image quality

[dry wave]

I would be also fine with a Flux light

what SD3 is supposed to be when it ever comes out

or some new and cool architecture like MAR

[bitter hearth]

the alleged samples of SD 3.5 2B looked great

[dry wave]

I don't trust anything SAI post anywhere ^^° I will just wait until its released

[bitter hearth]

yeah its hard to know

Hunyuan-DiT is slowly improving as well, they might yield a good model

the latest Hunyuan-DiT is not even that bad, so long as you do a tiled upscale or progressive upscale

it does need multiple passes

[pseudo owl]

I don’t see anywhere that it’s worse then normal dit, it just says that they replace normal attention with linear attention which loses some quality but then they replace ffn with their mix-ffn which makes it regain the quality.

[dry wave]

yeah, thats already ridiculous to me. Replacing attention with linear attention makes quality worse. Thats a good sign that linear attention is just not good for this task

using convolutional resnets can compensate that. Yes. We know. You can replace the complete network with conv resnets (see Paella). You will just not reach the quality of DiT

you see the problem in all their example images. They are full of errors and lack any global coherence

also their choice of making the VAE compression rate larger is a mistake in my opinion

generating images from latent using a l2 loss makes images blurry and low quality. To get around that you need either a GAN or a diffusion model. I think VAEs are usually trained with a GAN-like loss to mitgate this issue. But that has its own downside. You just get into trouble if you make the compression ratio too large

[bitter hearth]

linear attention seems to be better for very small things, smaller than the size needed for 512x512+ diffusion models

the scaling trends for linear attention seem to be okay for small compute amounts but it then falls off a cliff

which makes it useful for certain things but 512x512+ diffusion models are maybe too demanding

[pseudo owl]

Idk seems pretty coherent in the example images, not perfect but when are diffusion models perfect?

The vae part is true too imo, supposed to be comparable to sdxl vae which honestly is not too great compared to the 16ch vae. Can increase training speed and inference speed but still.

[bitter hearth]

I found that SDXL and SD 1.5 VAE still does well if your final image is 4k+ resolution
i.e. tiled upscale

then the VAE limitations are minimised

at 1024x1024 scale, the SD3/Flux VAEs are much better yeah

[gusty trail]

Their ae paper https://arxiv.org/abs/2410.10733

[bitter hearth]

if the rest of the model was good I would feel okay with SDXL VAE

not ideal but would be okay

a bigger problem with SDXL was the lack of ZTSNR, which pretty much every new model fixes

[turbid grotto]

[bitter hearth]

its probably dropping rly soon

I think people will be happy with it when it drops (despite lots of memeing)

the previews looked good

[turbid grotto]

I wouldn't mind if they trained for a bit longer if there is still benefit

[turbid grotto]

previews?

[bitter hearth]

they were on this server so they should come up in the search

[cunning lintel]

user revealedinadream_70414 claimed to have sd3.5 images, so search that username

it was a month ago #🆕｜sd3 message maybe #🍥｜anime message and #🆕｜sd3 message it seems, and a few more they posted were 3.5?? possibly

[turbid grotto]

thanks! Sadly, there is no complex examples, but it seems model was undertrained a month ago, so it will probably take some more time

[rapid pivot]

Those things are so weird to eat

[sage burrow]

Is there a new sd3 yet?

[turbid grotto]

8 months ago...
Which model made this? It is Flux level hands

Also, what happened to model from the paper, did it just disappeared or what? It looked awesome

[bitter hearth]

the 8B

[turbid grotto]

The one in api really that great? If so, why it took another 8 months and yet not ready?

[bitter hearth]

Ultra pipeline in the API in particular is very good yeah

[lucid swift]

they dont give it out because they sturggle with moeny

[dusky thistle]

yea that's basically like 8 years ago in AI years

seems like the plan is to just let 8B become obsolete, then release it

it was obviously ready to release 6+ months ago

[fleet meteor]

8B seems kinda small today that we got flux 😁 @dusky thistle

But tbh it looks like the smaller the model the more efficient somehow, look at pixart for example, or sd3 (which was badly trained but for its size it was good in some aspects)

[dusky thistle]

8b is big enough, i think

what flux has really done for us is put to rest the idea that we need models to be tiny for them to be adopted by the masses

no one cares about how small your model is

they can just quantize it

one of the arguments against releasing great/big models was also that we wouldn't be able to train loras

well, we can drop precision there too and still get good results, we can swap blocks and finetune 12b on a freakin 4090 (which i'm doing, it works)

[sacred jewel]

https://civitai.com/models/864650?modelVersionId=967520

[river vine]

Please generate a 1900*1200 sized wallpaper showing an artis representation of a neuronal network.

[muted dove]

https://tenor.com/view/person-of-interest-poi-the-machine-neuron-activation-neural-network-gif-23102996

[noble coyote]

Flux_GGUF and Magritte LoRA

[civic trail]

Come and get your Balloon...

[wary scroll]

m

[bitter hearth]

I like how there is always a camera lense hidden somewhere now

its like an easter egg

[wet rose]

Solution = “looking at the camera” or “looking at the viewer” and Negative Prompt = “camera”

[short thicket]

Mangled Merge Matrix is complete. Magic is coming along.

[alpine summit]

Flux

Flux+Minimax

[noble coyote]

DecadeTW Auto Prompt

[noble coyote]

"Super-Flux_GGUF"

From an idea by Olivio Sarkas

[short thicket]

I saw that video last night but haven't gotten around to adopting it. Do you see many improvements?

[noble coyote]

Really clear and sharp. Its like the Refiner stage after the Base in SDXL

[short thicket]

Cool. It looked pretty solid in the video. Do you see speed improvements?

[noble coyote]

Using GGUF_Q8 is always much speedier than Flux.Dev

[short thicket]

Unfortunately I'm stuck with Dev at least until I get all these loras merged. 143 to go. Then I get to play around after quantizing.