🆕｜sd3

[sacred jewel]

Supposeed to be a Faberge Egg... this is NOT a Faberge egg... 😛

Thanks, Flux. lol

[bitter hearth]

lol

interesting that it kinda knew there were gems

[sage burrow]

Flux shops at the dollar store sometimes 😉

[icy drift]

Doesn't matter that there's two spines on a book near bottom of center stack. No one was going to glance twice at the stock image anyway.

[bitter hearth]

even after you pointed it out I couldn't spot it at first 👍

[errant dust]

And it is why I used it in the end anyhow for my video's thumbnail

[icy drift]

In-focus background.

[sullen moss]

https://videocardz.com/newz/nvidia-geforce-rtx-5090-and-rtx-5080-specs-leaked?fbclid=IwY2xjawFipCZleHRuA2FlbQIxMQABHSNt7WGsimJGegM3-h6_Rk0aHb8axUgwKP7aDKiBVzrsnjdWGQtEJ6Vz0A_aem_mX4Rgvc_z89nYbVdDlZH6

[sacred jewel]

I am leaking the price:

$5090.99 USD

🤷‍♂️

The most I want is the extra VRAM... all else can stay the same as 4090 😛

[runic tusk]

Giving off these vibes:

[sacred jewel]

euler vs uni_pc_bh2

[grim crest]

its giving me a little of this vibe

[runic tusk]

https://tenor.com/xLnU.gif

[noble coyote]

GGUF_Q8 Flux

[muted dove]

From flux... 🤷🏻‍♂️

a close up photo of an intricately detailed Fabergé egg

[noble coyote]

Q8_GGUF Flux

[hallow lion]

Ah I just started with that one...

With controlnet on even the shnell flux GGUF will teach you patience.

Less than a year later I need another upgrade...

I came into all this at the tale end of SD15... SDXL was taking over so I upgraded and then SDXL ran pretty fast. Happy days.

then Came Cascade and Hunyuan and SD3 and now FLux

now i cry once more

[ebon geode]

preach 🙏 😔

cant even get ts to run rn

[hallow lion]

When will it end?

Never. It will never end.

[noble coyote]

Gone to LiDL ...

[sacred jewel]

Awesome 👌

[white dawn]

cat

[icy drift]

I'll believe it when I see it. 32GB VRAM would be a 100% legit reason to upgrade, so there's no way they'll give us that. They have zero competition, 70% profit margins, and more pools of AI money than anyone could swim in.

If accurate, my best Flux workflow would run 25% faster. 😕 (But we don't know the clock speed, so this might be a little off.)
It could be faster switching out models, if you had a CPU and storage drive that could saturate a 512b DDR7 pipe, which you don't.

Actually, if the 32GB VRAM let's you handle a batchsize >1 entirely on the GPU, this could be multiples faster. As much as 2-3x faster.

If I try a 4-batch right now, I get 141 seconds, or 35 seconds per image vs. 44 seconds for a batch of 1. A 20% speedup, if you don't mind being unable to cancel from the preview. (These are 4MP images though, 1664x2432.)

[hallow lion]

When AI models look more natural than real women.

[icy drift]

Protruding cheekbones are definitely a minority female skull distribution though... That's some supermodel level sigma there.

[noble coyote]

"From 'king-of-the psychotic' to master of the banal!!!" 😄

[dusky thistle]

just you wait... once i get the photography shit really down...

then it will get real interesting lol

[noble coyote]

🥳

[dusky thistle]

having a great time with coding up SDEs here

i remember hearing they wouldn't work with rectified flow... idk about that, they seem to be better than ODEs for RF just like with sdxl etc

[hallow lion]

These images say that your prompts are very... you must have a very deep a soul level connection to your creations.

[hallow lion]

Remember, we need to be able to summon up each and every part of our image. Like a set dresser on a movie set. to become more thean a gimmick/fad we need total control.

I mean down to the last mole on her skin and the last prop on the set behind her.

As film directors we need total and absolute control.

Then Hollywood will tremble.

Movies have people who worry about continuity.

[dusky thistle]

james cameron will come begging us for a job

[hallow lion]

I believe he will

in 3-5 years

https://tenor.com/view/star-wars-red-leader-garven-dreis-gif-4770703215780882083

[dusky thistle]

hah

[hallow lion]

But seriously

I dont mind promting up each and every aspect of my scene

but once I am happy with it it has to be consistently reproduceable.

I mean every aspect down to the last piece of furniture. Last mole

100%

No video gen is there yet

INCLUDING SORA.

[hallow lion]

I imagine being able to lock down a thing like "object 8" in the mind of the AI - like how we can now specify "face" for DINO ad segment anything.

And somehow keep "object 8" as "object 8" in future generations

that could be anything

like the main characters blouse.

We need to be able to lock down elements an reuse them

locally also.

It has to be locally otherwise it;s useless in ym opinion.

And when we can do that

Hollywood will die

Until we can reuse elements with 100% consitency they laugh at us

|But once we do

they will tremble

Imagine: use dino to select her skirt. and name it "skirt 2" and from then on the AI will use "skirt 2" as it is with 100% consistency.

And we can use this for any object in the scene.

In 3D software even tho it looks like ass and it is limited we can use the same objects over and over again with relative consitency...

We need the same here

We solved the lighting with IC-Lighting.

But not yet the shape.

[noble coyote]

"I sora video once; didn't see it again!!!"

Florence2/Flux i2i

[pseudo junco]

can you share the prompt bro

[dusky thistle]

At the edge of a frozen lake, a woman stands quietly, watching as the first light of dawn casts a soft pink glow across the snow-covered landscape. Her dark coat contrasts against the white snow, and her breath forms small clouds in the cold morning air. Her hazel eyes reflect the pastel colors of the sky as she gazes out over the serene, icy waters. The surrounding trees are dusted with snow, and the air is crisp and still, with only the faint sound of ice creaking beneath the surface of the lake. Her dark hair is tucked into a warm scarf, and a peaceful smile plays across her lips as she takes in the quiet beauty of the winter morning. The frozen lake and the soft dawn light create a tranquil, almost otherworldly atmosphere.

[pseudo junco]

thank you so much

[dusky thistle]

be sure to share what you get with it 😄

[pseudo junco]

Not as good as yours, but here it is.

[noble coyote]

Florence2/Flux i2i

[pseudo junco]

whats your sampling method bro

[muted dove]

2 passes of flux, adding a little latent noise and x1.5 latent upscale on 2nd pass

[pseudo junco]

thx

[noble coyote]

These are SD1.5 - IC-Light.
Can IC-Light be used on anything but SD1.5?

[muted dove]

No

[noble coyote]

Aarrrghhh!!¬!!

Back to Photoshop 😄

[muted dove]

You can use it for initial image to guide another model.

[noble coyote]

img2img it into Flux or Ollama? OK, cool

[muted dove]

Not sure if the light will stay in the same place though 🤷🏻‍♂️

[noble coyote]

OK, just hoping somebody is working on a Flux repo using IC-Light?! 🥳

[cinder vault]

a girl

[noble coyote]

A hedgehog's eyebrow

[errant dust]

Euler is by far the more realistic of the two images

[noble coyote]

A vintage Polaroid camera with a picture emerging from it. The photograph distorts reality, showing trees, buildings, and faces that twist into impossible shapes. Skyscrapers spiral upwards, melting into the clouds, while the horizon folds in on itself. A human figure stretches like elastic, caught between dimensions. The once-clear landscape morphs into a kaleidoscope of colors and surreal shapes, defying logic. The image emerging from the Polaroid frame contains this surreal world where reality and the impossible blur together.

[noble coyote]

Q8_GGUF Flux + LoRAs

[floral gate]

Imagen 3 >>> SD3 + Flux
(just not in censorship)

[bitter hearth]

nice lighting

[floral gate]

I know right

I made one photo by accident by typing in some random prompt I just meant to copy and paste somewhere else, and the result was proably the best ai image I had ever seen. I swear it looked straight out of an instagram post or something

@bitter hearth I found it, I searched up imagen 3 in the discord, I'm so glad I saved it here before deleting it

Like if someone showed me that, I would never believe it's AI

[bitter hearth]

yeah that one is like a photo

[pseudo owl]

yeah I believe its better in styles but from what I tested, flux seems similar in prompt following and better at text rendering. also imagen3 seems really censored but maybe they improved the safety system now.

and again not really fair to compare it to flux dev as imagen3 most likely uses a bunch of extra techniques and prompt enhancing to improve quality.

[floral gate]

That is true. I do like flux better for prompt adherence and also just being able to put like anything in, but other than that, I do think if you just compare model to model based on how you're able to use them (if that makes sense?), then I do think imagen 3 is better quality-wise

But like as you stated, everything else is kind of better.

[pseudo owl]

Yeah the only thing I dislike about raw flux dev is that its not very "creative" i suppose. You have to enhance the prompt using an llm, and the quality is usually much greater.

[bitter hearth]

classic side effect of distilling yeah
SDXL lightning etc have the same issue

[pseudo owl]

Yeah true, but just using an llm to enhance your prompt mostly solves that.

[bitter hearth]

yeah llms write way better prompts than I do

I think stuff like outpainting might be good for creativity

or cutting out characters and pasting on the background before refine

[fleet meteor]

What model is this? 👀

[bitter hearth]

its flux, cos of this

[daring spire]

Намалюй Логотип Pavlo Ruban School Діагностика ходової частини та Встановлення кутів коліс

[sacred jewel]

Flux.1 Dev

[noble coyote]

Florence2/Flux1.Dev img2img

Anybody understand how to use the Text Input box in Florence2?! 😄

[hallow lion]

nice

[noble coyote]

https://www.youtube.com/watch?v=syXk5CgABJg

[hollow yarrow]

cat

[rain current]

dog

[hollow yarrow]

draw a cat

[noble coyote]

Custard pie

[hollow yarrow]

how to generate a picture?

[noble coyote]

Go to artisan

[noble coyote]

#1237460438229450772 message

[hollow yarrow]

Thanks you!

[sacred jewel]

Plastic Fantastic

Trigger warning... (get it? lololol)

[sterile pendant]

Wait is that Finn from adventure time in the style of DBZ? Lol....

[urban arch]

Not sure if anyone mentioned this yet, but James Cameron is working with Stability AI now.

[sacred jewel]

Mentioned already...

[urban arch]

Ok, I didn't make it into the discord yesterday, so I missed everything.

[bitter hearth]

||badumtss||

[fleet meteor]

sorry i tried in flux schnell and it was way too simple

[fleet meteor]

Amazing, i wish i could get gens like that

[fleet meteor]

😁

For logos you have to be way more detailed in the prompt

[noble coyote]

A photorealistic image of a lighthouse standing on a cliff by the sea, with its beacon shining through turbulent weather. The light pierces through stormy clouds, symbolizing hope and guidance. The image captures the contrast between the dark sky and the bright, unwavering lighthouse light, representing direction in times of uncertainty.

[bitter hearth]

I like the bikes

[dull star]

https://civitai.com/models/800229/friendship-ended-with-sd-now-flux-is-my-best-friend

[cedar raven]

#artisan-1 Generate an acidic minimalist poster with rich colors

[hallow lion]

a scifi movie with these sort of vehicles would be epic

[sacred jewel]

Yup... agreed!

[sacred jewel]

You have to be in that channel

[sullen moss]

I can't seem to solve the problem of blurry inpainting in FLUX. No matter how hard I try, I can't achieve clear detailing, especially on the skin. SDXL handles this task perfectly. Has anyone found a way to get rid of this 'perfect smoothness' in FLUX?

Flux

SDXL

[noble coyote]

Try FreeU2

[noble coyote]

If you used the word "background" in your prompt then perhaps remove it. Reports of token problems surrounding the word background

[sullen moss]

hmmm

Thanks

[bitter hearth]

what am I looking at

[hazy kestrel]

😁

[hot isle]

be not afraid

[steel beacon]

Fans of Chronicles of Amber by Roger Zelazny, here's a Trump of Eric of Amber that you can regard if you want to try to contact him. 😄

[urban arch]

Wow, it's been 40 years since I read those books. You KNOW it going to be good when one of the main characters is named Random.

[sacred jewel]

Getting around to upgrading Forge since I left it a long while back...

[sacred jewel]

Do you really want to know? You will NOT believe the img2img source... 🤭

the prompt was nothing to do with the source image. I was just playing around with different input latent shapes and types...

[steel beacon]

I'm reading it again for the I-don't-know-how-many-th time as I look into playing the RPG with a new group. Steven Colbert got me to read it this last time after he announced last year that he's working on a TV adaptation of the books.

[noble coyote]

Florence2/GGUF_Flux img2img

[steel beacon]

Hey there, you know I like your stuff. Could you do some images on tarot cards? If you're interested in that, I'd like to see it in your own style but similar to what I have above: #🆕｜sd3 message

[noble coyote]

I'll give it a shot 🙂

[sacred jewel]

Architectural Sketching LoRA

[bitter hearth]

I'd like to play around with this type of shape

👍

[noble coyote]

^..^<

[sacred jewel]

Play Around, yes... I would love to "play" around wiith [that] model 😛

[bitter hearth]

Sadly...

There are no balls

[mortal kite]

"Pool Party"

[urban arch]

GODS I hope not! He is as WOKE as they get.

[steel beacon]

That's what I read, but it's been a year and a half and there's no new updates. I'd be happy if Neil Gaiman did it, as he was a friend of Roger Zelazny, or even George R.R. Martin, or even Jim Butcher, who are fans of Zelazny. I just don't want the books to continue in obscurity.

[sacred jewel]

https://tenor.com/view/right-yeah-right-yeah-oh-yeah-ohh-gif-14553729619893238098

[sacred jewel]

Dystopian LoRA

[sacred jewel]

Flux Citizen LoRA - Drake

[sacred jewel]

Flux Citizen LoRA- Aegis

[misty nimbus]

a cute girl

#artisan-1 cute

[atomic abyss]

a cute dog

a cute dog | sd3

Flux Citizen LoRA- Aegis

[noble coyote]

£10 cash

[muted dove]

DM me the next winning lottery numbers

[frail shoal]

There is a new model with a 16ch vae and 3B parameters : https://github.com/THUDM/CogView3?tab=readme-ov-file.

[noble coyote]

fLORENCE2/gguf_fLUX IMG2IMG

[frail shoal]

not convinced by the preview pictures, but waiting to test it in comfyui after it gets supported to judge myself

[sullen moss]

I hope that someday we will see the same quality in open source as we do in Kling

[dapper peak]

help

[noble coyote]

I need somebody ...

[sacred jewel]

not just anybody

[urban arch]

Give it a few more years for the WOKE Mind Virus to completely die off, THEN maybe we'll get a good adaptation.

[frail shoal]

it's text to image, not a video model i think

[noble coyote]

"Minimalist rugged oil painting in faded earthy green hues, capturing delicate details in vast solid patches. A post-apocalyptic city lies in smoking ruins, beginning to be overtaken by nature. Vines crawl up the sides of half-crumbled skyscrapers and leaves form in mats in the streets. Intense bluish mist slightly obscure the setting sun in warm salmon tones. Low angle cityscape from the ground on the other side of the river near the city."

[bitter hearth]

yeah it looks perfect

[bitter hearth]

its got everything
no distillation, 16 channel VAE and DIT

it doesn't have rectified flow however

actually I guess that makes it easier for us if its not ret flow

[young blade]

so does it feel like being able to NEG prompt in Flux isn't gonna happen for a bit? anyone having luck at all being able to neg something consistently in some way?

[pseudo owl]

Yeah I tested it using the API, and its actually pretty good. The hands have some issues but its actually nice. Probably what I expect sd3.1 2b to be like.

[mystic isle]

what scale is this, and what model?

trying to test somethinn

[mystic isle]

so cool!!! :0000 model????

[bitter hearth]

I use negatives in flux every time now
just use Skimmed CFG and Latent Mega Modifier nodes to reduce the CFG burn and its fine

[mystic isle]

having an issue where alot of my gens start to degrade or some out half baked...?

[sacred jewel]

Flux.1 Dev fp16

[young blade]

hmm, and you actually do notice the negs not appearing at least somewhat consistently?

[bitter hearth]

oh I just leave the negative prompt box empty

if you want to make negatives more powerful then you might find that delaying the negative a bit helps

so have empty negative at early steps and then the negative comes in

if that is not enough then perpneg helps loads, but its 50% slower

[sacred jewel]

Weird World LoRA

[sacred jewel]

Techno Punk LoRA

[spice ginkgo]

human

[noble coyote]

ectomorph

[dusky thistle]

i neg prompt all the time

[pseudo granite]

if cfg is 1 it will disable neg prompt

[noble coyote]

Florence2/GGUF_Flux img2img

[dull star]

I thought this was a plant until I realised that this is #🆕｜sd3

holy crap this looks photoreal for some reason

[civic trail]

Scale? 1088x960

[young blade]

love to see a screen cap of how you're injected a neg clip, i've tried just the standard method we've been using on SD for ever, but it doesn't appear to do anything

mostly struggling with watermarks and signatures on outputs

and jpeg artifacts

[fleet meteor]

Aaaa how you get that lightiiiing? 😁 👏

[noble coyote]

70+ text prompts from today's img2img using Florence2

How would I setup some nodes to feed each prompt one at a time txt2img into KSampler?

[sacred jewel]

This is how I have it... I use the nodes shown. I will post the format of the TXT file in a sec

For flux, no need for negative... but it is up to you how you want to set it up

It's important for the counter to match the limits of the file or it will fail (if you want to do what I do and set a batch of 100 or 500 overnight)

or whatever number you want to process 😄

My prompt collection is up to around 8500 prompts so far... so I usually just target one section of it by setting the limits on the number generator to be between say 2500 and 4000 ... it will only then pick from that range of the larger file... or a single line if I want...

[bitter hearth]

[noble coyote]

W/f giving error "string index out of range'"

[tight onyx]

Are questions allowed here?

[muted cargo]

is that a question ? Quick ! Jump on them !

[tight onyx]

I want to know if sd3 has IP adapter support or not?

[sacred jewel]

Make sure your index is not greater than the number of actual positive prompts.

What I do is use notepad++ to count the number of "positive:" lines and set my index max one below that number as the index starts at zero

You can also test it by setting min and max to the same number so it pulls the same index each queue and test to make sure you haven't gotten past.

There probably are better nodes to do this out there but I am lazy and try to use whatever I have installed 🤭

[sacred jewel]

I believe only the API version of the model supports IPA. I have not seen a locally runable IPA for SD3.

[noble coyote]

Positive lines? Or positive prompts?

[sacred jewel]

Technically same. But making sure each line with the "positive: " keyword is legitimately one single line without splitting into multiple lines.

So, pattern must always be

'''

positive: some prompt

negative: some prompt or blank

positive: blah or blank

negative: blah or blank

'''

and so on.

[sacred jewel]

Prompts can be blank.

But always the index must be a valid line number in the file. If the number is outside the range, the node does not seem to handle it gracefully and errors out.

DISCLAIMER: I could be wrong as I figured this out by trial and error. I could not find in depth documentation

[noble coyote]

I have set both min and max to 5 - but it still errors out?!

Workflow

[hallow lion]

wtf lol

[muted cargo]

yes there is

[tight onyx]

@muted cargo can you please point to a guide for that? I need to know the required models/files and workflow.

[muted cargo]

i m at work rn but it should work similarily to sdxl ip adapter

Actually it looks like the code is ready but nothing has been released yet ?

[sacred jewel]

Try this...

Put a hard carriage return between the positive: and negative:

----
positive: This is a prompt for the positivity on Earth

negative: Please for the love of all that is holy, don't do this
----
positive: another prompt looking forward

negative: nope, this is not to be done
----

Save this file as : "test_prompt.txt"

Load it in the workflow and set the random generator minimum to "0.0" and the maximum to "1.0"

If you are using the rgthree pack, run just the GROUP that has these nodes (if as in my workflow you group them together) so as to only run those nodes to test.

Keep running it until it breaks or it "shouldn't" break 😄

It should randomly flip back and forth between the two prompts.

Now, set the min and max to 0.0 or 1.0 and test some more. It should only pick that single prompt over and over.

Finally, break it on purpose, set the max to 3 or 100 or whatever... it should break as it is out of bounds.

Not sure if this is a thing but I believe having the separators "----" (four dashes) on the very top and very bottom, helps... but haven;t stress tested that.

[noble coyote]

Searge SDXL i2i

[sacred jewel]

@hexed dirge which one should we download? 😄 THis is for the new Giger 2.0 LoRA 😉

[hallow lion]

Car for Oil mogul who thinks its cool but really looks dumb and kitsch. XD

[sacred jewel]

Giger 2.0 LoRA

[noble coyote]

Dumb, Kitsch & Kuler 😄

[sacred jewel]

More Giger 2.0 LoRA

[sacred jewel]

Style of CAD Drawing LoRA

[sacred jewel]

Technical CAD Drawing V1 LoRA

[bitter hearth]

yeah they need CAD drawing to build all the WOAH! signs

[steel beacon]

Does anyone have a good inpainting workflow for Flux that won't degrade the image each run through?

[bitter hearth]

bit of a mystery at the moment

there's an inpainting control net that can help a bit

adding some coloured noise to the area you want to inpaint, in the shape you want the thing to be, helps a lot

if we can get restart sampling working on flux that will be a big deal, cos you can run restarts around the sigmas where main the inpainting changes happened

[signal shuttle]

Bro we got Flux 1.1 before SD 3.5 😭

[errant dust]

From Fal: Hey! We're super happy to partner again with Black Forest Labs to offer both the new FLUX.1.1 [pro] (aka 🫐) and the 2x accelerated FLUX.1 [dev] models. https://blog.fal.ai/announcing-flux1-1-pro/

[frail shoal]

Lol. We'll have sd3.1 in two weeks...

[alpine summit]

Flux 1.1

[sacred jewel]

Envy Sleek SciFi LoRA

[sacred jewel]

Is it available locally yet? Or was this API/On-Line?

[sacred jewel]

Envy Heavy SciFi LoRA

[alpine summit]

Sd1.5

[mortal mesa]

so that blueberry model was flux 1.1, i was secretly hoping it was SD whatever, oh well

[bitter hearth]

I thought it was Chinese

[sacred jewel]

A wine store in a spaceship? I am all in 😛

[old venture]

A black car with black background with red fire burning tier

[sage burrow]

Has anyone found whether this lora training works with nsfw?

[craggy crest]

who's is that?

[bitter hearth]

its Fal.ai

[craggy crest]

that's a really really really tiny mech

[errant dust]

Her performance brought the house down

[sage burrow]

FAL partnering woth BFL

[craggy crest]

well - does their TOS allow you to train nsfw?

[sacred jewel]

Personal sized 🤭

[sacred jewel]

she seems nice 😛

[pseudo owl]

Flux 1.1 pro kinda seems worse than flux 1.0 pro and even flux dev. It is cheaper though.

[sacred jewel]

Huh? Seriously? How do you mean?

[bitter hearth]

I saw a few Flux 1.1 Pro samples that looked really nice

[pseudo owl]

It’s more cartoonish and also overall slightly worse imo.

[sacred jewel]

I think I am under a rock LOL.. I have not seen anything... not that I get out much lol

[bitter hearth]

haha it was on reddit I think

[sacred jewel]

Interesting. I don't find that example so objectionable... keeping my mind open for now.

and for the record, if it is not free, it is not cheap enough for me 😛

[bitter hearth]

to my eye, flux 1.1 pro is stronger in that example

text and beard slightly better
but more importantly the background blur is nicer

[pseudo owl]

Supposed to be 5 people, 3 woman, 2 men. Both are wrong but dev looks better imo and has 5 people, while pro doesn’t.

[bitter hearth]

I don't mean to disagree but I think Flux Pro's image quality is a bit higher there too

[pseudo owl]

Well I guess the image quality is objective so understandable but there isn’t really any true improvement, they seem pretty similar in quality.

[bitter hearth]

its very similar yeah

in terms of subjective style I personally dislike the flux look overall
my personal taste is models that look like photos

but flux is much better at satisfying average human preferences, which is the goal

[sacred jewel]

Steampunk Illustration LoRA

[sacred jewel]

Agree on this one... until I can do testing myself, I will not criticize 1.1 yet 😄

[bitter hearth]

flux will improve a lot when more tooling comes out

[pseudo owl]

Both have cons and pros but in terms of overall quality they are similar but since flux dev is open, you can usually do much more with it.

[bitter hearth]

at the moment SDXL/SD1.5 can generate at 4-5k without upscale
and flux is like 2k at most

but this will improve

[pseudo owl]

Yikes how long will flux with 4k res take.

[bitter hearth]

oh the speed will get better too
Bytedance said 4 steps and lower hyper loras are coming

I currently use the 8 step one

[pseudo owl]

Yeah waiting for the 4 step Lora too.
There’s this too: https://huggingface.co/RED-AIGC/TDD

Also 8 step but supposed to be more realistic.

[bitter hearth]

oh thanks I didn't know this one

wow they beat my two favourites- PCM and TCD

[sacred jewel]

Cyber Room LoRA (I guess the subject needs to be a room for it to trigger but I still like the look it generates on regular images)

[sacred jewel]

This one is more relative to the output the LoRA is expected to generate

[bitter hearth]

yeah I use loras on the wrong subjects a lot

[bitter hearth]

here are loads of flux 1.1 pro examples https://glif.app/@angrypenguin/glifs/cm1to7ws0000gnlxccqgp1gt0

[sacred jewel]

Vintage Computer Ad LoRA

[bitter hearth]

looks like flux pro 1.1 does slightly nicer compositions
and much, much better blurred backgrounds

[sacred jewel]

My first Glif Flux.1.1 gen 😄

Flux 1.1 Pro via Glif

[bitter hearth]

I think they're a bit better

than the dev ones

[cinder junco]

Just saw the news about Pro 1.1. I hope they will release new versions of Dev and Schnell without too much delay.

[hallow lion]

yes please. dev is amazing

[dusky thistle]

you should change the text part of your test images to be "WOAHWOAHWOAHWOAHWOAHWOAHWOAHWOAHWOAHWOAHWOAH" and see if it can do it

four letters doesn't test its limits but that would 😛

[noble coyote]

I think that substantial amounts of text are still a challenge that AI has yet to fulfill ...

[signal shuttle]

Flux 1.1 pro

[errant dust]

Most, yes

[noble coyote]

Searge SDXL img+txt2img

[hallow lion]

Have a tremendous Friday yallz.

[muted dove]

With his most loyal "friends"

[sacred jewel]

AGV [ANTI-GRAVITY VEHICLE] FLUX (CONCEPT) LoRA

[pseudo owl]

Normal flux dev which only supports distilled cfg vs de-distill flux dev which supports normal cfg(same amount:3.5, same seed:0, same steps: 28)

I prefer de-distill mostly, seems to have some more detail imo.

[sage burrow]

Gliff 😉

[sacred jewel]

The dog is far less platicky.

[pseudo owl]

Yeah, it does seem like a slight improvement for me. I also need to test OpenFlux which also supports normal cfg, and if it's better or worse.

[sacred jewel]

Geometric LoRA

[sacred jewel]

World of Darkness LoRA

[sacred jewel]

Escher LoRA

[sacred jewel]

Damn that came out fantastic... wow

[low stone]

aurum

[noble coyote]

Searge SDXL, Clownshark and Future Cubism LoRAs

[hallow lion]

She just came from Korean skincare clinic.

[sullen moss]

Flux Pro 1.1

[sullen moss]

Ages test

[hallow lion]

im still in awe of flux hands.

i have low standards

no more

[sullen moss]

After testing the new Flux, I can say that now I look at Flux dev as a plastic analogue.

[dusky thistle]

3rd order RES sampling (recently implemented in RES4LYF)

consecutive seeds... getting really reliable quality now

[hallow lion]

crazy good, wis i could decipher your workflows XD

is that a consistent character?

altho the stains are not yet consistent ..

[dusky thistle]

nah it's not about the character

or the WF

it's the sampling

i've been writing new sampler code

[hallow lion]

wish it was local, not that mere mortal PCs can run it fast enough probably...

[proud drum]

What is new flux?

[sullen moss]

pro 1.1

[deft niche]

Featured:

Soft Spot
by keshi

Monthly albums from hillqryc

View your featured history with '/featuredlog'

[limpid thunder]

Last 7 days <Sep 28 2024> → <Oct 04 2024>

• Member counts
• 346119 ↗ 346151 ↘ 346061 ↗ 346071 ↘ 346070 ↘ 346069 ↗ 346072
• Action members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 84
• Message members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 62
• Reaction members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 34
  More details

[bitter hearth]

can I live in that

[pseudo owl]

It's actually supposed to be faster then flux.1 dev according to them(3x faster then Flux.1 pro and almost 2x faster then dev). Looking forward for a flux.1.1 dev.

[bitter hearth]

give prompt

[hallow lion]

alrighty then but it wont be released for local use so ehhhh

sadly any image/text/video gen that can't be used locally is more or less useless

you need a lot of tries to experiment

thats why i dont bother with kling and luma and miramaz or whatever

its pointless

I mean for casual user who just want funny inflatable cats it;s fine but for anyone trying to do aything real forget it.

Sora is obsolete already for this reason.

[mortal mesa]

imagine having a pen where you cant use its output

[hallow lion]

yes we need more control

especially when ti coems to video

CogvideoX is very good but not there yet. But it is very good and I am thankful.

[hallow lion]

Yes Clown. Your stuff is amazing! 😄

How?

😄

genius

imagine if your image scna be conistsent (character and background)

and then fed into one of these video gens

Hollywood will tremble

[hallow lion]

alderaan.

[sacred jewel]

sharp photo of a large military ship in the middle of a barren landscape. The ship is grey in color and has multiple levels with multiple antennas and other equipment on top. It appears to be a modern, sleek design with a pointed nose and a flat top. The sky is clear and blue, and the ground is covered in small rocks and pebbles. On the left side of the ship, there is a small tank with a turret on top, and on the right side, there are two smaller tanks. The overall mood of the image is desolate and desolate., ArtDecoVintageFuture

It used a LoRA as named

[sacred jewel]

Vintage SciFi LoRA

[sacred jewel]

Drawing Lesson LoRA

sometimes it triggers, sometimes it doesn't

[pseudo owl]

I was just looking into this: https://ljzycmd.github.io/projects/MasaCtrl/
Allows for consistent backgrounds and characters, it's training free as well. Originally for sd1.5 but supports sdxl now too. No flux support yet but should work I think.

[sacred jewel]

Construction Style LoRA

[dull star]

this looks very crisp, nice

[sacred jewel]

Pencil Drawing LoRA

[proven pecan]

SD3 + Pika

[sacred jewel]

Dieselcore LoRA

[bitter hearth]

the last two look like Midjourney style
except Flux has overtaken Midjourney in quality lol

[sacred jewel]

Yeah you're right... It is the Promptcore Dieselcore LoRA

[sacred jewel]

Blackships LoRA

[fleet meteor]

Do anyone know why I get cuda out of memory error with flux on windows 11 (24h2) when with windows 10 it worked fine? 🥺 Maybe I should install w10 again

[hallow lion]

Boring lora on steroids

[noble coyote]

i2i Searge SDXL

[sacred jewel]

Wired Cyborg LoRA

[sacred jewel]

Vintage Mono LoRA

[cursive frigate]

anyone have any new awesome VLM suggestions?

I am currently using this doreilly/minicpm26_q5_k_m
with Ollama, but its been out for over a month and was curious if anyone has a newer better one they could suggest?

[noble coyote]

LLaVA and bakllava

d/load via models at ollama.com

[graceful spindle]

Error code 128 stable diffusion
What is the solution to this error? I have been trying to find a solution for two days.

[noble coyote]

https://youtu.be/p26KHQZzuLM

[graceful spindle]

I have tried this and all the solutions on the internet and it did not solve the problem.

I have sd 1.5 and it works normally but I want to install sdxl 1.0

[graceful spindle]

I am trying to update the program to the latest version and I am also facing problems

[graceful spindle]

I did all the steps except downloading the integration package and copying the repository file.

I don't understand how to do this

[pseudo owl]

Looks nice, what model?

[dusky thistle]

flux, but it's more about the sampling

[pseudo owl]

Oh interesting, what sampler then?

[dusky thistle]

one i've been working on

[tawny edge]

how to increase realism?

I'm using amateur photography LoRa + noisify to add film grain

res: 1248x2048, no upscaling

[tawny edge]

whats your workflow/settings? can I dm you?

[hallow lion]

lol how would that 1 wheeled/legged? contraption even move. in what direction? :))))

[hallow lion]

Urbex Lora?

[noble coyote]

i2i Searge SDXL + LoRAs