🏞｜general-with-images

[hearty karma]

LMS K vs DPM2 K

[oak osprey]

waste of 19 minutes

[smoky oak]

ah damn, I hate when I do that, and I only do LoRA's

even then, its still really annoying haha

[oak osprey]

well an A100 80G is like "retrain it bro, who cares"

another 19 minutes later

[hearty karma]

i think for doing 'realistic' faces imma use dpm2 k now

:-)

[oak osprey]

yeah i switched to Karras a week ago

tis the bee's knees

[hearty karma]

i mean it definitely did better with the hand than ddim

didnt try it

[kind quartz]

it is very slow but best, ppl using more dpm++2M

[oak osprey]

karras sigmas are a pattern you can include in any sampler

i use k-diffusion if that helps

the one that you like the most

oh i just realised i can finetune like 80 models with different settings and batch the test generations

just see what works best over time while The Computer figures it out

no, ew

delete that question, m8

https://giphy.com/gifs/photoset-conan-obrien-zeqgtki9ifa7u

https://giphy.com/gifs/police-car-chase-RYjnzPS8u0jAs

[dense tapir]

I love it when people say there is no reason to tense up unless you are doing, or have done, something wrong. I guess they never had a police incident before. Ignorance is bliss for them.

Every single day

yes

I am shocked at the sub 35 crowd as they want to be governed harder.

[sterile temple]

Parents and schools put all these tracking and monitoring softwares on their kids phones and laptops, they constantly posting their best selves to IG. so who's to blame when they grow up the way they do?

[dense tapir]

if only I had a little more vram

yep

[sterile temple]

salad fingers

[dense tapir]

I still have that image on a disc somewhere

I loved Salad Fingers and so many do not get the reference now

3 brain went nuts, you know? irl, kinda sad

spoons

peanut butter, bologna, and a poodle

[sterile temple]

what about the days of flash sites, like weebls. Magical Trevor

[dense tapir]

Yeah, all just memories

Now everything is woke or else deemed insensitive white supremacist facist.

[sterile temple]

and youtube, before people could make money from it. so they just posted funny stuff, no sponsors, no faceless channels pumping out AI voice over drivel wikipedia articles

[dense tapir]

In today's world I didn't show my face either, as many others don't, because I don't want to be doxxed and some wokester come assault me, or worse. Used my voice though.

I am waiting for them to be AI from the creator to their voice. Will not even be able to tell soon.

If only I had at least 8gigs I could run depth_zoe in my chain as it is the best most times.

[dense tapir]

https://www.ladbible.com/entertainment/salad-fingers-inspiration-david-firth-567561-20221205

[oak osprey]

i broke 1.5

[sterile temple]

comparison of samplers I just ran

[dense tapir]

careful, I was told by a mod that censor box is considered circumvention attempt against clyde.

even if I painted on a bra it is considered the same.

[smoky oak]

censorbox?

oh, I see it

[dense tapir]

I guess this isn't even allowed 😦

[sterile temple]

Well I didn't want to rerun the whole thing again, plus I put NSFW in the negative to try and prevent the nudity

[dense tapir]

I know, I am just saying the level of logic happening with this stuff right now.

[kind quartz]

clyde today makes me angry....

[dense tapir]

It makes bad decisions due to how it was programmed.

It has a really tough time distinguishing shades

or outlines.

[frozen nebula]

psychadelic cyber necromancer

[dense tapir]

I have a question for anyone out there that might can understand this. I have my models as a sym link. 1_x model to the hdd, and 2_x to the nvme. Works in Automatic1111 but not in Anapnoe's fork. I don't get why since that would be an os thing.

anyone know why, and how to fix that?

[dense tapir]

I see more updates for CN

I am just going to sit these out

Don't have enough memory to do it at 768x1024

[kind quartz]

KISS

[ripe cedar]

Considering his mouth is covered, he won't KISS

[kind quartz]

Knight in satan service, Music group :

[kind quartz]

🙂

[dense tapir]

This one didn't change much

weird

[dense tapir]

It was the steps

[dense tapir]

This is kinda nice if I had more ram I could perfect it more.

6gigs is just not enough but soon

[kind quartz]

ctu?

[dense tapir]

I am on 2.1 so we don't have that yet

no tile, and no style

[kind quartz]

o.k. i am as well on 2.1 now, Artius model

[dense tapir]

good model

[smoky oak]

lol

The alarming part is there are new models that can do pokemon art so good its almost impossible to tell its AI

like look at this Flareon I just made in like 15 seconds

these art models are getting insane

[dense tapir]

@kind quartz

[south quest]

new to stable diffusion, and i just wanted to share an image i had been working on today

[smoky oak]

nice!

[south quest]

thanks!

[dense tapir]

That is one HUGE image

[south quest]

yeah, it took 7 hours

[dense tapir]

OUCH

[static tusk]

7h for 1 image. thats dedication

[south quest]

it was fun 🙂

[dense tapir]

which video card do you have?

[south quest]

nvidia geforce rtx 3050

[dense tapir]

Ahhh, had you said a 7900XTX I would have died inside.

[south quest]

no, im too cheap.

so i had to inpaint it slowly

[dense tapir]

Well, I am too cheap to buy a 4090 so I am going to a 7900XTX or the 7950XTX if that becomes a reality.

After the way Nvidia has treated us all I am done with them even if that means I am a tad slower.

[south quest]

mines just built into to my laptop, i cant change it anyways

[dense tapir]

Oh, I hate that about laptops

that is what they would love to do to us all.

[south quest]

get everyone on laptops?

so what did you do to get that image?

[kind quartz]

@south quest arent you running it on cpu?

[south quest]

i followed the first guide i saw on google

i dunno

[kind quartz]

7 hrs is too long for gpu 3050 even mobile probably

[south quest]

oh, i have the original

[dense tapir]

Not get us all on laptops but the GPU is soldered onto the mother board so have to replace it all just for a GPU upgrade.

[kind quartz]

I would suggest you different tutorials, suggested ppl knowing very much. Some tutorials on youtube are bad.
Are you using A1111?
@south quest

[south quest]

thats the one the guide said

so i will say yes

[dense tapir]

My image is using control net and prompting with embeddings, etc... It was so large I couldn't 1:1 it so the background was not 100%. Didn't matter I was seeing how it did with the subject anyway.

wish I had styles but if I did proably couldn't run that in my workflow anyway due to 6gb vram

[south quest]

i only have 4gb vram, ive gotten the out of memory error many times

[dense tapir]

I need some anime to test on

8gb on sd 2.1 is when you have just enough room for most stuff.

[south quest]

i wish i had double what i have now, just trying to increase resolution gives me errors 😦

[dense tapir]

Oh, the more the meerier

I train stuff so I really do need 24gb

With 15gb on colab I would run out

sometimes

[kind quartz]

@south quest --medvram should be enough

[south quest]

i dont know what that means

[dense tapir]

heck, there are some extensions that 3090s see them using 16-19gb

for 4gb --lowvram

makes every way slower doing that because it pages in and out

[kind quartz]

do you have webui-user.bat? @south quest
Edit it that way
set COMMANDLINE_ARGS=--medvram --xformers

[dense tapir]

telling ya --lowvram for 4gb

[south quest]

ok

i will try

[kind quartz]

in tech support they are telling 4GB med, 2GB low

[dense tapir]

even at 6 a lot of this is making me do --lowvram now instead of medvram

they are stupid then

or, rather, ignorant because some extensions are demanding

now base SD yes

besides, who has 2gb of vram in 2023?

[kind quartz]

Supposing Xenahli has not much extensions now. And --xformers are powerful in saving memory arent?

you would be surprised. On notebooks many ppl

[dense tapir]

hmmm, 2gb. My 7870 from 2009 had 2gb

[south quest]

webuiuser opens the same page, where would i type this line in?

[dense tapir]

oh, yeah laptops. I hate laptops so never ever keep up with what is happening with them.

[kind quartz]

.bat? there is that set Commandline_ARGS=
Already but without arguments @south quest

i dont like it eighter, prefer power of desktop

[sterile temple]

I love laptops! 😄

[south quest]

im confused as to where it would go

[kind quartz]

in your SD folder, there is webui-user.bat, that you are running SD
Right click on it and choose edit

[south quest]

oh

thats very helpful

thank you

[kind quartz]

np hope it helps with speed.

[frozen nebula]

last image for the night quite pleased with this one

only small atm, needs upscaling but my poor laptop needs a rest

[south quest]

i can actually change resolution now!

thank you so much

[kind quartz]

you couldnt before? Weird 🙂

[dense tapir]

4gb

[south quest]

yeah, i would always get the out of memory error

[dense tapir]

you'd be surprised

[south quest]

had to be 512/512 unless i uploaded an image

[kind quartz]

oh i mean it was unable 😄

[dense tapir]

Big bubba comes over to break my computer's knee caps at times even with 6gb

[south quest]

youve been more helpful than the random reddits posts i googled to figure it out

[kind quartz]

like missing sliders

[south quest]

should have came here much sooner

[kind quartz]

in #🤝｜tech-support there are good ppl

[frozen nebula]

If you have low vram try these cmd line args

--theme dark --xformers --opt-channelslast --precision full --no-half --always-batch-cond-uncond --opt-split-attention --lowvram

[south quest]

what does that do?

[frozen nebula]

Though you need to install xformers and precision can slow it down a bit but gets better results

[south quest]

just gonna save that so i dont forget

[dense tapir]

OMG

--precision full I can't even run

[frozen nebula]

This has all of them

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

[dense tapir]

it makes everything 32bits

[kind quartz]

not sure about --always-batch-cond-uncond

[dense tapir]

I use it

[kind quartz]

o.k. then

[frozen nebula]

Like i can run medvram but it sometimes errors out of memory on bigger tasks, lowvram and tiledvae lets you run controlnet and proper upscaling even on sub 8gb gpus

The other settings stop weird errors with images not completing

[dense tapir]

No, I did use it but recently removed it. Here is mine

set COMMANDLINE_ARGS=--theme dark --medvram --xformers --disable-safe-unpickle --port 9000 --api --opt-channelslast

[frozen nebula]

Yep tbh those are the main ones the others I use are to stop errors on my old potato laptop (bless it)

[kind quartz]

has anybody tried different theme?

[south quest]

Ok, i never teied the upscaler in stanle diffusion as even the restore faces button ran out of memory

[frozen nebula]

I recognise that pose 👀 😆

[dense tapir]

That is what my experiment turned her into

I need some anime to test it on.

[frozen nebula]

In case you wondered thats with the Allys newest model, it does really good poses and hands

[dense tapir]

I am on 2.1 though

[frozen nebula]

Ah yes, those are great for 2.1 tbf

[kind quartz]

i think memory on 2.1 are bit more demanding?

[dense tapir]

768 square is

[frozen nebula]

It is yes, well not really but you need bigger generations or it looks like hot garbage

[dense tapir]

Sometimes hot garbage is a treasure

just not in this case

[frozen nebula]

Well g'night all happy generating

[dense tapir]

night

[kind quartz]

somebody trash is other treasure 🙂
Or i spoil that? 😄

@frozen nebula night

[sterile temple]

I'm loving this prompt. The way it's getting the model to create a mix of realistic photo and abstract illustrations is something I haven't seen come out of SD before

This might be one of the most 'realistic' looking photos I've ever generated

[wispy nest]

ive been trying to generate strange eldritch cities

[kind quartz]

@south quest is it working?

[south quest]

Yes it worked, but after spending all my time on one image i dont feel like a second

(Right now)

[south quest]

Besides, i have sliders available to use now. So i dont know what i can do

[kind quartz]

512x512 or 512x768 or 768x512
Dont know what image to do?

[south quest]

Would just need to think a bit

I used to play with midjourney

And coming up with ideas always took me the longest

[kind quartz]

illustration of Poe's Raven.

[dense tapir]

no where in my prompt is a humanoid mentioned

[tired basin]

Can you dm me the prompt?

[dense tapir]

iow, I have no idea why I keep getting the women

[tired basin]

Biased model

[dense tapir]

probably

[sterile temple]

should be in the png info

[dense tapir]

Yep, bias is showing with this test

what do I do add woman to neg?

[sterile temple]

i found some keywords will put the woman in the pic no matter what. doesn't matter even if you put woman, female, girl in the negative, it just keeps showing portraits of woman

[tired basin]

Add woman, 1girl, female to negative and cross fingers

[dense tapir]

I thought so

[sterile temple]

did you use my negative prompt?

dull colorless washed out de-saturated bw sepia hands detailed face child boy girl, (deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck

[dense tapir]

yep, I have my 2.1 neg special sauce prompt.

sadly it had to be remade many times since 1.5

most times it works but for this experiment with the model throwing women into it I need something else

[sterile temple]

does anyone know what you call that black pocky texture it sometimes ads?

[dense tapir]

woman, 1girl, female

I have no idea wth that is

[sterile temple]

ET spaceship

[dense tapir]

Well, that is the model doing that

Said "no thots? then here, have a big ugly ball thing instead." 😦

obviously this is a horny model.

model is rmada

trying orangutan

Ahem

@tired basinwth? Orangutan

[sterile temple]

rmada merge 2.1?

[dense tapir]

the other was but that last one was orangutan

it didn't even follow my neg

I am just trying random ones now

that was sdartcompleteedition

[sterile temple]

1.5?

[dense tapir]

all 2.1

I don't screw with 1.5

well, rarely

[sterile temple]

just looked similar to some of the images I'm getting with the suburban dusk look

[oak osprey]

well they always say your screw-ups are your best learning experiences

i used 1.5 images as class data for 2.1 training

[dense tapir]

Works, cause I did that a few months ago but 2.1 is not good to train styles on

[sterile temple]

I was thinking someone should do that, if people like 1.5 so much

[oak osprey]

it doesn't do prior preservation if you use images not generated by the model, as regularization images

so don't use stock photography or 1.5 images for that

[tired basin]

To be fair Im surprised that came out that good! Im not really happy with Orangutan now, and couldnt even get an improvement merging stuff in 😦

[oak osprey]

you need to generate native outputs from a single prompt for regularization images, and just use that

and it should be a very different prompt than the training prompt, and data.

i had the same prompt in my reg and training keywords and the same data in both and it just burnt the model immediately breaking it

[dense tapir]

Bit odd it refused to obey my negative prompt and so far it was the only one that hasn't. I am sure I will run into others that will not as well.

You can do that with image2image

crispy when you do it

Found another

[tired basin]

I think ICBINP is way better than orangutan

[dense tapir]

old old 2.1 model called redshift

[dense tapir]

yeah, but that is 1.5

[tired basin]

Despite it being a “1.5” model it still works sweet at 768x768

[oak osprey]

using 1.5 as reg images causes this progressive destruction to the model

[dense tapir]

well, I don't like 1.5 because all my lycrois/locon/HN/Embeddings are for 2.1

[oak osprey]

it won't let me paste the next images due to Clyde

[tired basin]

Shouldnt need em lol

[dense tapir]

I go to 1.5 and I am naked

Naw, you do

[tired basin]

Depending on what youre making

[oak osprey]

this is one of my reg images for 2.1 training

[dense tapir]

That is the uniqueness these asshole companies are trying to take away from us by going to a web based "solution" only strategy.

[oak osprey]

i remade my whole dataset

[oak osprey]

balls

[sterile temple]

I don't have balls that look this nice

[dense tapir]

I have two but if they ever bioluminescent I am seeking a doctor.

be fun as I have my own light source until I croaked

[sterile temple]

I'm running an x y script with this prompt to see what the different samplers do. Definitely some clear winners for realistic portraits. But maybe a sampler that sucks at portraits will do better with this prompt?

[oak osprey]

Euler A Karras DPM++ Solver 2M CV Lit, is likely the current best scheduler

[dense tapir]

For reality I really like dpm++ SDE Karras at 12-20 steps but it takes so long

Alright this model is I dunno

[sterile temple]

hey, at least it's not a woman for a change

[dense tapir]

favours something I hope it doesn't

I wonder what that is?

[oak osprey]

05/16/2023 04:35:01 - INFO - __main__ -   Num examples = 200
05/16/2023 04:35:01 - INFO - __main__ -   Num batches each epoch = 25
05/16/2023 04:35:01 - INFO - __main__ -   Num Epochs = 116
05/16/2023 04:35:01 - INFO - __main__ -   Instantaneous batch size per device = 8
05/16/2023 04:35:01 - INFO - __main__ -   Total train batch size (w. parallel, distributed & accumulation) = 16
05/16/2023 04:35:01 - INFO - __main__ -   Gradient Accumulation steps = 2
05/16/2023 04:35:01 - INFO - __main__ -   Total optimization steps = 1500```

[dense tapir]

I hope computex isn't shit

I get this sick feeling AMD/Nvidia are just shit in 2023

[oak osprey]

you're so damn cheery today

[dense tapir]

Arsenic or Ricin take your pick. AMD or Nvidia. I see no difference.

[oak osprey]

ha. need i remind you, "No one got fired for using Intel"?

[dense tapir]

That is our last hope

[oak osprey]

ARC 69-69

i can't wait for this fucking training to hit step 100 already so i can try the model out

[dense tapir]

dittotune

[oak osprey]

she's sweatin lul

training the text encoder and the 768x768 model

[sterile temple]

God, why does Windows do this?? Goes at 2GB/s then drops to 10MB/s for no reason

DPM Adaptive is pretty good

[dense tapir]

it is alright but takes longer

one of the dpm (FAST) is anything but fast and sucks

3m while all others take <50s kind of speed

[tired basin]

Runs out of cached info so has to go to source to get it

[sterile temple]

🙂

[dense tapir]

moon

sun next

sun

mind you all of these has the neg

it simply is ignoring them

[oak osprey]

what seed is that? 4204206969?

[dense tapir]

LOL

PMJ next

SHIT

[oak osprey]

shit

[dense tapir]

Base 2.1

[oak osprey]

i can tell its 2.1 beczu there's no areoals

[dense tapir]

oh, lord

[fading whale]

how do you guys add negative prompt when writing the prompt here?

[oak osprey]

100 step increments on training

[sterile temple]

What did you do here?

[dense tapir]

changed the model to an anime one

[sterile temple]

but are you using img2img?

[fading whale]

when will free trial end?

[dense tapir]

No, no img2img

[dense tapir]

June 15

[sterile temple]

GPT4 with web access. Let's see if it can show me how to fine tune a stable diffusion 2.1 model

[fading whale]

tell us later if it can

[sterile temple]

going well so far 😅

[dense tapir]

The openAI gpt is 3.5 for free. For the real GPT4 I have to use bing

there is a big difference in what it knows between them

[sterile temple]

Looks like the annoying subscribe to our newsletter stopped it from reading, which is odd. I thought it would just parse the HTML file

Sure, based on the resources I've found, here's a step-by-step guide on fine-tuning a Stable Diffusion model:

Find a Set of Images for Fine-Tuning: Based on past experiences, it was found that synthetic/AI-generated images train Stable Diffusion models faster. For instance, images generated by Midjourney v5 could be used as they produce consistently good quality results. You can compile a dataset of Midjourney-generated images through various means, including scraping Discord channels where these images are shared​1​.

Clean the Data and Prepare Captions: This involves some data cleaning and decision making. You might decide to only use the images that were upscaled by Midjourney users and use the user prompts as captions. The reasoning behind these decisions could be that the images upscaled by users introduce human feedback, and using the prompts could make the output model easier to control. A Python code can be used to clean the data and prepare the captions​2​​3​.

Decide on the Training Protocol: You can use a tool like EveryDream 2.0 for training. The key parameters for training could be a constant learning rate of 5e-07, total of 100 epochs, saving a checkpoint every 25 epochs, and using 15% of the data as a validation set. Here are the full parameters:

it then spat out a bunch of json config data, but broke part way through

Here is the continuation of the step-by-step guide:

Fine-Tuning the Model: The training process could take several days depending on your hardware. For example, using a single RTX 3090, the training took about 5 days and used about 10GB of VRAM. The initial few epochs took about 2 hours each, but later epochs were faster, taking about 1.5 hours. This reduction in time is likely due to the model not having to save a checkpoint after each epoch. The training process was reported to be smooth, and no significant changes were made after each epoch. During the fine-tuning process, you may observe that the model slowly starts generating better images. You may also notice that some prompts work very well while others don't. These findings can be helpful for future iterations​1​.

Testing the Results: After the model has been trained, you can generate images using different prompts to test the results. The quality of the generated images can be a good indicator of how well the fine-tuning process went. You may find that some prompts produce high-quality images while others don't. This is normal and is part of the fine-tuning process. You may also observe that the model slowly starts generating better images as the training progresses.

Iterating Based on Findings: Based on the results of the testing, you can generate hypotheses for improvements and keep iterating on the model. For example, you might decide to change the dataset, adjust the training parameters, or modify the training protocol. It's important to remember that fine-tuning a model is an iterative process, and it might take several attempts to get the desired results.

Please note that the above steps are based on one example of fine-tuning a Stable Diffusion model, and the specific details might vary depending on your specific use case and the resources available to you.

🤷 I dunno. Is that helpful?

[wispy nest]

have you tried Phind?

[dense tapir]

Never heard of it.

[wispy nest]

Same idea as bing, but idk how comparable, never used bing.
(too lazy to login, my browser doesn't store session tokens)

well...it does, I just made it delete everything when I close browser

[sterile temple]

Phind is good

I used at work a lot before they blocked everything 😒

[wispy nest]

I tried it just today, it failed at my coding related question...but it wasn't simple one tbh, we'll see

[dense tapir]

Seems it is working on mine

well, it is right so far

[wispy nest]

although just tested another thing and it actually made script I was asking for, better then gpt3.5 did it, almost clean code too.

[dense tapir]

where is the rest of the code? export MODEL_NAME="runwayml/stable-diffusion-v1-5"
export INSTANCE_DIR="./data/data_disney"
export OUTPUT_DIR="./exps/output_dsn"
lora_pti
--pretrained_model_name_or_path=$MODEL_NAME
--instance_data_dir=$INSTANCE_DIR
--output_dir=$OUTPUT_DIR
--train_text_encoder
--resolution=512
--train_batch_size=1
--gradient_accumulation_steps=4
--scale_lr
--learning_rate_unet=1e-4
--learning_rate_text=1e-5
--learning_rate_ti=5e-4
--color_jitter
--lr_scheduler="linear"
--lr_warmup_steps=0
--placeholder_tokens="<s1>|<s2>"
--use_template="style"
--save_steps=100
--max_train_steps_ti=1000
--max_train_steps_tuning=1000
--perform_inversion=True
--clip_ti_decay
--weight_decay_ti=0.000
--weight_decay_lora=0.001
--continue_inversion
--continue_inversion_lr=1e-4
--device="cuda:0"
--lora_rank=1 \

--use_face_segmentation_condition\

[oak osprey]

that's to train a LoRA

[dense tapir]

apparently it is only giving me the parameters, which I already know, for the cli lora trainer

lora_pti.

I wanted a program. Oh, well

I mean code

[oak osprey]

https://github.com/nitrosocke/dreambooth-training-guide

[dense tapir]

nitro is a friend of mine since at least Oct of last year

[sterile temple]

It's not possible to convert a 1.5 model to 2.1 is it? My GPT4 instance seems to be under the impression that it is possible

[dense tapir]

we trained a lot of shit in DB

[oak osprey]

ask it about transfer learning

[sterile temple]

Yes, converting a Stable Diffusion 1.5 model to a 2.1 model is indeed possible. This was explicitly mentioned in a tutorial I found on fine-tuning Stable Diffusion models, where the authors converted their SD1.5 model into a 2.1 model for the fine-tuning process. Unfortunately, I ran into some technical issues when trying to quote the exact text, but the information is directly from a reliable source who has done the conversion themselves.

This conversion appears to have been achieved with the help of the EveryDream team, which suggests that there may be specialized knowledge or tools needed to successfully make the conversion. If you're planning to do this yourself, it might be a good idea to reach out to knowledgeable communities or experts in the field for specific guidance on the conversion process.

[wispy nest]

No idea if it works or not

sources looks sus, probably not

[dense tapir]

I am finding that myself where the AI is making stuff up.

[wispy nest]

yea

[dense tapir]

I especially find that with the OpenAI site

[kind quartz]

you mean prompt?

[sterile temple]

https://followfoxai.substack.com/p/guide-to-fine-tune-your-own-general

I think it got the impression of 1.5 to 2.1 from this article. But I don't think they talk about that

[kind quartz]

all angry snowmen are smiling...

[kind quartz]

resting china emperor?

[dense tapir]

yep

[static tusk]

chatgpt likes to make up things all the time

[kind quartz]

whats this? @dense tapir 😄

[dense tapir]

In hell the snowman is king

[kind quartz]

hell will freeze

[dense tapir]

If a snowman can exist in hell of course he would be so bad ass that even Satan would bow to it.

[kind quartz]

interesting how it mixing arabic and roman numbers on one dial 🙂

[dense tapir]

Look at the rez on that

[kind quartz]

here perfect all dial. I-XII

[dense tapir]

Not perfect as the Roman Numeral for Four is IV

I hate it when I see IIII as it shows they have no idea wtrf they are just in some Chinese sweatshop cranking them out.

[sterile temple]

Those poor GPUs working away tirelessly in Chinese sweatshops

[kind quartz]

@dense tapir perfect. At dial are mostly used IIII instead of IV. There is several explanation

(I,II,III,IIII) (V,VI,VII,VIII) and (IX, X, XI, XII) is one explanation, another is IV means Jupiter starting letter and therefore forbiden to use.

interesting

[dense tapir]

Don't care about any of that the real roman numeral system explicitly says how it works and IIII is 100% flat out wrong.

I took latin in school so that shit was drilled into my head via text books and teacher

[kind quartz]

Simply said, watch dials mostly using this form, so result is o.k.

[dense tapir]

not all

the ones that use IIII I never purchased

[kind quartz]

i have not write all

[dense tapir]

I know you didn't, my friend, I was just saying you are right

Latin surely isn't, lol

Sad it is a dead language

[kind quartz]

yes, as well as old greek

[dense tapir]

I am unfamiliar with that one. I know Latin died off after its fall and the 600-700 years of our Dark Ages but was ancient greek for the same reason?

Look at V. In Latin that was their U so when, in English, we say double u being W notice it is two V (two Us).

[kind quartz]

yes i know that i was wondering

tried to check deforum, man it will be lot of studying.... @dense tapir

[uncut helm]

is it possible to erase a part of an image and then tell it to fill it in with something else?

I want to change the first image so it has chains like the second image without changing the entire thing

oh this channel is not for discussing how the platform works right? 🤔

[kind quartz]

i dont understand much. Probably what are you looking for is inpainting or add another layer, this is not SD thing that layer way

[uncut helm]

wdym by "this is not SD thing that layer way"

[dense tapir]

I do plan on goingto video but I need the vram for that.

[kind quartz]

add layer, it is photoshop things

or krita gimp.

[uncut helm]

uh? I can give the entire gimp project to SD?

[kind quartz]

No... check inpainting in img2img. I am not sure what are you trying achieve and my english is bad realy bad.

[uncut helm]

I just want to add chains to the first image without changing anything else

[sterile temple]

draw the the image of the chains on a separate layer with a white background. Then try img2img or use controlnet to make some images. Then add remove the white from the images and add back to your original image?

[uncut helm]

yes but then the chains won't blend in with the portal

[kind quartz]

if you inpaint in img2img chain as you have on right image and then turn it into "chain" i think it will look great.

[uncut helm]

uh this is really weird

[surreal thistle]

How do I fix her eye?

[kind quartz]

@surreal thistle in extras tab, two bottom things are for enhancing face and eyes

[surreal thistle]

I’m using Playground AI.

[kind quartz]

aha i automaticaly supposed a1111

[wraith karma]

Damn, can't send my image because it's considered explicit

[kind quartz]

i got same issue today with inocent image

[wraith karma]

Well, mine kinda is explicit

[surreal thistle]

Can you do this in Playground as well?

I can't use the official Stable Diffusion website because it's too pricey.

[kind quartz]

i dont know playground at all.
I got localy installed it and it is free.

[surreal thistle]

My laptop is potato.

[kind quartz]

o.k.

[surreal thistle]

Playground is the best free option.

In my opinion.

[kind quartz]

check if there is GFPGAN there or codeformer.
These two things doing so.

[surreal thistle]

Some girls I made.

They look OK except maybe the 5th one?

[kind quartz]

it is common ppl having different eyes.

[surreal thistle]

So it just makes it more realistic.

[kind quartz]

i will try to do something with it

[surreal thistle]

Playground AI makes some pretty good girls, I'd say.

It's also not bad at landscapes.

[kind quartz]

realy dont know if better or worse

[surreal thistle]

Actually probably worse.

She's more like a doll now.

[kind quartz]

because eyes?

[wraith karma]

Had to completely blur it to send

[surreal thistle]

Maybe.

These ones are also good.

But the skin is kinda weird on the last two.

I like how it looks like a mixture of a painting and a photo.

[kind quartz]

the beauty of local install is you can change what you want

[surreal thistle]

Still pretty fake.

[kind quartz]

it is tool. I am not expert on it. This just changed eyecolors. I am more fan of creating aliens 🙂

[surreal thistle]

What do you think about these?

[clear path]

nice images

[surreal thistle]

That's Playground AI for you.

[clear path]

🤡

[clear path]

is that sd?

[surreal thistle]

The eyes look kinda realistic.

[surreal thistle]

Yep, that's their new in-house fork.

[surreal thistle]

https://tenor.com/view/its-not-real-dwayne-johnson-seven-bucks-the-rock-its-not-the-truth-gif-20130196

[kind quartz]

your isnt as well real, its watercolor, isnt it?

[wraith karma]

No nsfw channel here, right?

[dense tapir]

Correct

[wraith karma]

I guess this is as far as I get with sharing my art then

[kind quartz]

Btw this egghead's words i can take seriously.

[dense tapir]

Clyde will smack you down even with SFW sometimes

[wraith karma]

It's funny that yesterday this passed

Since both are variations of the same image

[surreal thistle]

Yeah, but it looks more believable.

[kind quartz]

i changed nothing but eyes, note.

[surreal thistle]

Yes.

The sheer impact of eyes on the entire picture...

[dense tapir]

Hey, Purple, where have you been hiding at?

[surreal thistle]

Exams.

Very soon.

Must prepare.

[dense tapir]

Ahhh. Been a bit since I last saw you around here.

[surreal thistle]

But also must generate pictures of sexy girls.

[dense tapir]

LOL

[surreal thistle]

Good colibri.

[surreal thistle]

Proof that you can make a hot girl without showing boobs.

[wraith karma]

img2img is crazy good

[surreal thistle]

Do you think anything can be improved here?

[surreal thistle]

Joke images go brrrr

It DOES make frogs really well, maybe because enough data was provided?

[surreal thistle]

Which images of mine do you fancy the most and why?

[kind quartz]

guess decade

[dense tapir]

1968-1974

[kind quartz]

@dense tapir when played with duna interiours i cant get ppl so nicely spread

yes 70s

[wraith karma]

Generating anything in wide format seems like something SD doesn't like, huh?

[kind quartz]

probably you need hires checked.
Otherwise duplicates. But i personaly dont like hires...

[hearty karma]

why no like hires?

[kind quartz]

because different image and when small change strong artefacts. Dont know why. I mean latent denoise

[kind quartz]

btw hairfashion tag

@dense tapir reminds me movie Brasil

[dense tapir]

I loved that film.

[kind quartz]

yes in usa they must have happyend so movie was bit shorter 😄

[dense tapir]

yeah, I hate that

[wispy nest]

anyone can help me with stable diffusion here?

[dense tapir]

Never heard of it

[sterile temple]

freaky

[kind quartz]

@wispy nest technical issues there is chanell

[sterile temple]

I like the first one better, feels more alive

[kind quartz]

what a magical helmets 🙂

hires

supposed hires fighting with doubles

[dry crow]

What was your base resolution?

[kind quartz]

oh yes wide as well @dry crow i cant upscale much.
1536x512

o.k. just other way is to go bellow 512 to keep ultrawide and have it working on 8GB

[surreal thistle]

How did you generate 3 women at once? Playground can't do anything because when I put in "more than one woman" it gives me just one woman.

[dense tapir]

2.1

[surreal thistle]

Playground has 2.1 but it's restricted to 512x512.

[dense tapir]

oh, that might be why

[surreal thistle]

No, I don't use the actual SD on Playground.

I use their fork.

[kind quartz]

tiles

[tired basin]

If Harry Potter series was made by Dreamworks

[kind quartz]

looks like harry potter friend

[tired basin]

I'm trying to make the 3 in the Cartoon model 🙂

[kind quartz]

i somehow cant handle it at all, that cartoon model 🙂

[tired basin]

Tiled VAE not work for you?

And to finish the trio

[dense tapir]

I asked SD for a camel caravan and got this

[kind quartz]

been playing with hires... i am doing something wrong.
Why for example with low denoise there is this

[tired basin]

latent upscaler creates a crapload of noise - either use realesrgan/ultrasharp or a crapload more steps, or a higher denoise

[kind quartz]

o.k. @tired basin thank you!
Could be as well probably ram friendly classical upscalers.

[tired basin]

@sterile temple your prompt on the cartoon model 🙂

[sterile temple]

Nice!

[kind quartz]

Seems i can afford only esrgan_4x

[tired basin]

Gonna try ICBINP and Gibbon with a Ribbon now 🙂

[kind quartz]

Model and Loras?

[sterile temple]

Try the prompt from this, some pretty pics coming from this prompt

[tired basin]

Model is https://civitai.com/models/66347/rcnz-cartoon-3d - no Loras used

[kind quartz]

yes i got this one. But with not somehow well results

[tired basin]

ICBINP 🙂

[tired basin]

Optimus Primate

[kind quartz]

Another Alice in wonderland book? Alice in quicksand. 🙂

[tired basin]

Gibbon with a ribbon

Very cool prompt @sterile temple !

[sterile temple]

nice

[tired basin]

looks like a movie screencap

[kind quartz]

@tired basin its that cartoon model

[tired basin]

is the graininess intentional?

[sterile temple]

Time flies

[kind quartz]

A chaos theory.
Deep depth focus.
Highly detailed.
Cinematic, high quality color photo.
Ektachrome photograph.

changing only first line. And chaos theory gives me portrait...

not sure if cinematic making grainines

[tired basin]

I would assume so

I would also assume that a lot of the input images for the training would have had human characters in them

[kind quartz]

yes for sure.

[tired basin]

It does know a monkey though 🙂

[sterile temple]

Lost in thought

[kind quartz]

very artistic 👍

[sterile temple]

One of the best guitars I've seen from SD

Usually it makes a complete mess of the fret board

[kind quartz]

probably because most is out of focus 🙂

This is cartoon model and i know what you mean exactly

[sterile temple]

Change of heart

[south quest]

those command changes really made these images faster

[kind quartz]

how long? And what is native resolution? If not better lower and upscale lately

[south quest]

that was only 3 hours

[kind quartz]

🙂

[south quest]

🙂

[kind quartz]

what sampler and how many steps? Do you know?

[south quest]

euler a, and 50

alot of 50s

[kind quartz]

oh 30 should be enough

for euler A

[south quest]

i just dropped down from 70 as it was def too much

[kind quartz]

even 10 gives you sensible image on Euler a
Read about it, afaik 30 is value past which euler a doesnt benefit much.

[south quest]

huh, ok

[kind quartz]

Try it and you will see. EulerA 80-100 is suggested for inpainting i think

[south quest]

80 just gets weird and blocky on me though

[kind quartz]

10,15,20,25,30,50

[south quest]

30 does look the nicest

[kind quartz]

yes try it, you spare some time and gpu and all

[south quest]

will do for next one

[kind quartz]

good is just with very few steps you can have idea about result in euler A. Problem is it has different result than all others samplers.

[wraith karma]

Euler A can give good images even on something like 24

[south quest]

i did go through all the other samplers once and i liked the version euler a gave me more than the others

[kind quartz]

best but slowest are probably dpm++ karras

[south quest]

will give it another try later

[dry crow]

3 hours for generating in SD?

[south quest]

no, for making the image

[dry crow]

Ah okay

[south quest]

if it was trying to draw one image in 3 hours id probably turn it off and not make any

[kind quartz]

i though so for generating 3 hrs 😄

tiled diffusion i like it. No memory issues. But i know my fault

[wraith karma]

Guys, do you think I should get AUTOMATIC1111? I've been using SD with Invoke AI for two days

It seems okayish for now

[kind quartz]

i know only a1111 so i cant compare. But is easy to use.

[kind quartz]

i like such mysterious things. I mean somehow hiden

[smoky oak]

How many Sen alts did I miss while I was gone? lol

[sterile temple]

1 I think

[smoky oak]

I have a question for you all

if I am using a model that has a tag that is like

character (movie they are from)

how would I go about triggering that tag?

Cause the example I have just triggers the character name and movie separate

[kind quartz]

As a person not speaking English well sometimes getting surprising images.

[oak osprey]

@dense tapir want to see my 2.1 fine-tuning results so far?

[smoky oak]

They are pretty dope for a 2.x model

[oak osprey]

seed: 420
prompt a puppy, hanging out on the beach
steps 35
sampler: PNDMScheduler (default SD garbage scheduler)

@smoky oak i forgot to mention that i'm using the default scheduler which traditionally needs 50 steps to start to converge, and this is 35 steps

so i kinda gimped it in both tests and it's still pretty damn good imo

the key has been to manually set up my dataset and crop / centre everything at 768x768 by hand, carefully... using low batch size and a very low learning rate, and training the unet and text encoder simultaneously

[smoky oak]

Yeah, I was confused why you were like "oh yes, 70 steps"

I was like bro 👁️

until I realized you are using basic pipeline

[oak osprey]

70 steps was probably me referring to the training steps

i had the dreambooth script kicking out a ckpt every 10 steps so i could investigate thoroughly why it's burning so quickly

i did a lot of micro-testing like that to figure this out. i went to bed far too late

i have a lot of test results i can't really share in here

this is an interesting result from 400 steps of training on really bad data where it was uniformly downsampled/stretched to the proper resolution

200 steps later it was beginning to distort before it just turned entirely into a couch cushion

[smoky oak]

@frosty haven

it is insanely good at transferring the info of one image to another

[frosty haven]

Yeah, I saw this YouTube on it. Still isn't exact. Things always change in every generated image. It'll add a hat or flower or a few trees or an eye will change color. Still can't just rotate the camera position in the exact same scene. The AI only sees things 2 dimensionally.

Yes, I've been using ControlNet ref-only. And that's where my interest is fading.

[smoky oak]

oh yeah, your expectations are severely unrealistic if you expect that

[frosty haven]

I have no expectations. And as such they are not unrealistic.

It's just that this creation tool is a VERY limited one-off dumb-luck generator. Good for PIXIV NSFW images, but not really much else.

[smoky oak]

alright then, well if you are hoping to get that level of control, then bust of luck, cause thats happening no time soon

[smoky oak]

just because you don't have the skills/use case to use it doesn't mean its not useful

just means its not what you want/need

there are people that have made full publish quality projects out of SD that look just as good as traditional media

[kind quartz]

@smoky oak need help, can you help me pick image?

[smoky oak]

I can try, but I have to pee so I will brb

[kind quartz]

o.k. i post it here

[smoky oak]

back

speedy pee lol

[kind quartz]

for pow.

[smoky oak]

oh, let me check what the POW is

[kind quartz]

i am for one of those on right

[smoky oak]

what is the POW?

[kind quartz]

cinemon something catchy. It is competition here

WEEK#31 Cinemotion

[smoky oak]

hmmm, I am not a huge fan of any of them, but I like the last one the most of them

[kind quartz]

Picture of week?

[smoky oak]

prompt of the week

I know what it is, just didn't know what this weeks prompt was

[kind quartz]

o.k. me too, but i am terrible in judging 🙂

@smoky oak Thank you very much!

[dense tapir]

Yes.

[oak osprey]

#🏞｜general-with-images message that's this

[dense tapir]

Not bad. btw, I had someone who trains a lot on 2.1 hit 0.12x loss. I am awaiting to see what they used.

[oak osprey]

loss is a shitty indicator of model health

totally burnt the shit out of 2.1 last night and made it no longer function anymore and it was a pristine .5 loss, the same it started with

i was messing with fine-tuning a fine-tune though, so, YMMV

my point is merely that the original model had .5 and mine had .5 and mine didn't do the thing anymore i assume because the text encoder got absolutely rekt

loss refers to the autoencoder's ability to reproduce its inputs without any modification occurring

[crimson dawn]

@clear jacinth this was using depth controlnet

[earnest dune]

in inpaint, should i use latent noise when im adding a new object to the picture

[crimson dawn]

major changes i use noise 👍

[hearty karma]

I asked a friend earlier today if I could borrow one of their line arts to mess with on AI

and it looks fairly good

[kind quartz]

and what is difference latent and latent antialiased? In Hires?

[dense tapir]

Under no circumstances have I ever been able to get anything out of 0.5. 0.45-0.5 is where the style falls into oblivion and the training session is just a waste of time.

[oak osprey]

interesting

[dense tapir]

TE for 2.1 is what fights you and why 2.1 sucks for training on. Same damn dataset that fails (high losses, which means it isn't learning btw) in 2.1 pull it into 1.5 and POOF, it just works. I tried that twice and I have no trained since and refuse to waste my life doing it on 2.1 on Colab ever again.

[oak osprey]

well the same shit breaks 1.5 in the same ways for me

there's different implementations of DreamBooth and some are more broken than others.

[dense tapir]

Well, mine is styles and I found training on 2.1 for subjects to be easier but styles forget it.

[oak osprey]

oh

i am adding features

not styles or subjects

[dense tapir]

Yes, 100% and I firmly feel the entire Kohya shit is fubar for 2.1

[oak osprey]

im going to go and do another set of tests with frozen text encoder for a majority of training until last epoch

[earnest dune]

it looks like it is sick :(

[oak osprey]

once i find an epoch i'm happy with, go one earlier, and unfreeze TE

[dense tapir]

a lot of the schedulers he implemented are half implementations, or were when I used it, per the original papers.

[oak osprey]

the 2.1 docs for DB recommend this, btw. unfreezing the TE only at the end, to bring it up

it is overtrained very quickly compared to the unet, which i think is quite robust tbh

[dense tapir]

yes, I bet that would help

[oak osprey]

you can't apply a different learning rate to the text encoder vs the unet which would be pretty damn cool

[dense tapir]

For Lora/Lycoris/locon we do not get such a luxury

[oak osprey]

i am working on adding the concept of areolas to SD2.1 without overfitting it or ruining its generalisation

not to publish the model but just because i'm curious if i can do that

it seems like a lot of people think it is not doable

[dense tapir]

we can train unet then go back and train TE but it will not be the same picking up exactly where it left off per the dev

[oak osprey]

you can set your torch and numpy seeds for dreambooth training, fwiw

gives more predictable results

[dense tapir]

we can't as the seed is only for images created

[oak osprey]

the seed is for random numbers generated

[dense tapir]

we don't have that

[oak osprey]

i do

[dense tapir]

Should help.

[oak osprey]

yep, it helped me identify some issues last night

made a ckpt every 10 steps

that used a lot of disk space.

[dense tapir]

yes, all this randomness is like asking a tech to fix your PC, or car, and it sometimes does it. We are screwed trying to find what the problem is without the ability to reproduce it at will. Hence, the same seed being used here.

That is what I do not like with noise based art because noise, at its core, is chaotic. Chaotic good, or evil, is subjective.

[crimson dawn]

@clear jacinth heres the loop backscript approach. i combined it with controlnet too here. as you can see it kind of jazes it up over time

#5 is my fave. Johnny #5 i'll call it

[dense tapir]

Damn newegg, as they keep MIR my case/chassis for the same price it was pre MIR. I don't screw with MIR any longer. 3 times the MIR was extended in a month.

they just extended it for another week. 😦

[oak osprey]

training restarted 🙂

looked at my reg data the script made for me on the last try, and it was terrible. so, redoing that now with better prompt

[wraith karma]

SD tires me so much for some reason...

[mystic marlin]

First of the day from me. lol

[wispy nest]

Hi

Bye

[pallid fractal]

hello im trying to install sd on miniconda and i keep running into a problem

any help?

[oak osprey]

https://images-ext-1.discordapp.net/external/lsBBam3ewH1ygQVp3tH217f_XEDEe3-IIRVWuwtzyWA/https/tripleback.net/public/discord/1684264939.0630774656fca339aee58847ddb9df5f0eeef7a.png?width=1004&height=1004

[golden sorrel]

[dense tapir]

Yes, it is not allowed

use webgui get notified to soon, regardless what they currently say, be banned. watch

[golden sorrel]

should i invest into a gpu server with google cloud

$400 bonus is promising

[dense tapir]

I wouldn't

[golden sorrel]

damn

how about paperspace

a bit slower but still does the job

• doesnt shut down whenever it wants

[dense tapir]

I never pay as I prefer local and free on paperspace hasn't been a thing inforever

most went to runpod if they pay

[oak osprey]

yep, still does that

local 4090 running on solar

the whole front array's output is basically going right into the 4090

it has made 412 class images so far

[golden sorrel]

damn

[hearty karma]

First time messing with different size gens

[crystal bloom]

testing a pixar art style model

[hasty nova]

Hey everyone! I just thought of a really neat idea for when I get around to making v3 of Digital Diffusion! I want you to send me your best gens from either v1 or v2 of the model, preferably through dm so I can find them easier, along with the prompt. I will include the images in the training dataset for v3, which I am thinking will be a complete retrain. It doesn't matter the size or aspect ratio. I am hoping this will increase token diversity and such, and possibly improve prompting in general, since I am kinda overhauling the text encoder with these models.

[dense tapir]

Damn, the lights are out due to a storm.

Running on ups to finish some work.

[hasty nova]

Actually on second thought you can probably send an image of at least 768 resolution from any model and as long as you send the prompt with it I won't care too much.

[hasty nova]

I also got a storm going here as well, you in Texas?

[oak osprey]

@dense tapir go on, dox yourself

[hasty nova]

Texas is large enough that it doesn't help much with anyone tryin to find ya

[oak osprey]

lol

10 steps in i can make the lord of the rings go surfing

i'm too powerful

@dense tapir if we don't find a way to train this stuff correctly, this puppy gets burned 😦 literally. the model stops making the pupper

it's interesting that feeding SD the entire movie The Hobbit made the puppy's hair go kind of auburn

sometimes, Sanity Puppy looks at me like, why are you doing this to me, papa

[gritty cedar]

wdym feeding it the whole movie?

[hasty nova]

He took "action frames" from the movie and trained it. Around 446 or something frames if I remember correctly.

[oak osprey]

it's learning

[hasty nova]

did you caption the images?

[oak osprey]

my script just trains a keyword

[hasty nova]

oh, so lotr

[oak osprey]

yeah

[hasty nova]

are you training on top of DD still?

[oak osprey]

aye

[hasty nova]

sweet

[oak osprey]

was thinking earlier, could pull e621's images and all its tags from their public collections, and run that all through BLIP fine-tuning to improve its image tagging abilities

build on top of the Salesforce BLIP model, which is already quite impressive

this is the default scheduler and everything at 35 steps, so, not trying for impressive details here

but definitely seeing the changes starting to compound

[lucid hamlet]

Death&Rebirth

[rain gazelle]

evenin'

[ripe cedar]

Anyone know if ComfyUI has img2img color correction?

[wispy nest]

i found this extension in automatics webui extension list https://github.com/Physton/sd-webui-prompt-all-in-one
and apparently it has some tencent api stuff built into it and it bricked my venv
reinstalling now

https://tenor.com/view/rachet-cat-attack-death-gif-22457670

thought it would be cool to be able to edit prompts easier

but this shit is bloated with api and proprietary code

[smoky oak]

https://tenor.com/view/whimsyy-aaja-angry-cat-funny-jokes-gif-26279289

howdy doody you crusty cretins

[wispy nest]

https://cdn.discordapp.com/emojis/818003174786334730.gif

alo

[kind quartz]

is this o.k.? Just want to know if height 384px as source for hires, is enough for 2x resiz, or anything bellow 512 is in any case bad.

[south quest]

i wanted to recreate my favourite midjourney image with stable diffusion inpainting, and wanted to share the final result

[proper scarab]

niiiice not to be nitpicky but the graphics look like early Xbox game.

Give that hottie some NKMD upscaling and some extra textures and it's 👌

[south quest]

i have no idea what nkmd is

[proper scarab]

upscaler for adding natural textures

although I'm not sure if it would work on cgi tbh

[south quest]

if i was to download nkmd, would i be wanting the superscale v2?

[proper scarab]

o: I've never heard of that one. link?

[south quest]

https://nmkd.de/shared/?dir=ESRGAN/Models/Realistic%2C Multipurpose

found a reddit link to this

[proper scarab]

woaaah I didnt know there were that many

I would say try em all. Im about to test some of them myself. I dont even know which one I have

if that one doesnt make it look better, look for another upscaler called nickelback

since the image is already big I would just run it on a 1:1 upscale, or if ur daring u can do a CTU 1:1

[south quest]

i already set it to upscale, soooo im just waiting for my 15k* resolution image to finish

hmm, just made it somewhat darker

[dry crow]

Which upscale technique did you tried ?

[south quest]

chainner

with the

4x_NMKD-SuperscaleV2_46k

i cant upload such a large image, but as far as i could tell it just shifted the brightness down a bit

[kind quartz]

Just asked, is 384 as source for hires good enough or anything bellow 512 is bad source? Thank you in advance!

[dry crow]

Ah okay, didnt tried chainner yet.
You could try using HiresFix or sd upscale script

[proper scarab]

holy moly macaroni.

[smoky oak]

oh look, its another sen alt

[proper scarab]

thats base prompt no upscaling no nothing, on top of being a dreamboothed girl

I think I finally got dreambooth down for torch 2.0, I really wish we had a dreambooth discord though.

this power...it's pulsing through my fingertips...

https://tenor.com/view/thanos-infinity-war-avengers-gif-14003260

[smoky oak]

wonder how long till you get banned again

[proper scarab]

what is this guy talking about

[smoky oak]

sen, you really think its not blatantly obvious that its you lmao

[proper scarab]

https://tenor.com/view/huh-heh-interesting-confused-confusion-gif-22475884

[smoky oak]

brand new account, just joined, same plastic looking gens, and being happy about torch 2 working lmao

[proper scarab]

I have no idea what this guy's talking about

blocking this bozo

[smoky oak]

sure lmao

you should be banned again in no time anyways

[proper scarab]

https://tenor.com/view/rip-bozo-gif-22294771

[smoky oak]

just stay gone bro, stop coming back

[proper scarab]

Ugh I take it back, it's actually kinda undertrained...

[oak osprey]

oh it's Senran Wrap

[smoky oak]

yeah lmao

[proper scarab]

idk who these people are mistaking me for. I literally just came from the other SD discord lool

[smoky oak]

its actually pathetic how hard you are trying to fake it

came from the other SD discord, yet your account was made today, alright, Sen

[proper scarab]

Interesting discovery, this time, swift didn't get disfigured after training the model...

[smoky oak]

and now you are talking about taylor swift again, how blatantly obvious can you be

[wraith karma]

That must have been a shit ton of work, amazing

[south quest]

yeah, i spent about 3 hours on it. Even though it was just inpainting over my existing image

[wraith karma]

only 4 toes btw

[south quest]

the original had 5, but i couldnt get stable diffusion to keep all 5

possibly because they were slightly merged together on the original

[kind quartz]

@tired basin ICBINP i like it

[balmy onyx]

yoo

[wispy nest]

Guys why I have this error?

[dense tapir]

My ip does that already. So....

[dense tapir]

This had nothing to do with a storm since it was just a sprinkle. My city lost power yesterday and I got mine back in 11h. 12k without power scattered all over the city but my part of the city was where most lost it.

[smoky oak]

ah, looks like I wasn't the only one that lost power

recently at least

[dense tapir]

Yes

[smoky oak]

we should finally be getting internet again today

[dense tapir]

Well, my ups stayed on for 2.5h then I went out to come back to it having used all its power. Not sure what sucks so much power from it when all was turned off but one cled clock.

My battery has 4/5 bars now. I hope it goes 5/5 else I will have to buy a new one soon

but up now for 4h so probably not enough time to recharge though this is the time the battery goes.

[smoky oak]

I don't have a UPS, but I may get one when I have money to throw around

[dense tapir]

Here, if you don't have one you will lose the electronics.

I finally had to get one. Battery replacement was, pre bidenflation, 36 bucks, no idea now.

OpenAI GPT is pretty stupid and refuses to listen to my commands.

[smoky oak]

yeah, chat GPT has gotten worse and worse as they cut it down more and more

[dense tapir]

I tell it what I want. Perfect. I then say use all of my prior requirements only leave the last punction mark off from each response. Leaves it off and forgets what I said I wantred. Tell it to do that but do this as well and it tells me sorry, and explains this is how I will do it from now on. Sends me responses with the last punctuation mark. Go back to scorn it and tell it what I want and it goes back in a circle.

It simply is worthless except for very simple stuff.

I swear it used to be better.

[sterile temple]

3.5 memory isn't very good

[dense tapir]

Well, even when I tell it to do XYZ it does XY but ignores Z. Makes me so aggravated

[smoky oak]

older versions use to be better with that

also, I know that the free version has a much more limited memory cap compared to the paid version

[dense tapir]

After 30 minutes of fighting it I gave up and thumbs down it and told them what was wrong.

I would never pay for this having only a glimpse of the free version.

It is that bad.

[smoky oak]

certainly wouldn't be the first time that an impressive AI got worse to try and upsell you. AI dungeon was so damn cool before they made it so terrible to force you to pay

[dense tapir]

Normally that pushes me away.

When it doesn't follow my instructions to a t then I have no incentive to pay

[wispy nest]

Need more Shrek

[dense tapir]

Don't worry a new Shrek is being made right now

I mean the movie/film

[dense tapir]

Damn, lights just flickered a couple of times.

My new drive arrives today, YAY. I think magician will clone the drive? Not sure.

[wispy nest]

For real?

[dense tapir]

For real with Mike Myers

[wispy nest]

Ahh it's released in 3 days

[dense tapir]

yep, he has been working on it for a bit

Eddie Murphy I hope is in it too

[wispy nest]

Sounds hott

Can't wait for new Shrek come out, imma wear a suit and buy five star dinner

[dense tapir]

LOL

Yep, clones, so cool.

[oak osprey]

oh, good

they got KID CUDA

he only travelled a few inches but that puppy went 4500 steps

fine-tuning 2.1 with supremely low learning rate

for 5000 steps, this was a painting

now she is real at 5500 steps

[dense tapir]

@smoky oak I just asked it "what were the instructions I gave to you?" I wanted to see what it was doing/thinking and I get back "Apologies for not following your instructions accurately. Here are five more examples that adhere to your specified requirements:" . HUH?!? Sheesh.

btw, even that didn't follow my instructions, nor the 5 it gave.

[oak osprey]

yeah they had to increase capacity and i think they did that by changing its sliding window length

the API still has the ability to consume the same token count but it does so less effectively than when GPT4 first came out

[dense tapir]

Yeah, I can tell it became dumber.

Just 2-3 months ago it followed what I said pretty damn well.

[dense tapir]

Bing's GPT-4: I’m an AI language model and my capabilities are limited to generating text-based responses to user queries. Is there anything else I can help you with? Worthless to me.

It is just a search engine. Sad, as it worked a week ago.