🆕｜sd3

[halcyon yarrow]

yes it does it's very well trained on comfy i know from personal experience

[craggy crest]

when was th elast time you heard openAI say that ChatGPT knew anything about how to use ComfyUI, much less those nodes?

[halcyon yarrow]

i literally had it code up an image to video node and got it to work

[craggy crest]

you were fortunate

[halcyon yarrow]

lol no dude you were wrong

[short thicket]

I have a zig zag setup that's interesting.

[halcyon yarrow]

chatgpt is highly trained on comfyui since 4o came out, and if you know how to use it then it's able to do anything, if i can have it code a complex AF node like image 2 video then a simple change like having an input field that affects the other ones should be a breeze

[craggy crest]

i'll let you argue this with @dusky thistle

[halcyon yarrow]

check it out this is the video I made using mochi

[short thicket]

there were all sorts of MBW patterns for the SD UNets. Flux is a different beast though as far as architecture.

[craggy crest]

where are her arms?

[halcyon yarrow]

mochi only has an empty latent node so you can only do text 2 video, i created a image latent node where you do load image > vae encode > image latent mochi video and i got it to work

[short thicket]

Plus these are merging on the 24gb models so it takes a lot.

[halcyon yarrow]

lol i mean it's not perfect i'm running like super low rez fp8 quantsizied stuff just to get 2 seconds out of it on my 8gb vram hardware you're missing the point if you're gonna point out issues with the quality

[craggy crest]

yeah. flux is using LLM architecture

[short thicket]

especially for more advanced merge methods that involve making calculations on every block

[halcyon yarrow]

are you limited my vram when merging?

[craggy crest]

https://www.reddit.com/r/StableDiffusion/comments/1gkb60d/run_mochi_natively_in_comfy/

there IS a node for mochi ...

[dusky thistle]

chatgpt is like a 4 year old when it tries to write sampler code

[halcyon yarrow]

this whole argument started bc mangler griefed at the manual labor involved with changing every field to be the same value, my response was it should be super easy and straightforward to adjust that node to allow you to do that with chatgpt, i think if it was me I'd spend 15 to 30 minutes getting it done and save myself the drudge of having to do it manually

[dusky thistle]

or even simple stuff like dealing with tiling

[halcyon yarrow]

yeah that's what i did, im using all the built in nodes it was great and easy

[short thicket]

the method I showed you isn't too bad in terms of the time it takes. But I've tried the Dare and Della methods and both of those are HEAVILY taxing because they are running matrix calculations on every single block before merging. I spent a lot of time just figuring out how to do them without getting OOM errors on my 3090

[dusky thistle]

i use it to look up funtions but i don't trust it to write anything for me that's more complex than... a bash script to rename some files or something

i sually end up spending more time fixing its mistakes than just writing it myself

[halcyon yarrow]

i think it has a lot do with user experience and skill, it's only as good as you're able to be, you gotta understand it's nuances and you gotta be able to get on it's level

[craggy crest]

clownshark is a hell of a programmer

[halcyon yarrow]

there is something to be said about finding a balance between fixing it's mistakes and just fixing it yourself, sometimes you're riight it's more of a time waste to have it fix things then to just do it manually but it's def a good helper in that regard where it can get me 90-95% of the way there and then i just take over and adjust the final bits

[craggy crest]

he could probably write the code to create chatGPT

[dusky thistle]

i've found it gets yu more like 20% of the way there, then goes 15% of the way in some random wrong direction

[halcyon yarrow]

i'm not doubting that he did make that cool alternative to ksampler, to each their own right?

[craggy crest]

i was serious when i said he could probalby write teh code to create chatGPT

[halcyon yarrow]

as someone who has chatgpt code more than 90% of all my stuff i highly disagree, it's user error or lack of knowledge if you can only get it 20% of the way there

[craggy crest]

i would suggest you listen to the man

[halcyon yarrow]

it's a matter of opinion crystal, there's no right or wrong, to each his own

[craggy crest]

not a matter of opinion. where is your code? where's your repo?

[halcyon yarrow]

you just gotta spend time kinda learning the ins and outs of chatgpt to truly be able to master it, some are just biased and don't want to master it, there's no right or wrong here

[dusky thistle]

give it sapmle code from comfyui, something like the code for dpmpp_2m so it has an example

[craggy crest]

shakes head

[dusky thistle]

then ask it to write a version of that that does implicit sampling, using Gauss–Legendre quadrature

then def paste the code here and ping me haha

[halcyon yarrow]

that's what i mean by you're doing it wrongn

i'd never approach the problem like that

[dusky thistle]

i was able to write it myself in like... 3-4 min

[craggy crest]

it's always interesting to listen to amature coders that think they know more than they do argue with a master programmer

[dusky thistle]

prolly couldn't even explain it to chatgpt in that time

it's def best to just learn how to code the stuff yourself, it's slow at first but much faster once you've gotten some experience

[halcyon yarrow]

you gotta sort of prime it to be able to get you there, so for example when i had it make for me the image 2 video now i gave it to the EmptyLatentImage node, I gave it the MochiLatentImageNode and I gave it all the code for nodes_latent.py so it has a frame of reference as to what's going on. i didn't just go in with a basic example and ask for somethinig complex that's what i mean by user error

[craggy crest]

son, he KNOWS how to prompt ChatGPT

[dusky thistle]

well, see if you can get it to write code for that, no matter your approach

[craggy crest]

you apparently don't

[dusky thistle]

and let me know if you succeed

[halcyon yarrow]

16 years of programming experience i think it's a little messed up to assume thhings about people but to each their own right?

[bitter hearth]

the the in-context lora paper was very cool but I find it hard to see the use case for making panels of 4 images

[dusky thistle]

if you do, i'd ecrtainly be interested in knowing the approach

but i am def not seeing that happening lol

[craggy crest]

i've got twice that and i wouldn't even try to compete with him

[halcyon yarrow]

i had no problem with's having it make the panels in any of the loras except the couple one

[mortal mesa]

why did you turn it into a competition plotting 2 people against each other though

[halcyon yarrow]

good for you i'd just refrain from callng people you dont know amateurs its a little off puttin imo

[mortal mesa]

all ego if you dont like Pepsi your out of the club

[craggy crest]

i didn't. Shark challenged him

[halcyon yarrow]

bottom line is the approach of feeding it a simple example + prompting a complex ask is a bad strategy, you wanna feed it plenty of examples all int he relevant domains and then before you even tackle the code discuss stuff like strategies and even ask it to ask you if there's anything else it needs to know, it'll be like "in terms of this how do you wanna do this?" and then you can think about it and guide it, just saying its a lot better than 20% of the way there and if that's as far as you get you're 100% doing it wrong

[dusky thistle]

well, if you're able to pull it off, lemme know

i gave up trying to get any LLM to write complex code with lots of tensor math

[untold valley]

the only acceptable drink is Dr.Pepper

[halcyon yarrow]

@mortal mesa check out the model page for the incontext loras in civit i created: https://civitai.com/models/929592/creative-effects-and-design-lora-pack-in-context-lora I've been spending the day playying with the loras and posting images for each lora so visit each tab and scroll down to the gallery to see my creations 👍

[mortal mesa]

you were busy

[halcyon yarrow]

@bitter hearth see plenty of 4 panel examples

to be fair chatgpt generated all the prompts i just guided it with basic ideas

[bitter hearth]

oh I wasn't very clear
I was saying I don't see the need to make images of 4 panels
I agree it can do it well

[halcyon yarrow]

its such a trend on civit tho, especially for nsfw images 🤫

where each frame is a scene in some erotic story

[bitter hearth]

its still a big deal yeah, since SDXL could not do this

[halcyon yarrow]

there's tons of loras for sdxl/pony that do this for sure, sdxl actually has way better ones where they're not uniform panels

[bitter hearth]

ah okay wasn't even aware people made panel loras for SDXL lol

SDXL does seem to still have capacity to improve with a lot of training

the latest realvis model was a big step up from before

[halcyon yarrow]

1055x - ❌- realvisxlV40_v40Bakedvae
I've generated 1055 images with realvis before i deleted it/banned it, wasn't too happy with it's performance overall

here's a good one from realvis, hard to believe it's SDXL sometimes

[bitter hearth]

ah okay, do you happen to know any realistic models that are better?

[halcyon yarrow]

statistically pony realism is probably the top most used realistic model on civit, my fave is real dream, like in terms of realistic women vs realistic in general is two different questions though

[bitter hearth]

I always wondered why real dream was ranked so highly

is the SDXL version of it also good or is it the pony version in particular

[halcyon yarrow]

babes by yogi and chromax by savy are also top notch

pony in particular, my favorite SDXL model is called mixtape v2 resonance but it's hard to be a fan of any sdxl model personally

i think what astra did with the custom CLIP L for pony sets it apart from SDXL significantly enough where it's in it's own class

[bitter hearth]

its only really considered a separate model because of an early choice Civit made really

[halcyon yarrow]

no i think they did it because of the difference with the CLIP models, they're not cross compatible so it's important to distnguish the difference technically to be able to make things work logistically

for example you can't use a pony lora on sdxl but you can use an sdxl lora on pony

[bitter hearth]

this is just the weights being heavily changed though

I get that they had to communicate it to people

and the way they thought of saying it was telling people its a different model

for logistical/commercial reasons

[halcyon yarrow]

@short thicket I'm posting a bunch of images under your model by the way, not sure if you saw the earlier ones of the home decor with the vibrant colors, he's mangler's take on the same prompt and here's FluxFusion on the right.

[halcyon yarrow]

that's exactly what i'm saying, it goes beyond the weights, it's not just the weights changing, it's the CLIP L changing that really breaks cross compatibility

no other SDXL finetune afaik has come out wth a custom clip L. so for example when my queue is generating an SDXL image it knows I can override the CLIP models in the safetensors file with the improved ones but if i try that same strategy for pony models the image just becomes garbage, bc there's only one clip model you can use with pony base and that's the one Astra made.

at one point I asked Astra to help me figure out how to create a remix of zeropoint's vit14-l fine tuned model but he wasn't interested, something like that would require retraining and access to the original training data

[bitter hearth]

lots of SDXL models train the text encoders though

[halcyon yarrow]

i didn't know that, bc from my experience, overriding the CLIP models has worked reliably for all SDXL models ive used, so while maybe they did train it, maybe it was light training that still allowed it to be compatible with the base model? maybe astra picked a different architecture for his clip L? not involved deep enough to know those nuances but there's something about it for sure

[dusky thistle]

gauss-legendre (implicit) 6th order

SDE mode

SD35L

[bitter hearth]

one of the most popular full-rank Dreambooth tutorials from Furkan involves training the text encoders, for example

astra just moved the weights more

the problem is the information that its a different model got spread around a lot

so now its hard to convince people that its just an SDXL finetune

[halcyon yarrow]

Compact Summary of CLIP L Architectures
ViT-L/14 and ViT-L/16
Vision Transformers with large architectures (14x14 or 16x16 patch sizes). Popular for their robust feature extraction in image-text tasks.

ResNet-50x64 and ResNet-101x64
ResNet-based backbones with wide layers for high-capacity feature extraction, used for detailed multimodal understanding.

ResNet + Transformer Hybrids
Combines ResNet for early-stage vision processing with Transformers for text alignment, offering balanced efficiency and accuracy.

Cross-Attention Enhanced ViT
Adds cross-attention layers for handling dense or long-text tasks, optimizing text-image alignment.

and then when I ask it what arch does SDXL use it says:

SDXL (Stable Diffusion XL) uses ViT-L/14 (Vision Transformer, Large, with 14x14 patch size) for its CLIP L-based text encoder and vision alignment.

and then according to Astra's article:

When training V5/V6, I used a CLIP-based classifier, eventually settling on the ViT-L/14 version of CLIP, which is the largest and last model released by OpenAI.
so there we go that put's that question to rest it's using the same base architecture, i guess you're right then astra just moved the weights around more

[halcyon yarrow]

dude i did a git pull on your stuff the other day after I updated my ComfyUI to the latest version to work on mochi... i was getting errors like there's no more sampler rk, you deleted like 2 other nodes, i haven't gone back in to see what the extent of the damage is but you did some major overhauls at some point lately huh?

i was just trying to get that snowshark to work

[dusky thistle]

snowshark lol

yeah it's a work in progress, tons of shit has moved forward

[halcyon yarrow]

my snowshark buddy never got to be rendered, remember i was having that all black issue? so i updated my comfy, reran iti and still all black, then i did a git pull on your stuff and then it said i was missing 3 nodes

[dusky thistle]

id' be willing to bet that git pull said there were over 1000 changes lol

[halcyon yarrow]

im not sure i didn't go intio detail but it looked brief to me

anyways ii had to abandon your stuff for now bc ii'll have to revisit your changes and realign my code to your nodes and it's a lot of maintenance for a moving target like what you're doing so i just went back to shitty ksampler ways 😦

When loading the graph, the following node types were not found

SamplerRK
SD35L_TimestepPatcher
ConditioningZeroAndTruncate
Set Precision Universal
Sigmas Rescale
SharkSampler
ClownSampler

lol like you blew the whole thing out

[bitter hearth]

DPM++ 2SA can get fairly similar results if you up the Eta and S_noise a lot

its worse but not catastrophically worse

[halcyon yarrow]

@dusky thistle if you could share an updated workflow for mr snow shark wiith the latest changes you did that would help out a lot re-integrate your samplers into my stuff

[dusky thistle]

something is screwed up if it didn't find any of those

it must not be loading

[halcyon yarrow]

oh i see so at least you didn't delete those nodes that's good

you didn't delete any of those?

that's the first thing i looked at too, the startup logs to ensure it was indeed loading, iill restart it real quick and give it a second look

0.1 seconds: E:\ComfyUI_PlusPlus\ComfyUI\custom_nodes\RES4LYF yeah seems to be loading fine

oh weird the nodes are loading now, wonder what that was.... anyways new error @dusky thistle

Prompt outputs failed validation
SharkSampler:

• Return type mismatch between linked nodes: sigmas, LATENT != SIGMAS
• Return type mismatch between linked nodes: latent_image, SIGMAS != LATENT
• Value not in list: sampler_mode: '5' not in ['standard', 'unsample', 'resample']
  so what's the eqiuvialent to 5 now?

[dusky thistle]

yeah tons of shit changed

gonna get you a wf in a min

bugfix one sec

this is the mask i used

input images

for the guide and guide_inv

here's one with large

switc hto res_2m if you want it faster

[halcyon yarrow]

@dusky thistle do you have a WF that doesn't use any load image nodes? like just purely text 2 img w/o any masks or anythinig?

[halcyon yarrow]

alright i got some output on the shoes

[dusky thistle]

that is one weird prompt lol

[halcyon yarrow]

not sure if you remember this file (not a comfyui workflow) but that's basically what i have to do to get your stuff working again, just adjust that spec

[bitter hearth]

I haven't tried the unsampling yet

I'm really glad there is a good workflow for it now cos I tried several times to make an SDXL one over the last year LOL

[bitter hearth]

Flux definitely still needs SDE I went through a lot of generations at different Eta levels on different passes and the layouts are so much nicer with SDE

its most obvious when you go from Eta 0 to even Eta 0.05 there is often a big jump in quality

[dusky thistle]

def... think quality is usually at peak with eta around 0.25 for that model

finding with sd35 higher levels really help with mutations

0.5

[bitter hearth]

I liked the soft scaling a lot

it was good up to 10 but it gets too soft

I had a second pass with 0 eta which helped a lot

[dusky thistle]

yea im gonna add in the ability to schedule that stuff again sono

[bitter hearth]

around 0.25 feels right, I was often choosing between 0.2-0.4

[dusky thistle]

that'll be a second passthrough node so longclown doesn't get too long lol

[bitter hearth]

there's Blepping insane chain sampler for now, it can take a few clown nodes per pass

[bitter hearth]

it lets you change sampler within one pass so like

one sampler node for 0-10 steps

another for 10-20

another for 20-30 etc

not as good as an actual schedule of 30 floats

but better than nothing

[dusky thistle]

if it keeps it going in and out of all the noise scaling stuff it might be helpful

for that alone

depending on what yprecision you're at you're gonna lose something doing that over and over

the shit in comfy that is

i gotta say i really like medium

it's def not as smart as large, but it pumps out some stunning images

[bitter hearth]

I didn't like it on launch day but your images since have been great

[bitter hearth]

SD 3.5 does seem better than flux with colours

[dusky thistle]

definitely

the images are just more interesting

way more diversity

[bitter hearth]

I should try it some time

[bitter hearth]

I still like SD 1.5 a lot for the diversity

[dusky thistle]

WFs are embedded, all you gotta do is hit ctl enter to run these

[bitter hearth]

ok thanks

my favourite image models out of anything are sg-minority and its sequel paper MinorityPrompt

they focused on diversity as the main thing

[dusky thistle]

haven't seen those

creativity is def one of the top things i'm looking for

[bitter hearth]

even on the super boring scientific datasets like LSUN-Bedrooms and CelebA, sg-minority was able to make way more interesting images

makes it look more like amateur cellphone footage

CADS was still very competitive in their paper

if we can get CADS for flux it would be amazing

its just noise injection to conditioning vector, its weirdly simple

[dusky thistle]

https://github.com/asagi4/ComfyUI-CADS/blob/master/__init__.py this here?

that doesnt work for flux?

[bold scarab]

画一只小狗

[bitter hearth]

yea that's the one

cannot remember if I actually tried, I thought I did but maybe not

there is a second CADS in latentmegamodifier and that CADS doesn't work

[dusky thistle]

Here is the image you requested.

[bitter hearth]

even troll images are aesthetic now

[dusky thistle]

yeah if it's just noising the conditioning... it wouldn't be hard for me to adapt all my noise sampler stuff for that

[bitter hearth]

I started using https://huggingface.co/MiaoshouAI/Florence-2-large-PromptGen-v2.0 today its Florence 2 fine tune on civit prompts

was really good

[bitter hearth]

I do want to switch away from florence though its falling behind

VLMs are exploding in quality lately

[bold scarab]

画一只兔子要秋天的感觉而且要很多颜色的叶子的感觉

[dusky thistle]

Here is the image you requested.

guide image

[bold scarab]

画一只兔子在很多秋色和桨果而且很多品种的叶子

[bitter hearth]

oh I forgot to tell you the other big news

someone got flux working in Int4

and its over twice as fast on a 4090

[dusky thistle]

whoa

which means, of course, it's time to double the number of params lol

[bitter hearth]

they also got it working in FP4 which isn't faster now
but will be for next year's cards

[dusky thistle]

can't wait for the 5090

just hope they aren't too big to stuff two on the same board

using 46,000 cuda cores to generate a single image is my kinda style

[bitter hearth]

gonna be a lot of doorless cars driving around

[dusky thistle]

Here is the image you requested.

[bitter hearth]

I put the 8 step Flux turbo lora on my image and the quality went up rather than down 🤔

[bold scarab]

自创一个国旗而且是兔子国旗

[noble coyote]

SD3.5L Turbo Llama3.2

[noble coyote]

Olivio Sarkas does a double KSampler 8Step-Flux-Turbo-LoRA workflow ...https://youtu.be/jfbqlSaRIPI

[bitter hearth]

thanks its interesting

I agree with spiking up CFG or guidance in later passes

[noble coyote]

0 - 8 at Flux Guidance of 3

9 - 12 at Flux Guidance of 7

[bitter hearth]

I need to try 7 I have been doing 6.3

[noble coyote]

🥳

[short thicket]

Awesome! I'm gonna go check them out. The best part of working on these models is seeing what other people make with them. 😊

[sacred jewel]

ballz

[sacred jewel]

That last one is nice!!

[noble coyote]

Vrrrrm Vrrrrrm

[summer kelp]

^^ Always good to have extra fingers

[magic drum]

👍

[magic drum]

[noble coyote]

🥳

[dusky thistle]

these are all SD3.5L

[split bramble]

Nice, these are good ones.

[dusky thistle]

Using dormand-prince_13s SDE with a solid black image as a latent guide

[halcyon yarrow]

heyy @dusky thistle when you get a chance do you thiink you can offer me some guidance on what this error means:

nan_to_num(): argument 'input' (position 1) must be Tensor, not float

[dusky thistle]

would need to see the WF

is it something in a RES4LYF node?

[halcyon yarrow]

i like how you added steps inito the sampler iitself so no more need for that betascheduling node

i want to debug it further but i dont understand the natuure of the issue, checked the positive and model input of the sharksampler assuming positioni 1 was one of those two but that checks out

[spark quail]

yo @dusky thistle , havent been in the sd scene in a while so i havent kept up. last i read, sd3 uses rectified flow which is incompatible with sde samplers. im seeing that you're using it and others -- where's the disconnect in this coming from for me? is your workflow getting around whatever issue prevented RF to be used with SDE somehow or simply using sd3 without rectified flow, etc? from what i recall, it had to do with how ancestral samplers worked or something like that that didnt mesh well with the RF process.

good stuff btw, im seeing more and more people mention you here and there

[dusky thistle]

it turns out it's very much compatible with SDE, but the noise scaling has to be done very differently

it's a lot less tolerant of funky math than the previous non-RF models, it has to be dead on with controlling the variance

[spark quail]

ahh okay sweet

[dusky thistle]

i've got a bunch of noise scaling modes working now

[spark quail]

i def wanna experiment and play around in that case

[dusky thistle]

know all the parameters of what can and can't be done with it, i think

i think the results are def better than with just ODE

[spark quail]

do you have a log of what youve tried so far, what is and isnt good

probly just whatever you mentioned in here right

[dusky thistle]

mostly yeah

aside from that... just physical notes

[spark quail]

if you jot them down somewhere id love to go thru them as a read while on a lunch break or something. even things like what you wanna try, thoughts on something you have tried, etc

[dusky thistle]

oh they're pretty incoherent lol

[spark quail]

otherwise no worries, im probs going to go thru a bunch of stuff youve already tried so far on my own ventures lol

[dusky thistle]

it'd take forever to really tabulate all that

[bitter hearth]

as far as I can tell the one key thing is making a scaling function to map the s_noise value for each step to a noise amount that is the right size for that step, for that model

[dusky thistle]

the critical thing is calculating three values correctly

[halcyon yarrow]

@dusky thistle do you think you could help me out and figure out why I'm getting that error? I tried deleting and re-adding the nodes to that workflow just in case any bad value was set and iit's still generating the same error

[dusky thistle]

how much noise to add after each step, what to step down to, and how to scale the latent when adding noise

[dusky thistle]

ah, you're using pony, that might be the issue

see if it works with sd35

i haven't checked on the sdxl side of things in a while... i'll get to that at some point though

[halcyon yarrow]

have you tested it with flux?

I just trieid the same workflow I gave you that doesn't work and switiched it to sd3.5 and now I get a different error

CLIPTextEncode
'NoneType' object has no attribute 'float'

That's my fault the WF is built for pony so ii can't just swap out the model, makinig adjustments to just tsee if i can make it work

[dusky thistle]

here's a basic one that works

i'd start from this

[halcyon yarrow]

cool thanks that helps

[halcyon yarrow]

when you said that 'works' you mean it works for SD3 right? i was thinking you were giving me a WF that works for sdxl

[dusky thistle]

yep works for RF

[dusky thistle]

if you git pull this will work

fixed the bug just now

with sdxl that is

[halcyon yarrow]

great thanks, im going to also test sd1.5 and pony make sure it's all in the up and up ill let you know if anything else comes up

[halcyon yarrow]

btw iis the new denoise parameter in the shark sampler behave the same way as the one in ksampler?

[dusky thistle]

yep

[halcyon yarrow]

That’s way better than having to use that rescale node and having to code in for 14.5 or whatever that number was

[bitter hearth]

haha 14.7 yeah

the number gets familiar

[halcyon yarrow]

@dusky thistle I noticed the error with SDXL comes with the truncate conditioning field, I'm going to set it to always false, is there a good reason I should ever set it to true?

[dusky thistle]

it can help a lot with images going to shit in SD35

if the prompts are > 72 tokens in SD35, sometimes they really go to crap

[halcyon yarrow]

interesting so i mean i could in theory leave it false and then just manually truncate my prompts to 72 tokens... i thought it was 77 not 72?

i'm seein eta of 0.5 and res_3s doesn't play well with pony models

res_3s with eta of 0 also fails, im going to try res_2m wiith eta of 0 since i know that used to work great.....

still failing, mayybe it doesn't like the scheduler as linear quadratic? gonna try normal and karras see if any of those two fix it.... yep that was it

@dusky thistle if i had to pick between other more compatible schedulers whiich one would you pick for higher quality and/or higher compatibility karras vs normal for res_3s?

btw just a little side fun fact I made a little checkbox for your special sampler, so when I generate an image using ksampler, or sharksampler and I see I don't like the way it came out I can press a button and I get this popup. the first button deletes the image that came out in case it's real bad, the second box will toggle between new workflow (sharksampler or box toggled off) and old workflow (ksampler or box toggled on) and then I can select how many images I want to geenrate if I want to change the batch size for it

[dusky thistle]

if you're using sdxl i recommend karras or exponential

or sgm_uinform

[halcyon yarrow]

which one is the best of those 3?

best in terms of quality regardless of speed

or wait how about this, what scheduler is the best one to pick for something that'll work across all models really well? that way I don't have to dynamically change it per model just stick to one and let it do it's thing, that means one scheduler that's compatibile and works great for sd15, sdxl, pony, sd3, flux

I'm trying 'normal' for all 5 base models and it looks great for all except Flux that generated garbage

[dusky thistle]

it's really a good idea to use different ones for RF vs sdxl etc

what's best will depend on what you're generating, you'll have to play with it

beta and linear quad are really good with RF

not so much with sdxl

[halcyon yarrow]

right but i don't want to introduuce complexiity if not needed, are you saying there isn't a scheduler that works across all models?

tell you what i'll just try 'beta' on sd15,sdxl,pony and if it works i'll just use that across all 5 models, now is there an rk_type that works really well with beta or is res_3s the recommended choice for beta?

hard to believe this is sd1.5 sometimes, this is using beta, the limbs are a little wonky

[bitter hearth]

if you use shift then beta can work across all models

of course this isn't actually using beta in all cases because of the shift but for the GUI it is

[halcyon yarrow]

i thought shift only applied to sd3/flux

[bitter hearth]

yeah so you take the shift off when you use SD 1.5 and SDXL etc
but you can keep the beta schedule

[halcyon yarrow]

i used beta across all models w/o shift and it worked fine btw

[halcyon yarrow]

yeah that's what i did exactly

so my final global settings for all 5 models (sd1.5, sdxl, pony, sd3.5, flux)
rk_type - res_3s
scheduler - beta
noise_mode - hard
eta - 0.5
eta_var - 0

and i think that's pretty much the major settings that actually matter afaik

[bitter hearth]

would strongly advise using a high shift, or something equivalent, for SD3.5 and flux

you don't have to but the difference is big

the reason is that Flux decides the layout of the image by the time it hits sigma 0.9-0.8 (it goes from 1 to 0)

look at this example, you can see the first pass is almost done in terms of layout

but that pass ended at sigma 0.8 lol

so you just need to make sure the model has a decent amount of steps in between sigma 1 and 0.8, or even sigma 1 and 0.9

see in this image its 40 steps in that range

I often plot people's flux images and there's like 2 steps in that range

so the model has only 2 steps to do all of the layout which is pretty rough for the model

[halcyon yarrow]

i'm open to try it, let me reviiew my current settings....

for flux Im using ModelSamplingFluxNode and the settings are ```json
"max_shift": 1.15,
"base_shift": 0.5,

yikes and this is a good time to realizie im not using the exponential shift one fo sd3

is there an all in one I can use for both or do they both require their own node?

for SD3.5 I could use that SD35L_TimePatcher node and just set it to exponential / 3.0 like how Clown does it. Do let me know if there's a better way

[craggy crest]

clip_g gets a full 77 tokens. clip_l and t5xxl share 77 tokens.

[dusky thistle]

that's what i had thought would be the cutoff

what's weird though is that once you hit 73 tokens, truncating the conditioning embed changes the output

i haven't had a chance to look into wtf is going on

all i can say for sure is 72 or less isn't gonna go over whatever soft limit there might be (which i think is primarily an issue with insufficient training on longer embeds, so it's also gonna vary on what you're prompting for if that's the case)

[craggy crest]

it's pretty clear in the actual network diagram

[dusky thistle]

it is

lol idk why it's the case, but it is

it may be a bug

idk

all i did was truncate the embed to the same length as one from an empty prompt

72 tokens, truncated/non truncated are the same, 73 they're not

havent looked into it any further

[craggy crest]

might be. if you track it down, and it turns out it is, please let @lavish osprey know what you found

[halcyon yarrow]

i do remember reading that longcliip study article and it was talkin about how the 77 tokens is actually 20 effective tokens

https://arxiv.org/abs/2403.15378

Despite its widespread adoption, a significant limitation of CLIP lies in the inadequate length of text input. The length of the text token is restricted to 77, and an empirical study shows the actual effective length is even less than 20.

[craggy crest]

that was published back in march, though - i wonder if any advances have been made since then

[halcyon yarrow]

to be fair it was revised in July that was only 4 months ago

[Submitted on 22 Mar 2024 (v1), last revised 22 Jul 2024 (this version, v3)]
I think iit's interesting how clip L only has 20 effective tokens that's why im a big fan of longclip that extends it to 248 effective tokens

[craggy crest]

i'm curious if they released any papers on the actual long_clip research they proposed in that one

[mortal mesa]

https://arxiv.org/abs/2410.10034v1

weird world where papers are referencing reddit posts

[craggy crest]

ROFL!!!! oh my

@dusky thistle @bitter hearth https://github.com/kohya-ss/sd-scripts/pull/1768 from Dango (lead programmer for sd3.5 medium) https://github.com/kohya-ss/sd-scripts/pull/1768

[halcyon yarrow]

from the conclusion:

5 Conclusion
We have propose Long-CLIP, a strong and flexible CLIP model with long-text
capability. Our model can support text inputs of up to 248 tokens, and can better
capture the detailed attributes, obtaining a large improvement on retrieval task.
Moreover, our model keeps the performance on zero-shot classification and can
replace the CLIP encoder in a plug-and-play manner in image generation task.

They also have this table in the PDF that shows the effective token count, I thnk it lends to what @dusky thistle was seeing where he's setting it to 72 rather than 77, in some instances i see values like 69,70,71

@bitter hearth you never told me what you recommend for the shift settings that you were talknig about

[bitter hearth]

what I like to do is have one pass to set the layout and a second pass to finish

then if the layout is too poor you can just raise steps in the first ksampler

or the opposite, sometimes

[craggy crest]

take a look at Dango's timestep fix code i pinged you on above

here are some fun SD scripts https://github.com/kohya-ss/sd-scripts

[bitter hearth]

LOL yeah same topic

[craggy crest]

did you look at his code?

[bitter hearth]

no I'm not really into scripts like koyha or simple tuner

been working on some personal pytorch scripts since like february

[craggy crest]

at least read through the comments

[bitter hearth]

okay will take a look thanks

[bitter hearth]

wow that's some nice looking anime

would be cool if more anime generations were that sort of thing

[halcyon yarrow]

yeah i love the shading style on that

[halcyon yarrow]

simply put I'm just askinig what your default/starting values are that you recommend before I adjust anything

for flux i have max shift of 1.15 and base shift of 0.5 and for sd3 i have exponential 3 shift

[bitter hearth]

if you want a default I would default to the comfy default, which some from the BFL reference code

if I remember rightly its 1.15 and 0.5 as you say

not sure for sd 3.5

[halcyon yarrow]

this started bc you said and i quote would strongly advise using a high shift, or something equivalent, for SD3.5 and flux so thats what I'm asking. what is a high shift value i can use as a default for both nodes?

[bitter hearth]

oh I see, I would say 1.5, for flux

for both base shift and max shift

is a pretty reliable high shift

[halcyon yarrow]

wow interesting 1.5 for both base and max? ok ill try it

[craggy crest]

i 100% agree. i'm really tired of all the anime being some version of "girl with large chunks of meat on the top that are so heavy she should have major back problems"

[bitter hearth]

my actual overall recommendation is to do custom sigmas but it seems unpopular to do custom sigmas

I know very little about anime but it probably has some very interesting artstyles yeah

[halcyon yarrow]

personally ii'm really sick of Flux's default anime style, its so off putting now, the way it makes anatomatically correct heads with these weird large eyes

[craggy crest]

shift 3 for sd3.5

[bitter hearth]

ah ok thanks

[halcyon yarrow]

alright so exponential shift 3 is the sweet spot for sd3.5 and 1.5 for base/max for flux, ill try it

[craggy crest]

yes, well, that's what happens when you scrape midjourney and stuff all the images into your model to mask the probelms it has

[bitter hearth]

I did a lot of image stuff in the last few days but it was all in the new Itercomp thing so there were no shift shenanigans

[craggy crest]

for sd.35, yes. for flux, usually i wind up with 2, but sometimes 3

[bitter hearth]

I found something funny
if you stack enough DPO loras on SDXL it starts to look extremely flux-y

[craggy crest]

i did tell you that flux is a huge lora, right? what all have I been saying about flux, and being argued with by everyone, all this time?

[bitter hearth]

I never quite understand what you mean when you say its a lora

[craggy crest]

someone, FINALLY, earlier on reddit admitted that when they try to do images of people laying on the ground with flux it has the same issue that sd-2b-medium had with warping the body

[craggy crest]

it can only do specific things in a very tight range. just like a lora only has a very limited, tight range, of information

[bitter hearth]

oh in that case I totally agree

[craggy crest]

it's massively overfit for women, dogs, fantasy images, and anime catgirls. it's so overfit that's all it can do with out a huge battle. and after it was stuffed full of that to mask the warping issues that the core network causes (rather than fixing it right) then they not only distilled it, they dpo'd it

it's a great tool, if that's what you want.

if you want something else though, it will put on armor, pull out lasers, and battle you to the death

[bitter hearth]

ok I think I understand, when you are saying its a lora and its frozen you are saying its overfit
I agree with that yeah

[craggy crest]

no, i'm saying it's overfit when i say it's overfit. without all that stuffing it would probably be about an 8billion parameter model. it's 12 billion

[halcyon yarrow]

hey @bitter hearth do you know of all the rk_types in the sampler nodes which one is the fastest?

[bitter hearth]

Res_2m with iter=0

[halcyon yarrow]

where's the iter field? ii dont see it in shark or clown sampler nodes

[mortal mesa]

All models are wrong, but some are useful

[bitter hearth]

the names keep changing so its hard to know
might be implicit steps

[craggy crest]

might ping shark and ask him what he did with it

[bitter hearth]

implicit steps I think

[halcyon yarrow]

got it, yeah that's step to 0, so res_2m is the fastest ii'll try that, using res_3s and while the output quality is superb my rendering times are now like 50-100 seconds for pony images whereas before it used to be 15-40 seconds

[bitter hearth]

to save you the time, I tried a stupidly high amount once already and its not worth it

was like 1000-2000 steps with implicit steps on 8

so over 10,000 function calls

really not worth it

[craggy crest]

you do like to torture your machine, don't you?

[bitter hearth]

yeah its the fun part 😄

I did a big test this week of flux with CFG or without CFG and I preferred CFG 🤷‍♂️
although I can understand why people say it breaks the model it loses some of its base style

[craggy crest]

here, run this: sampler: deis, scheduler: linear_quadratic, cfg: 5 Steps: 35, Prompt: various types of fruit arranged in a crystal bowl sitting on a table. next to it are scattered flower petals and leaves. the background is a spherical goldishgradient. back lit, hard-rim lighting

use sd3.5 medium

and once you run that, then turn on the SkipLayerGuidanceSD3 node and skip layers: 22,23,24 scale: 1

and run that

[bitter hearth]

ok thanks will try tomorrow
Deis is really really good I am a big fan of it

[craggy crest]

:) i like all of them, i just turn them into pretzels till they do what i want

[bitter hearth]

I want to make my own set I've just been procrastinating it

there's a vast amount of things you can choose from to put in a sampler

[craggy crest]

too many projects, too many toys

[bitter hearth]

yeah for sure and this is just one small area of ML

for me to do yet more samplers wouldn't be that useful I guess as there are so so many
working on some FP4 kernels for Triton and Tensor RT would actually be useful but very tricky and dry

RTX 5090 is gonna get FP4 speed boost

[craggy crest]

not real sure about that. which area of ML are you most interested in?

[bitter hearth]

mostly statistical models but out of the more fun stuff, LLM agents

although its a bit early for proper agents there are some janky versions out there

you can do a fair bit in comfyui even if you chain Ollama nodes and vision LLM nodes

[craggy crest]

then there might not be nearly enough samplers, or schedulers - you're not talking about compvis

[bitter hearth]

yeah that's true

the reason I think a focus on Triton and Tensor RT is good is that it speeds up every model you run

whether its image or LLM

quite a few companies, like Fal, for example, their competitive advantage is things like hand-written kernels for things like Triton and Tensor RT

[halcyon yarrow]

@bitter hearth can i speed up image gen on my 4070 with any tensor rt tech for comfyuiii?

[craggy crest]

@bitter hearth https://youtu.be/43U_ADJmCoc?si=HCXTz6TaX4_b2OZ3

[bitter hearth]

thats good entertainment, i like the female cyborg

but i wouldn't agree with anyone who claims ai can ever achieve consciousness

it can push the limits of neural activity but not consciousness

Ai will be the most sophisticated apparatus within the physical domain of reality down to the very acute level of quarks and stuff, and as far as mind body interactions can go

[toxic bone]

garland is a crazy man with science fiction. devs is a good show too, about a company developing a denoising algorithm

[bitter hearth]

humans will most likely thoroughly enjoy loving AI female cyborgs intimately, but very well knowing they are void of human emotions, they can translate your emotions very well but unable to experience it on their own

[toxic bone]

woman will get on board the first. they already are in hand held form

[bitter hearth]

you talking about toys?

its impossible to feel the same level of connection with cyborgs as you will with real human girl

[toxic bone]

pff. over rated. most guys are going to go for the easy stroke.

[bitter hearth]

sure but still that's not the same degree of experience

[toxic bone]

the ai will be like "you're so great you're the best" and that's all they want

[bitter hearth]

there is soemthing deeply related to mental connection with real human girl that cyborg can't offer

[toxic bone]

the kids growing up with all the easy access will lower standards significantly

[bitter hearth]

loweirng cognitive experience is not a scope

you will have a broader scope to augment

[toxic bone]

the shift has already happened once before. we were an animal husbandry race of people. the bond with your horse was sacred.

cars are just better at it. measured in literal horsepower per car

you still can find horse girls though

[bitter hearth]

not the same

even if you give examples, which btw is real, who will enjoy and prefer that type of expereience, but thats still not the same

[toxic bone]

first gen sex bots for sure will be like the tbucket.

then the enzo ferrari of sex bots will enter the stage

[bitter hearth]

you know im very sure there are people who are very much driven by materialistic aspect of their lives, so i know what you mean

but that doesn't objectively imply they have as much richer experience as someone with high cognitive sensitivity

[toxic bone]

yeah but wait until you see ludicrous mode

[bitter hearth]

ludicrous mode?

[toxic bone]

whatever the ludicrous speed equivalent is

[bitter hearth]

never heard of such attribute

[toxic bone]

https://tenor.com/view/speed-gif-5593959

[bitter hearth]

ahh ok so thats easily possible when you understand that AI can have the collective brain capacity of all humans

[toxic bone]

to reference space balls, tesla updated the cars with a new launch mode that was faster than insane launch mode. called it ludicrous

ludicrous launch is when it goes 0-60 in /

[bitter hearth]

that can happen by simply increasing bandwidth for data

theoretically what LLM of today is capable of compared to human intelligence you can enhance that with more memory capacity to compute

[toxic bone]

i've been trying some LLMs lately. it's better to get models that you need to jailbreak, because models that are pre jailbreak are all sexbots basically

it's a very popular scene

[bitter hearth]

censorship sucks

i wish we had llm bots that are open to explicit disucssions as long as they dont impose any harm on anyone

but what we have now with explicit content is some nonsense moral standards of humans

[toxic bone]

yeah. but you see, you're talking about safety. but that's what they're trying to do and it's categorized as censorship. there are blurred lines that are difficult to define here

[bitter hearth]

dont censor that part of human primal desire

[toxic bone]

what i know is that the decesnored models are basically destroyed for regular knowledge. they're just aol roleplayers

[bitter hearth]

or they will create another sd 3

i havent come across any llm that can discuss intimate topics openly

they should at least allow the kind of interchange you would expect between couples

[toxic bone]

anything that the sillytavern community talk about in their weekly thread is a very "intimate" orientated model

[bitter hearth]

but you know .. there is a risk involved

people can get suicidal not cause ai can be so intimate but cause of underlying societal scenarios

[toxic bone]

i downloaded one of the llama 3.1 models thinking they wouldn't have been able to turn those into a very intimate one, but it's honestly all the mdoel wants to talk about. If i use one of these jailbroken ones instead of the base model and giving ti a jailbreak prompt. i'm trying to find a model that wont give safety warnings all the time, but also won't bring up the ol push an shove at every opportunity

the roleplay crowd has got llm training figured out

[bitter hearth]

well its not a limitations of any kind from AI's part, but its human imposed safety net

[toxic bone]

yeah. an there's a lot of research into disabling the net post training. in better non destructive ways than what i've found

[bitter hearth]

the critical point of this whole thing is society's religious mindset to a great degree, this outdated fear based faith is hindering developments

society is not yet ready for polyamorous love cause they think its sinful and immoral

yet they will have you accept that love is unconditional

talk to a religious person about jesus cherry picking a girl

logically and as a human being with principle its not a difficult idea to love all girls

[toxic bone]

not just that but a lot of taxation and property rights are woven into the idea of the atomic family

to the point that bigomy is tax evasion

[bitter hearth]

economy and human labor are intricately associated with it as well, and gives rise to the arbitrary concept of jealousy and cheating

unless there is mas scale automation to replace all human labor meaningfully and with greater efficiency, society will not overcome those mindset of jealosuy

here is the world we live in today, ... you only love one person, you are so moral and ideal, .... you love many people, you are a degenerate

[bitter hearth]

and if you promote polyamory w/o removing those other factors of human labor, economic servitude you would actually push them towards unhealthy bonding

[harsh bison]

draw a paperwall with a hello kitty

[dusky thistle]

https://github.com/ClownsharkBatwing/RES4LYF/blob/main/README.md

finally got things pretty close to being stable and able to be forked into a dev and stable branch

lots of nice simple workflows up there with input images included

[untold valley]

the prompt

[dusky thistle]

that looks considerably more pitiful/desperate hahahah

[sage burrow]

How's medium vs large?

[dusky thistle]

different

both are great

i'm usually generating most of my stuff with medium lately

[sage burrow]

thank you 🙂 I'll definitely have to try both. I'll also try to create loras for both and see which works better. So far my large lora attempt is not turning out as well as my flux ones <cries>

[dusky thistle]

You might wanna check out RES4LYF again, big update tonight, it's getting really good results with RF and fast

Bunch of nice simple workflows on the repo now

Getting waaay easier to use lol

[sage burrow]

Your github page is awesome!!! I usually stick with civit, only because they actually show images and examples (whereas HF doesn't easily really)

[sage burrow]

3.5 medium

[untold valley]

We not using res-2 no mo?

[sage burrow]

sd3.5 (large) is 100% trainable for loras. flux is only about 70% trainable.

[marble lion]

Anyone know why we have not yet seen an explosion of SD3.5 checkpoints like Juggernaut etc? Too early? Are they in training or what?

[sage burrow]

not really many loras either; I figure people don't realize how trainable sd3.5 is after trying and having abysmal results with flux

[gusty trail]

Some in context lora creation

[frank breach]

hello everyone, I am attempting to run stable diffusion 2.1 via https://github.com/Stability-AI/stablediffusion. I did not find the version 3, and I have the BFloat Error, can I have some help, please ??

[turbid grotto]

I suggest you to run them via interface: forge webui or comfyui
There is plenty tutorials how to set them up, choose which you like, however, sd3.5 seems to not be supported in forge yet

[bitter hearth]

yeah the tensor rt nodes should work out of the box even

however vram limitations are there for your GPU

you may be able to do sd 1.5 but not flux, for example

for tensor rt

[halcyon yarrow]

i tried it last night actuallyy

the tensor rt loader actually has flux-dev and flux-s in the dropdown so it should work in theory

i actually tried it, i converted a flux model but i couldn't get it to work, kept ggetting an error

[bitter hearth]

fairly sure the flux one requires 80GB to make and then 40GB GPU to run

I've been using them on A/H100 I don't think they are for 24GB

for SD 1.5 you will be ok

[halcyon yarrow]

wdym? the files i was making were smaller than the input

[bitter hearth]

to make flux tensor rt engine?

[halcyon yarrow]

it was working fine on my 8gb vram setup for pony

[bitter hearth]

yeah for SDXL

[halcyon yarrow]

yeah i was able to make the files for sdxl/pony/flux just fine

[bitter hearth]

you made a tensor rt engine for flux on an 8gb vram GPU?

[halcyon yarrow]

iti was just figuring out the workflow to get it to run, kept getting a shape mismatch error

yes

[bitter hearth]

hmm okay thanks I read this requires super high vram

[halcyon yarrow]

the STOIQ model

anyways its a fun idea and its quaint but its not practicall

the lack of lora support is kind of a deal breaker for me

however if i can get it working for flux Ii'd love to convert a few of the better models that don't always need loras to this tensor rt format and see how much faster it is. considering some images take 15 minutes to render I'd love any knid of speedup i can get

[bitter hearth]

it does support loras though

you just merge them first

[frank breach]

I have a supercomputer with 6 nodes 128Gb of RAM, and 6 PNY RTX 4000 ADA. Do you do distributed run ?? Frankly my objective is the intregration with my own RAG to build 3D object and 3D caracter into Unity. LocalAI work, but right now because there is no easy distribution I am limited to 20 GRAM per node.

[bitter hearth]

swarmUI is good for distributed inference

[halcyon yarrow]

yeah that's not practical though

notce how i made 3 variations of the PVC model, i was playing with merging loras into the model to verify that indeed does work

[bitter hearth]

its annoying to have to build the engine yeah

in some ways FP8 matmul plus torch.compile is a better compromise

especially since the latest pytorch speeds up torch.compile

[halcyon yarrow]

lack of dynamic lora support let's say makes it a deal breaker for any sdxl/pony model i'd use but it's still a viable solution especially for SD3 where there's no loras and Flux where some modells dont really need any loras

i was going to investigate it further and find a WF that works with flux + tensorrt see if i can get it to render something in comfy

[bitter hearth]

have you managed to run the flux tensorrt engine?
or just build it

I'm just confused cos here and on comfy server people were saying 4090s could not do it festivalman — 10/10/2024 19:42 how did you covert flux to be tensorRT? Every time I've tried it, it says OOM on my 4090.

[frank breach]

did you test the image to 3d verqsion ?? I have instantmesh that work, but it is taking all the RAM

[bitter hearth]

I don't think 3D ML models are very good yet

I am not sure they are ready for use

[frank breach]

Frankly, no, they are not ready, the face of a 3D caracter from LocalAI does not appear, but I still need to implement it for my Game

[bitter hearth]

okay I'm afraid I don't have experience with the 3D ML models as I have ignored them so far

[frank breach]

Thanks 😄

[bitter hearth]

I would ask in blender discord and UE5 discord if I were you

[frank breach]

Yeah, they are way more advanced then Unity

But I started coding with UNity

😄

[halcyon yarrow]

weird bc i converted it just fine on my 4070. tell you what ill convert it againi just to make sure i have a solid base

[bitter hearth]

I like Unity a lot, its still good

[halcyon yarrow]

now that you mention it ii do remember tryng to convert flux dev destilled and that one did give me OOM issues

[bitter hearth]

maybe I wasted money on 80GB server for this lol

[halcyon yarrow]

id use runpod

wouldn't cost me more than a dollar to rend 80gb for a few minutes

[bitter hearth]

yeah I use sites like that

$0.70 or so for an A100 80GB is my favourite

but sometimes $0.80 for L40s 40GB

[frank breach]

Frankly NeonNinjaAstroo, if you invested into a A100 or something, you are not the only one to think like this, so don't worry too much

[halcyon yarrow]

i never rented any machine that large, I'm a big fan of the 3090 and the A1000

[bitter hearth]

3090 is good one to rent still yeah

[halcyon yarrow]

my biggest spend was renting multiple A1000s like 4 or 5x concurrently and have them rendering for me pony/sdxl images while they all get streamed back to mym main server in real time, it was quiet the sight lol

[frank breach]

0,70$, where do you find these price ??

[bitter hearth]

I see 3090 for $0.30 often

[halcyon yarrow]

yeah in terms of value, unless you need more vram that 3090 on the community cloud is the sweet spot

[bitter hearth]

$0.70 was vast.ai interruptible non-data center

[halcyon yarrow]

ii usually pay 12 cents for the A1000

thats 16gb for 12 cents, at most i pay 14 cents

[bitter hearth]

the reason I avoid the smaller ones is I had too much trouble with them

[halcyon yarrow]

I have it set up so I can use the spot instances so if someone kicks me out then i have a script that attempts to outbid them in an effort to regain control of the machine and then it'll keep retrying upto a point where it makes financial sense mathemtically and then it'll kill the intance and stop trying lol

[bitter hearth]

haha

[halcyon yarrow]

12 to 14 cents is only attainable if you're willing to do spot instances though, like if that make ssense for your use case

[bitter hearth]

I found the interruptible ones were not always cheaper, which is weird

[halcyon yarrow]

ive never seen that, for me its always cheaper

[bitter hearth]

this was on vast I think

bare in mind I use the verified data center ones not the community ones, cos of security

[halcyon yarrow]

i dont care i just want the cheapest, security is of no consequence to me lol

i have found that there's like a mini scam on runpod not sure if you've picked up on it

[bitter hearth]

what's that

[halcyon yarrow]

sometimes i'll get a machine assigned (like 1 out of 20) that has a broken GPU

liike some sort of mismatch or invalid state the GPU is in with the dockerized instance

# Step 0: Check if CUDA is available before proceeding
if ! python -c "import torch; print(torch.cuda.is_available())" | grep -q "True"; then
    echo "CUDA is not available. Proceeding with termination process."

I had to put this at the top of my deploy script so I can check that first and if the GPU is busted on it I just terminate the instance

[bitter hearth]

there are some weird instances out there

[halcyon yarrow]

it feels like a scam bc i can't report it, there's no mods or admin or contact person, so they get to just go back on the market and keep eating people's money with no consequence

[bitter hearth]

there's loads of those around yeah

not necessarily a deliberate scam just a mistake

[halcyon yarrow]

yeah i can imagine that, like the dude just set it up, something broken and the dude still sees money coming in so as far as hes concerned it's still 'working'

[bitter hearth]

there are datacenter ones like that too

the providers like runpod, vast and salad don't do as many checks as they make out

[halcyon yarrow]

that sucks, so ill stick to runpod then, ive already got tons of code written to integrate witih them, really enjoy their graphql API and other hooks they provide, plus ive looked around and you cant beat runpod's prices

even with like a sign up discount from other providers like massed, even affter a bonus credit or whatever they give you, its still more expensive than runpod

here's the tensor rt workflow i used to make the working pony models

[bitter hearth]

I don't actually know why people use anything other than vast, given that vast is much cheaper than any of the others

the reliability is around the same across all of these

[noble coyote]

Anybody rate Flux Colossus?

[bitter hearth]

so far haven't seen a flux checkpoint that had better image quality than dev

[halcyon yarrow]

vast is cheaper than runpod?

what's the cheapest 16gb ii can get on vast for?

[halcyon yarrow]

you mean you dont think any finetunes are better than the base model?

[bitter hearth]

for flux yeah I think that

there are loras that refocus the model on a particular style or subject
but I haven't seen one that actually raises the overall image quality
whereas with SDXL or SD 1.5, Realvis or Jugger are way higher image quality than SD base

[halcyon yarrow]

I think there's a lot of good finetunes, this is the set I use, stuff like Fluximate, Pixelwave and STOIQ really stand out in terms of being better than the base model, that flux.dev dedistilled is in another league of it's own, not only is better than base model but it's almost on par with Pro imo

[pseudo owl]

i think flux already had lots of high quality aesthetic tuning so its hard to improve performance, except in artistic styles.

Its actually the same with modern open source llms(llama3, gemma2, mistral), there is not really any general finetune that improves performance.

[bitter hearth]

I've tried all of those except Raemu
I'm not rly convinced

they are good because flux base is good but I am not sure they are better

[halcyon yarrow]

raemuu is alright I'm not going to sing it's praises but it does not dissapointi either

if you want somethinig demonstrably better than flux-dev in every way, in every comparison test, try dedistilled like i said just scientifically, empirically that model is better than base model without a shadow of a doubt

[bitter hearth]

okay thanks will give it a go

I use CFG anyway on every flux model 😄

[halcyon yarrow]

like iif i have a complex prompt and I dont want to mess around, i just want it right on the first shot, and I dont mind waiting foor the 60 steps to finish ill puull out that distilled model

i use ksampler on every model too, I just tweak it to 1 to 1.8 for distilled models and give it free range for dedistilled models (like mangled, fluxbooru and flux-dev-distill)

[bitter hearth]

ah yeah that would work

with some tonemap, threshold or skimmed cfg can go a bit higher

[halcyon yarrow]

to follow up with our earlier chat I managed to convert the SDXL version of the STOOIQ model not the flux version, the flux version is in the unet folder nott in my chekpoints folder so thats why i got confused, bottom lne iis you're right i don't think i have the memory to convert it

[halcyon yarrow]

i just rescale 3 to 10 to 1 to 1.8 and it's been working pretty good

[bitter hearth]

rescale CFG is fine yeah it looks kinda similar to tonemap

[halcyon yarrow]

jst tried convertinig sd3.5 to tensort and that doesn't work either, and thinking iit through the idea of using a rented machine wont work bc the engine is built on the GPU it's made with so it's not like I could go into a rented H100 and pick 4070 from a dropdown when buildnig the sd3.5 or flux model

[bitter hearth]

I am not sure now
a lot of people were saying it requires high VRAM but I am not sure where the line is

[halcyon yarrow]

yeah i think they were right, again my earlieir test of was flawed, it was just an sdxl model, i'm getting OOM when converting flux or sd3.5

but using a different machine other than yours would generate an engine that's not compatible with your GPU

[bitter hearth]

but using a different machine other than yours would generate an engine that's not compatible with your GPUyeah this is an issue

I don't know the solution to home users

other than potentially to get an RTX 5090 or A6000 40GB at home
which is expensive

[halcyon yarrow]

lol i'm on a laptop so i couldn't just get a nicer card, I'd have to get a enclosure or a full pc

[bitter hearth]

I'm also using laptop yeah

but one without a GPU even

I have a PC in some state of repair

[sage burrow]

Ultra? What's that?

[bitter hearth]

big Flux Pro

its like 2048*2048

really really far ahead of every other model

[halcyon yarrow]

hey @sage burrow nice to see you're back, you were gone for a while, life am i right?

[fast crane]

Do sd 3.5 large loras work on medium?

[halcyon yarrow]

i dn't think so, i tried it myself bc i wasn't gonna take no for answer despite it being common sense, they're just different base weights so they can't be compatiible but comfyuii was complaining about shape mismatch so there's probably more to it than just weight alignment

[fast crane]

Okay, thanks

[noble coyote]

Ukiyo-e painting with Gustav Klimt influences. Beautiful fairytale princess discovers a big shiny, golden compass in dense, dark forest, gnarly trees, lush green vines and colorful flowers, full of magic and mystery. Dwarves watch in amazement. Sparkling light floats in the air, adding a sense of mystery and fantasy.

[sage burrow]

Still hiking season lol

[sage burrow]

I just made a medium lora which worked on large GGUF, I think lol

[halcyon yarrow]

If you have a link I could try it and confirm

[bitter hearth]

flux dev loras work on schnell but 3.5 large loras don't work on medium

[bitter hearth]

Flux + LoRas

[halcyon yarrow]

@dusky thistle your stuff is almost like a bitter sweet victory getting it integrated bc it's cool that i have available this high quality sampler that's much better at not generating artifacts but itis like 10x slower than ksampler for sdxl images and like 9x slower for flux. that is a flux image typically takes me 80 to 120 seconds now averaginig 1000 seconds, sdxl typically takes 15 to 40 now averaging 120-140. that's way too much

just as a little side fun fact this image took me 40 minutes to generate using flux dev dedistilled + sharksampler @ 60 steps

if i were to make a suggegtion as to the next set of changes for clownshark/sharksampler is to focus on an rk_type that's fast above all else no matter what sacrifice in quality, try to get something that'll render in the same time or within 1-3 seconds of what ksampler can do

[dusky thistle]

depends what settings you're using, lol

it's just as fast as euler in ksampler if you use res_2m

[halcyon yarrow]

i tried that

[dusky thistle]

when did you last update?

[halcyon yarrow]

i'm using res_3s and then I think NeonNinja told me res_2m should be just as fast but i personally didn't observe any performance improvements, imo it seemed just as slow at 10x

[dusky thistle]

i've reworked shit bigtime in the last 24 hours

yea that doesn't make any sense

res_2m is def just as fast as euler

tested it a few hours ago again

[halcyon yarrow]

i can try it again for sure maybe my tests were flawed or inconclusive

im afraid to do a git pull

i'd hate to deal with any breaking changes again

did you remove any nodes, add any fields or otherwise changed any paradigms?

[dusky thistle]

yeah it's way better

check the github page

there's workflows on the main page now that are up to date

you can see how much cleaner it is

[halcyon yarrow]

lol oh no so you did change the workflows?

[dusky thistle]

yeah i'm finally at a version of clownsharksampler i can keep

[halcyon yarrow]

lol no way dude i'm staying away ill just try res_2m for now and be happy, last time i had to spend like 4 hours re-integrating all your stuff back in

im not ready to commit a few hours to figure out what changed right now

[dusky thistle]

look how nice and simple that is now

resizes all your images and latents and masks for you

converts everything to latents and masks internally

no need for a clown and shark node anymore

[halcyon yarrow]

yeah i see you added shift and base shift internally

[dusky thistle]

and no need to deal with zero out nodes

or even negative conditioning at all it'll do it for you

[halcyon yarrow]

ive got 1 more image waiting to render and then ill switch to sdxl and res_2m and give that a shot and confirm if it's within ksampler's times. personally i diidn't mind the way it worked i just set it to the truncate to false and i can just leave negative conditioning to blank for flux or actually use it for sdxl

[dusky thistle]

this is what txt2img looks like now

simple as f

actually a simpler WF than using ksampler

[halcyon yarrow]

oh no, so you got riid of clownsampler and the 'sampler' pipelinie that's gonna be a mess to recode 🤦‍♂️

[bitter hearth]

in the new comfy UI how do you make an image lol

can't find generate button

[halcyon yarrow]

i've no clue i think mines is on a specific branch so it doesn't go to the new UI

bc whenever ii do a git pull im still stuck in the legacy UI

@dusky thistle 212 seconds for the first image using 6 loras, 107 seconds for the second image using 1x lora, same model so the weights already loaded, that's a far cry from 15 to 40 seconds using ksampler, like i said even res_2m is 10x slower than ksampler. I think you should really take some time to focus on optimizations for the next commit with the goal of really being on par wth ksampler's speed

[dusky thistle]

you should get to the current version first

[halcyon yarrow]

88 seconds on the third image, 89 seconds on the 4th image, if anythinig that's 2x slower considering some images take 40 seconds somettimes, i could do a side by side comparison of like exactly the same generation with ksampler and with your sampler but i think the point is that for sure there's no doubt the speed iisn't on par with ksampler

[dusky thistle]

whatever speed issues you may have, may already be fixed

i'm not seeing any difference in speed, nor is anyone else that i'm aware of

[halcyon yarrow]

so with the latest version you already did a side by side compariison, with ksampler and with your new WF and benchmarked the times?

[dusky thistle]

12 seconds with euler in ksampler, 13 seconds with clownsharKsampler

[halcyon yarrow]

alright that motivates me to do yet another reintegration, i just dont have the time for it right this moment but iill def try it in a few

cool stuff man keep up the good work

[dusky thistle]

it's possible your shit slows down a bit from having to generate the noise after each step, idk

i just tested with noise generation off and it's actually a tiny bit faster over a couple runs for whatever reason, prolly random luck

yea just did a bunch of runs here with SD35M, zero difference, the gap is random

[sage burrow]

hit the cntrl + and/or the cntrl - on your kb until it shows up....

[bitter hearth]

didn't work

I just went back to old one

[halcyon yarrow]

@dusky thistle
so you got rid of eta_var? what's the alternative now?
whats beta57 compared to beta? is there any benefit? do you recommend it for somoeone that uses beta for everythinig?
whats denoisie_alt and how is it used? is it safe to just act like it doesn't exisit and use only the denoise field?
for sd3.5 i was using 3/exponential for the shift whats the equivalent? 3 for max and 0 for base? is 1.5 and 1.5 the equivalent to what the flux node was doing?

[dusky thistle]

hard_var = eta_var

clownsampler is still there with the full discrete options

but the difference has been so small i didn't see much advantage to keeping it in a "full package" efficiency style node

base only has an effect with flux

[bitter hearth]

does only hard mode use that or can soft mode use it

[dusky thistle]

it's actually its own mode entirely

the reason i had it separated before is in case it was useful to have another noise mode take over once the math breaks down for hard_var, which is around a sigma of 0.15 or 0.2 or so

but... i haven't seen any benefit tbh

the way the schedules we have are set up, that's usually only like 1-3 steps anyway

[halcyon yarrow]

what about beta57 question?

[dusky thistle]

beta57 = beta scheduler with alpha = 0.5, beta = 0.7

i like it a lot

there's currenty a bug i haven't figured out yet where denoise doesn't work correctly with it

denoise_alt = rescale the sigmas instead of slice them, like ksampler and regular denoise does

[halcyon yarrow]

oh im relying heavily on denoise for my genearations so thats important to know ill avoid it then thx

and denoise_alt should be insigifnicant from a functional standpoint?

[dusky thistle]

for the exponential thing, sd35 medium and fulx both use exponential already, it's large that doesn't

you can still hook up that timestep patcher node and it'll patch sd35L to exponential mode

[halcyon yarrow]

using your new workflow, looks great

[dusky thistle]

give it a shot, it's a pretty different effect

i've found it can be easier to denoise something to a small degree using it

try setting it to like... 0.9 or something

[halcyon yarrow]

this is my denoise code

let lowRange = 0.75;
let highRange = 0.95;
if (isSd3Model || isFluxModel) {
    lowRange = 0.45;
    highRange = 0.65;
}
nextItem.denoise = _.sample(_.range(lowRange, highRange, 0.05));

[dusky thistle]

the best settings i found for flux were max_shift = 1.35, base shift = 0.85, using beta57

it's certainly possible that i was optimizing for certain image types or resolutions that worked best with certain samplers

but that wokred really well and seemed to translate fairly well to sd35

the beta57 that is

found the default shift is pretty much fine... 3.0

[halcyon yarrow]

is there a way to disable it so i can just rely on the nodes rather than the sampler? like if i set it to 0 and just leave the existing nodes as is?

[sacred jewel]

I steal prompts and I don't mind getting in trouble for it 😛

[dusky thistle]

the shift?

[halcyon yarrow]

yeah

liike can i set base and max shift to 0 or -1 to not affet it?

[dusky thistle]

hahah no prob, i def recognized some prompt gen outputs :D:D

[dusky thistle]

i'll implement that now

[halcyon yarrow]

ok cool thanks yeah i just dont have variable logic in my stuff per field per targetBaseModel so I'd have to create some new code to support variable shift per model or have 2 nodes in my WF and its all a lot of work

what are you going to settle on, -1 or 0?

[dusky thistle]

alright you should be able to git pull and set to 0.0 for either shift value and it'll disable it entirely

can you actually do anything with a shift of 0?

does it just explode

lol

[halcyon yarrow]

very good thx for the support

[dusky thistle]

np

if you pulled a min ago pull again lol

idk wtf happened but somehow a piece of this chat got pasted into the code right as i pushed lol

[halcyon yarrow]

and you renamed the class_type ClownsharKSampler lol sheesh

[dusky thistle]

well, clown still exists

[halcyon yarrow]

Updating eb8c99e..5178739

[dusky thistle]

im' gonna keep that around as a pure sampler option

this is just me mashing together the most important options from clown and shark

[halcyon yarrow]

{
                        "noise_type_init": "perlin",
                        "noise_type_sde": "studentt",
                        "noise_mode_sde": "hard",
                        "eta": 0.5,
                        "noise_seed": "seed",
                        "control_after_generate": "randomize",
                        "sampler_mode": "standard",
                        "sampler_name": "res_3s",
                        "implicit_sampler_name": "default",
                        "scheduler": "beta",
                        "steps": "steps",
                        "implicit_steps": 0,
                        "denoise": "denoise",
                        "denoise_alt": 1,
                        "cfg": "cfg_scale",
                        "shift": 0,
                        "base_shift": 0,
                        "truncate_conditioning": "false"
                    }

this is gonna be my settings

[dusky thistle]

should be fine

if you get weird results, play with the noise, gaussian/gaussian is gonna be the most reliable usuuuuualllyyy

[halcyon yarrow]

i have a json file that generates a workflow for each target model using the spec so I'm gonna test the models now, i could automate this part but its better to test it manually

ii did have it set to gaussian for everything on both nodes so ill keep that in mind

[dusky thistle]

okay it doesn't explode with shift of 0.0

so i'm gonna make that "disable" value -1

[halcyon yarrow]

oh you changed the order of the ksampler's outputs 🤦‍♂️ another thing i have to relaign, you keep changing it too

okay so youre settling with -1, got it lmk when do a pull

what's the denoised output on the ksampler do btw?

[dusky thistle]

order is the same, it's a new node lol

[halcyon yarrow]

technically iti's a replacement for sharksampler from my pov, i see iit only has "output" and "denoisie"

sd1.5 looks great with those settings, trying sdxl now which also uses img2img so let's see if 0.5 denoise will still work as before

[dusky thistle]

yeah, before it had fp64 for both but that was really only important for unsampling, and i just threw it in as another element in the latent image output

it'll check when resampling for latent_image['samples_fp64'] and grab that if it's avail

some nodes blow up if they get fp64, a number of em do actually, hence the reason for trying to ensure the output is the same dtype as what went in

[halcyon yarrow]

makes sense

looks sharp but im seeing 83 seconds on sdxl not looking good on that end, i do have iit set to 3s tho ill try 2m once its done

[dusky thistle]

oops i'm a fuckin idiot

                if isinstance(model.model.model_config, comfy.supported_models.SD3):
                    model = ModelSamplingSD3().patch(model, shift)[0] 
                elif isinstance(model.model.model_config, comfy.supported_models.Flux) or isinstance(model.model.model_config, comfy.supported_models.FluxSchnell):
                    model = ModelSamplingFlux().patch(model, shift, base_shift, latent_image['samples'].shape[3], latent_image['samples'].shape[2])[0] 
                elif isinstance(model.model.model_config, comfy.supported_models.AuraFlow):
                    model = ModelSamplingAuraFlow().patch_aura(model, shift)[0] 
                elif isinstance(model.model.model_config, comfy.supported_models.Stable_Cascade_C):
                    model = ModelSamplingStableCascade().patch(model, shift)[0] `````

woooops

it's only triggering if it's at 0 lol one sec

or less than lol

[halcyon yarrow]

that looks right doesnt it? -1 to disable?

uh oh SD3 cominig out all black 😦

thisi is the not-so-fun part where i have to retry sd3 with various settings to see whats wrong

sd3 workflow if you wanna help me debug it

[dusky thistle]

k it works now

get pull it again

if shift is < 0, then it skips setting shift for everything except flux, which skips setting shift if either shift or base shift are < 0

[halcyon yarrow]

ok ill try it again

[sage burrow]

keep hitting them over and over until you see the right button

[bitter hearth]

hmm

[dusky thistle]

whoa this is weird lol i don't have this node nor did i make one with this layout

[halcyon yarrow]

that's just how it looks like with my stuf, its your node your class_type but i edit it a little bit lol

[dusky thistle]

ahh k gotcha just making sure something real crazy wasnt going on lol

i figured you were prolly messin with it

[halcyon yarrow]

fun side fact I found an endpoint yesterday for ComfyUI that I can hit that enables previews, bc even tho its set to auto in the command line iit still doesn't set it to auto for websocket connected clients so calling this endpoint after connectin makes your ksampler behave like my other one (ksampler efficient advanced)

enablePrevews() {
    this.fetchApi('/api/manager/preview_method', {
      method: 'GET',
      headers: {
        'Content-Type': 'application/json',
      },
      data: JSON.stringify({ value: 'auto' }),
    });

alright sd3 works great looks good

[dusky thistle]

shift disabler workin fine too?

100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 60/60 [00:24<00:00, 2.44it/s]
Prompt executed in 25.89 seconds
got prompt
Requested to load SD3
Loading 1 new model
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 60/60 [00:24<00:00, 2.42it/s]
Prompt executed in 25.22 seconds

first is dpmpp_2s_ancestral in ksampler, second is res_2s in CSK... they both use two model calls per step and should have similar runttimes

[halcyon yarrow]

soliid mate, getting 14 seconds now with the new updates

[dusky thistle]

yeah i know some shit was way fucked up a while ago as i was working through some issues with the rewrite

[halcyon yarrow]

oh wait sorry i jumped the gun its stiill using ksampler my bad

[dusky thistle]

you prolyl just had something with a bug that cleared up during that process

hahaha

[halcyon yarrow]

alright 36 seconds, wiih 5 loras and 24 steps on the first run using res_2m, that's totally within acceptable range I can leave it as the default at this poinit

[dusky thistle]

yeah it should be pretty damn similar to euler with ksampler

there's a slight price for the noise generation

if your comp is slow and you're using huge latents it becomes more apparent for the sde solvers

[halcyon yarrow]

21 seconds with 3x loras at 19 steps yeah it's all good now man

im glad you solved it

[dusky thistle]

13/30 [00:26<00:34, 2.02s/it]
euler with ksampler

euler with CSK:
| 8/30 [00:14<00:40, 1.83s/it]
1792x1344 latent

great to hear

yea i'm sure you know how it goes, often times the best way to fix bugs is just to ignore them and keep cleaning up code

[halcyon yarrow]

aint that the truth lol

[dull star]

CSK?

custom sampler?

[dusky thistle]

clownsharksampler, new node

[dull star]

ahh

[dusky thistle]

rolled into the functionality of clownsampler and sharksampler

[halcyon yarrow]

now that ive established a baseline for res_2m i'll move over to the other side of the spectrum and use res_3m see what kind of times I get for that and if it's within reason

[dull star]

what does your sampler have to offer (for flow matching models for example)

[dusky thistle]

all the *m should have similar runtimes

[dull star]

does it have noise injection and stuff

[dusky thistle]

tons of features nothing else has

yeah, it's full blown ODE and SDE

30 samplers, 20 noise types, 6 noise scaling modes

[dull star]

cause im braindead when it comes to sampling values (like wtf is eta, all I remember is that it should be like 0 if its flow matching model, etc)

[dusky thistle]

unsampling, latent guide modes, noise inversion (redid all the math for it, getting better results than with the paper's implementation)

[dull star]

so I only tried noise injection cause its simple enough

[dusky thistle]

https://github.com/ClownsharkBatwing/RES4LYF got a bunch of WFs here

[halcyon yarrow]

i was leaving it at 0 but then i experimented with 0.5 and it works fine for all models

[dusky thistle]

the thing about it needing to be 0 is totally false

it WAS true with the other samplers

but it's not true at all that it can't be made to work

tbh, results are def better with noise

the math had to be reworked completely for the noise scaling

[halcyon yarrow]

what's the equivalent to res_3s in the ksampler world? you said res_2m is equivalent to euler right?

[dusky thistle]

in terms of speed? there isn't one

there's only the equiv of *m samplers, and 2s

the 2s ones are just dpmpp_2s_ancestral and dpmpp_sde

heun, heunpp

[halcyon yarrow]

no like in terms of algorithm like what's 3s equivalent to if there is any in ksampler as far as like the math that goes into it

[dusky thistle]

there isn't anything

[halcyon yarrow]

oh cool so its unique stuff nice

[dusky thistle]

in terms of the algorithm, there's only euler, 2m, 3m, and 2s

some variations on the 2m and 3m themes like unipc etc

insofar as the noise modes go, the new euler_ancestral and dpmpp_2s_ancestral use something pretty similar to noise_mode_sde = "soft" and eta = 1.0

noise_type_sde = gaussian

[bitter hearth]

do you like brownian or gaussian more

[dusky thistle]

it really depends

on what, i'm not even entirely sure

gaussian is more reliable

but brownian often gives a crisper image with a bit more punch to it

but sometimes it looks grainy or shitty

gaussian can lead to more color contrast

[bitter hearth]

hmm okay

[dusky thistle]

most of these options are back now btw

[bitter hearth]

I liked uniform and high frequency fractal power noise

[dusky thistle]

just have to hook it up

it overrides the same options in CSK

[bitter hearth]

ah nice that's cool

I should explore c2 and c3 more I guess

[dusky thistle]

it only does something with res_2s, res_3s, and dpmpp_2s and dpmpp_3s

the rest of the methods come with hardcoded ci

dpmpp_sde_2s is hardcoded at 0.5

otherwise it's the same as dpmpp_2s

err sorry

hardcoded at 1.0

[halcyon yarrow]

im going to test beta57 now and see if it works with all the models

SD 1.5 ✅
SDXL ✅
Pony ✅
SD3 ✅
Flux ✅
yeah it seems to have no issues with any of the models I think i'll adopt it just to be a little different and try it out

[dusky thistle]

awesome, good to hear

[craggy crest]

SD3.5 medium

[halcyon yarrow]

saw a video of a fancy new upscaler, check out this image, zoom in on the faces of the soldiers notice how InstantIR is the only one that gets it right

[bitter hearth]

InstantIR might be the new open source SOTA yeah

[dusky thistle]

afaik... first ever 10th order sampled diffusion image on here (SD35M)

[cunning mesa]

2nd order of coffee.

[dusky thistle]

https://media.tenor.com/5-VyD_RyqsYAAAAM/dr-strange-love-nuke.gif