PBRFusion 2 (Legacy)

[shut tide]

Here's a more extreme use, the upcsaler ran twice (64x64 -> 256x256 -> 1024x1024) on a texture from oddworld munch's oddysee, but at this point its essentially a completely new texture
https://imgsli.com/Mjg5OTUz

[mental geyser]

waipt

https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union

controlnet union came out with beta, with tile

yesterday

[shut tide]

ooo

[mental geyser]

ill try this out in a bit

[shut tide]

me too (i am very uneducated with controlnets but will try my best)

[mental geyser]

need a wrapper

[mental geyser]

i was mistaken no wrapper

well a special node

[shut tide]

im still always skeptical of custom nodes that have less than a few hundred stars ever since the LLMVISION incident

[mental geyser]

valid

[shut tide]

i'm getting an OOM error on a 12gb card when i add the controlnet node, will fiddle more with it later

[mental geyser]

ok i was able to get gguf q3 to work without OOMing, but still some problems with union result, plus yeh generation time is still abysmal, gonna wait on a dedicated tile model.

I also see more buzz on reddit about more vram savings tn, so it just needs more time to mature as well

[shut tide]

yeah, the controlnet model itself is apparently 7gb with bf16. Just seems like we'll have to wait a lil longer for more individual controlnets

[wary forum]

💀

what the heeeeell

[gritty dew]

did I just add rest api?

I just copied the nodes from image and adpated it to PBRFusion

[gritty dew]

No I didn't

But I will keep trying

[gritty dew]

Well It keeps trying to run the upscaler before it actually gets the texture idk how to fix that

[gritty dew]

Alright so removing the get and set nodes seem to be the only way around this

but unfortunately I don't have enough vram to run this

[gritty dew]

Everything looked to work before I crashed due to lack of system resources

It does need a bit of cleaning up

I've never felt so limited by 12gb of Vram

[gritty dew]

I GOT IT TO WORK

it all works

needs to be a low res texture though

actually the depth isn't working properly

well just that one texture it seems

[mental geyser]

Yep that's the main problem I haven't made one

Working on a fix to just load capture without viewport

[gritty dew]

The difference in one click is amazing

This would work so well on a 4090

tbf it is just the upscale that eats the VRAM

[gritty dew]

So do you already have a workflow modified for Rest API?

[mental geyser]

I haven't done an official one, a batch load version and a one click would be different projects

[quick zenith]

Fun fact the Remix node to get Textures should allow you to get from the selection if you want to do textures 1 by 1, otherwise you can do your entire project. It should just be 1 switch

[mental geyser]

I have seen that thanks 🙂💙

I meant more on the making a kit app although Ig I could maybe put in a toggle viewport button somehow

[wary forum]

1 click?

Are you about PBRFusion?

[rustic talon]

@mental geyser btw if you want to pull the PBRify models at runtime instead of packaging them in with the zip, you can use the direct links from here:
https://github.com/Kim2091/Kim2091-Models/releases/tag/4x-PBRify_UpscalerDAT2_V1
DL: https://github.com/Kim2091/Kim2091-Models/releases/download/4x-PBRify_UpscalerDAT2_V1/4x-PBRify_UpscalerDAT2_V1.pth

SwinIR M:
https://github.com/Kim2091/Kim2091-Models/releases/tag/4x-PBRify_UpscalerSIR-M_V2
DL: https://github.com/Kim2091/Kim2091-Models/releases/download/4x-PBRify_UpscalerSIR-M_V2/4x-PBRify_UpscalerSIR-M_V2.pth

also yes i totally forgot, but i released a SwinIR Medium model as well lol

[mental geyser]

Thanks 🙂 yeh I remember Swinir being mentioned at some point

[mental geyser]

[mental geyser]

ok im tired of waiting on tile, ill try it myself, (only trained on textures instead) ill start on sdxl tho, @rustic talon what was the script you made that did the downscale stuffs

[rustic talon]

https://github.com/Kim2091/helpful-scripts/tree/main/Dataset Destroyer

this one has succeeded it in features and reliability since then: https://github.com/umzi2/wtp_dataset_destroyer

[mental geyser]

you reccomend 2nd one ?

[rustic talon]

yes

but the default config is kind of a pain in the ass

this is my config for game textures: https://github.com/umzi2/wtp_dataset_destroyer/blob/master/configs/custom/kim/game_texture.json

[mental geyser]

ty kim!

[mental geyser]

This will be harder then i thought, i will have to write some of the training code from scratch...

[mental geyser]

training

||on sd1.5 to test first||

[neon tulip]

YEAHHHHH BOIIIIII

[mental geyser]

I think this is at least matching quality of sdxl model, using kim's span model as preprocessor.

I need to work with it more,

Ive also started a fine tune of sd1.5 with textures. This is new territory for me and ive only ever seen loras of texture stuff on models so well see how this goes

[neon tulip]

HE COOKIN CHAT

[mental geyser]

||POSSIBLY COOKIN * ™️ ||

[mental geyser]

fine tune is doing okay

left is fine tune

still has some more trainin to do

[gritty dew]

So what is this?

[mental geyser]

another another side project lol, dont worry im still working on upscaling, as we speak actually lol

thought being a tuned sd1.5 towards textures to help upscaler, but i may be going another route already

[mental geyser]

🙂 🙂 🙂

so this is brand new research paper stuff

after 6 hrs of stupid ass dependnecy errors i fixed alot of stuff

i just got this running natively so i gotta wrap it into comfy now but heres some sneak peeks

https://imgsli.com/MjkzMzA1

https://imgsli.com/MjkzMzAz

https://imgsli.com/MjkzMzAy

[pale field]

what does it do exactly? are these textures generated by AI image models?

[mental geyser]

no this is upscaling

[pale field]

this models definitely works a lot better and outputs cleaner results on certain geometry like straight shapes or lines but lacks in reconstructing small divets and holes

overall im impressed with this model of yours

[mental geyser]

aint mine, its just brand new

[pale field]

what model is it then

[mental geyser]

1 step generation btw kinda crazy

https://arxiv.org/pdf/2406.08177v2

[pale field]

do you intend to have any support for Llava for automatic prompt generation? i know SUPIR works on prompts to works effectively

[mental geyser]

this is using recognize anything which is doing quite well

[pale field]

that's good

[mental geyser]

speed

512 -> 2k 30s

256 -> 1k 7s

128 -> 512 3s

i think tiling is slowing down 512 to 2k ill look into it more

[rustic talon]

it looks very impressive, but in all of those results it kinda loses some of the finer detail over larger surfaces

kinda worried it'd lead to very flat? outputs

[mental geyser]

true! more to tweak and experiment on but bed time now haha

[rustic talon]

certainly has promise 🙂

and goodnight!

[pale field]

That's fast

[rustic talon]

indeed

very close to normal models

[distant gate]

This is such an infamous texture at this point 🤣

[distant gate]

https://imgsli.com/MjkzMzQ0
https://imgsli.com/MjkzMzQ1
https://imgsli.com/MjkzMzQ2
https://imgsli.com/MjkzMzQ3

Amazing work Night, I havent touched any models since May but after seeing the comparisons people posted I am stoked to use this. Enjoy some comparisons of these buttons! Thank you man!

This is a really interesting case https://imgsli.com/MjkzMzUw

[distant gate]

My feedback similar to happysufigee's, when there is guestimation that needs to occur, opting for angular/straight would be preferred if possible. Right now in this picture since there is so much guesswork, it ends up very fluid/uneven looking for where there should be panels.

[mental geyser]

Great results, glad you liked it 🙂

[mental geyser]

Ty for the feedback, hopefully I can find the magical formula(s)

https://tenor.com/view/calculation-math-hangover-allen-zach-galifianakis-gif-6219070

[mental geyser]

so, this is a diffrent paper, 50 steps slower and takes up quite a bit of vram but probs worth it

https://imgsli.com/MjkzNDI3

[neon tulip]

THAT’S SO GOOD

[pale field]

This is the one

Best one so far

How much time did it take?

[mental geyser]

131 seconds, but mostly cuz it went into my system memory cuz i ran of of the vrammmm

i wonder if it could be optimized with tensorrt

[pale field]

This one is crazy good, what's it called

[mental geyser]

https://github.com/cswry/SeeSR

[distant gate]

Yeah... sorry I hasnt the clue what Im talkimg about sometimes 🤪 . Would love to see the other model if you can run it faster 🙂

[mental geyser]

Curious to see how it does on this pic, img SLI doesn't let me download small version

[rustic talon]

ya that's a ton better. very very impressive

[rustic talon]

ah no wonder lol. a lot of friends told me this was better than SUPIR, but i never got around to trying it

[rustic talon]

imgsli does have JPEG compression though, so it's even worse than the original input

[mental geyser]

https://imgsli.com/MjkzNDY1

[rustic talon]

i think if it had the original it'd look a lot better

[mental geyser]

yuuuup

[distant gate]

Oh lord I should read chat before posting, but yeah this is the orginal in png

And this is the dds. When I did the conversion, I used this dds variant

[mental geyser]

i ahve no idea how to convert this lol

[rustic talon]

it's probably just the alpha layer causing issues

throw it away

[mental geyser]

forget yesterday?

[rustic talon]

recently i forget everything

[mental geyser]

🎶 well make the great escape 🎶

[rustic talon]

this is with alpha removed

[mental geyser]

thanks pal

[rustic talon]

pretty sure that's what arro posted already lol

just realized

[mental geyser]

oh

yeh

tbf he did edit it after and say, ifigured it was jpg

bout the same

https://imgsli.com/MjkzNDcw

[distant gate]

I wont lie I love the original model because of its creativity with adding details

[mental geyser]

i cant win

lmao

FINE ARRO gimme a sec

[distant gate]

https://tenor.com/view/patrick-scheming-plan-gif-10212416

[mental geyser]

first im clearing out my recycle bin full of 400,000 txt files

[rustic talon]

who caaareess about performance!

just run both and blend the results together

[mental geyser]

...

[distant gate]

https://tenor.com/view/put-them-together-build-assemble-put-together-put-up-gif-15623019

[mental geyser]

no no no he wants creativity

HES GONNA GET IT

[mental geyser]

ok nvm my idea just made it noisy, its like its made to stick to original or sumthin

[distant gate]

What would I need to tweak in Comfy to have it upscale to 4096/4x by default? Or like the older iterations, is 2k the standard?

[mental geyser]

https://cdn.discordapp.com/attachments/1263287747905327236/1268637243308773386/image-2.jpg?ex=66d75600&is=66d60480&hm=710f2a7fad0f012e514437d82759318515b479fd44d7e48e973c11f7200b333f&

Change the 2 s

To 4 s

Or 8 s for 8k 👀

so the SeeSR upscaler doesnt do very good above 4x. so i will still use tile as well, need to mess around more

[distant gate]

👍

[distant gate]

Last question how would you go about preventing the suffix of 0000X_ from being appended to the file?

[mental geyser]

it shouldnt be appending anything with default workflow

[mental geyser]

what does this look like for u

[distant gate]

Empty same

[mental geyser]

its not appending that for me ...

click save on right

and send that file here

[mental geyser]

running that doesn't do same for me either

how do you have duplicate files with same names

[distant gate]

Disregard that, picture from the recycling bin lol

[mental geyser]

ah

im not seeing your behavior, or see a reason it is

and now im botherd

youre coming straight off the zip right?

[distant gate]

Yep

[mental geyser]

not ideal but worse case run a script that deletes {-10001-} from end of all base names,.... i cant freakingwrite under scores apparently

[mental geyser]

@primal olive seesr on ur texture

https://imgsli.com/MjkzNTQ0

[primal olive]

goddamn

this may be one of the best results yet

[distant gate]

Did a 4x upscale of a generic reference image from MC. This looks like so awesome/stylizied it could be it's own texture pack
https://imgsli.com/Mjk0MTUz

[mental geyser]

Most things look good but what did it do to stone and cobble

[rustic talon]

this looks like the realisticrescaler model

it adds a lot of fine grain and generally does a good job, but kinda fails with finer detail

[mental geyser]

yeh the grain that comes in is my least favorite part of the upscaler

seesr does a much better job in that regard

[rustic talon]

well, to be fair, the author somehow managed this level of refinement on an SR arch from 2017 😛

[mental geyser]

well yeh, not discrediting previous impressive work. 🙂

[mental geyser]

ive done it! the ultimate compromise

down from 2 minutes to 13 seconds, 256 -> 2k

https://imgsli.com/Mjk0MjIz

going from 50 steps sd2.1 to 10 steps sd turbo, i see similar if not sometimes better quality

and then ran through kims span upscaler to get to 2k

512 -> 2k is still bugged (at least with my 10g of vram) it takes over 1.5 min, tiled vae decoder is being weird i havent looked at it yet 🫠

[pale field]

Amazing, what's the time for the 10 step result

[mental geyser]

For 256 4x, 13 seconds even with Kim's scaler

[distant gate]

This looks amazing, can you try the one in the referenced message again

[mental geyser]

I just turned off my PC and headed to bed lol I will tomorrow

[mental geyser]

@distant gate does a lot better actually

https://imgsli.com/Mjk0MzQ1

[primal olive]

Can you explain a bit more? Is this using PBRify? Or your SD based upscaler

[mental geyser]

https://github.com/cswry/SeeSR

It's seesr, if base res is below 512 since it's only doing 4x, then it will run through a Kim model to get to 2k

So

512 -> 2k seesr

256 -> 1k seesr -> 4k Kim's span model, downscale to 2k

128 -> 512 seesr -> 2k Kim span model

Is the current thought, I can change upscale amnt but need to test if quality suffers it does at 8x in seesr

[wary forum]

why so serious

[pale field]

It's either SeeSR or Caesar

[wary forum]

(just zoom it)

[mental geyser]

?

[wary forum]

I'm about the details

[rustic talon]

looks great. great thinking on the workflow

[distant gate]

This output is nearly scary, its so good at improving details. Almost not just a upscaler but a detailer too lol

But for the ultimate test... do the 4x upscale

[mental geyser]

lol

[distant gate]

That is definitely... an interesting generation

This brings up an interesting case study. Can you train a model to understand when to replicate a pattern, aka block textures

[mental geyser]

im not following like be more stable on same texture grass in background?

[distant gate]

Dirt pattern is a good example, differs in the distance. And differs on the ground (far back three blocks)

[mental geyser]

mmmm maybe a segment controlnet in addition to diffusion model

for upscaling, dunno

[distant gate]

Well either way, I want my grubby paws on this asap. I would love to hurl countless old ship textures at it to see the detail it can add

https://cdn.discordapp.com/emojis/1117656665970454570.gif?size=80&quality=lossless

[mental geyser]

im making the comfyui wrapper as we speak 🙂

[distant gate]

Man you really are the

https://tenor.com/view/goat-tongue-gif-26615227

[mental geyser]

why he lickin so much

[distant gate]

Just typical daft/dumb goat behavior. I've had to deal with worse... I'd have to deal with a pack of like 30 or so raiding a barn whenever they heard the sound of a door open. Pure chaos

[mental geyser]

lmao

[mental geyser]

I'm so confused, I got the wrapper working, but then it was giving a worse output, no big deal I probably did something wrong, went back to other environment..... Where it was working, now same problem, I even have a snapshot of working files those don't fix it, I've made a whole nother environment doing same steps as I did before.... Same result I'm severely pissed off rn

[mental geyser]

Something must have severely bugged out, in a good way?

I found the parameters to fix but still have no idea what happend cuz these weren't touched before

Also switched back to sd2.1 and 20 steps at a similar speed, which improves quality in most instances, also don't gotta worry about sd-turbo license

So maybe this was a good thing

But I'm still mad

[rustic talon]

😦

[mental geyser]

I have the problem you were having now @distant gate lol, at least now i can troubleshoot it

[mental geyser]

this is getting a lot closer to being finished, wrapper is working great and is in a working full PBRFusion environment. New comfy version has broken a couple things and still have some housekeeping to do but nearly there 🙂

zip size is going down from 39GB to 25GB 🙂

well thats without venv

so that may be 30GB

[distant gate]

How goes the update?

[mental geyser]

i fixed your problem lol

i have changed max batch size before and now cant find it

[distant gate]

Looking forward to this new model

[mental geyser]

https://tenor.com/view/cat-catdriving-gif-11227672289482918211

[distant gate]

https://tenor.com/view/right-now-eric-cartman-craig-tucker-lola-south-park-gif-22863637

[mental geyser]

what gpu do you have again arro?

[distant gate]

3090

[mental geyser]

btw, with the changes ive had to make since last thing you saw, the upscaler is much more of an upscaler and less so on the making up details

https://imgsli.com/Mjk1MTM1/1/2

the pic named ComfyUI is new version

but as you can see turbo would hallucinate alot of things

and ive lost that way to make that result so it what it is (not without trying to get it back)

but new one def performs better imo, look at how much more detail is in 20 new compared to 10 old

https://imgsli.com/Mjk1MDc4/1/2

you can play wih the guidance scale paremeter to let the model have more freedom tho

[distant gate]

I mean hey this is what you showed a few fays agos, its what Im hooked on. I just cant believe how it can go from this level of blur to this level of detail. All whilst being soo accurate

And I would be cautious with saying it doesnt 'make up details'

It def does, but its the approach that matters. Is it making up using real world examples, or is it making up just based on geometry/color, etc.

Certainly this model has a great approach

[mental geyser]

im saying that that right result is gone, zip, in the void

[distant gate]

Yes I get ya on that.

Also here is some more great examples of spaceship textures, which are very small and would require a great deal of "making up detail"/enhancement

[rustic talon]

is this still using an upscaler after the diffusion generation?

i do like the new model. it's a bit soft, but it's a lot more faithful which is good imo

[mental geyser]

i will in final workflow, here no

[rustic talon]

that could help with the softness then

[mental geyser]

https://imgsli.com/Mjk1NjI2

[distant gate]

Interesting result, not gonna be harsh on it, really doesnt have much to work with

[mental geyser]

https://imgsli.com/Mjk1NjI5

yeh and its all like grayscale lol

[rustic talon]

that's particularly good

most traditional models would turn it into a hair-like texture

[mental geyser]

well if kim says its good

[rustic talon]

try this please

PNG format

this would also be interesting

none of the normal SR models i've tried can handle these well

[mental geyser]

https://imgsli.com/Mjk1NjMw

[rustic talon]

that's very impressive

that's definitely a good base to work on top of now

if you want to test a nightmare scenario:

[mental geyser]

https://imgsli.com/Mjk1NjMx

lol

[rustic talon]

i was hoping for better, but tbh that input is just horrid lol

i'd say it did pretty good considering

[mental geyser]

hes trying his best 🙂

if i let it fly with creativity

[rustic talon]

another nightmare

[rustic talon]

lol. that's interesting

[mental geyser]

https://imgsli.com/Mjk1NjMz

its a dot painting

[rustic talon]

yeah...

so many textures from Barnyard are like this

probably over 60% of them are just unable to be worked with or upscaled. really bad quality

[mental geyser]

😦

waipt i wanna try something

[distant gate]

Bro this like wtf levels of upscale. The middle right looks so high quality

[mental geyser]

https://imgsli.com/Mjk1NjM4 normal looks decent 2

[distant gate]

Ill get you the OG normal in a sec

Can you upoload the new bump directly here

[distant gate]

Going to do a imgsli with the og to generated

https://imgsli.com/Mjk1NjQx

[mental geyser]

ayooo

not bad

[distant gate]

This is wild the difference in approach from the developers vs the model

Not in a bad way

Looks great!

[mental geyser]

i love that its doing this correctly

on the dome

[distant gate]

Indeed

[mental geyser]

lol \

[distant gate]

https://tenor.com/view/thats-it-yes-thats-it-that-right-there-omg-that-thats-what-i-mean-gif-17579879

[mental geyser]

i am stuck on something that should be ez lol ( not stuck ig per say i just cant find it........)

thats how it goes

[mental geyser]

https://imgsli.com/Mjk1NjQ2

[distant gate]

Defiently helps, but there is an odd discolorization

[mental geyser]

eh can also run it through a color match

[rustic talon]

the SPAN models would try to reverse the color shift present in DXT1 compression

but the diffusion model already does that

[mental geyser]

hmm i can skip the color match in seesr and let span handle

[rustic talon]

@mental geyser what if you were to train a 1x model on pairs of low step and high step outputs. this would be a fairly lightweight model. think SPAN or maybe RealPLKSR dysample if necessary?

it could clean up artifacts and overall provide sharper results

i'm not sure how fast the model would need to be to outweigh the performance hit of just... running more steps to begin with

[mental geyser]

worth investimigating

[rustic talon]

for an idea of speed, you can refer to this set of benchmarks: https://github.com/the-database/traiNNer-redux/wiki/PyTorch-Inference-Benchmarks-by-Architecture#1x-scale

realplksr dysample is drastically slower than SPAN, but also is near DAT2 quality

[mental geyser]

so this is seesr doing no color processing

https://imgsli.com/Mjk1NjQ5

[rustic talon]

hm

so

if you want a version of this model that does almost no color modification, i have it hosted on github here: https://github.com/Kim2091/PBRify_Remix/raw/main/Extras/Alternate_Models/4x-PBRify_UpscalerSPAN_Neutral.pth

it's V4 but a checkpoint of it near the end of training that exhibited minimal color shifting

i put it up for people upscaling normal maps and similar, but it should work nicely for this

the initial color correction from seesr seems to be more precise than what my model was doing

[mental geyser]

https://imgsli.com/Mjk1NjUy

nice!

wait wow i must be getting tired

its a bit better

[rustic talon]

to me the original model did better reconstructing the details

but the color is a lot better here for sure

ah wait it's a difference in the seesr output, not the span model

old:

this one

[mental geyser]

wow yeh

[pale field]

can i request a texture upscale please? dont have my PC setup ready to test it out so that why

i wanna see how well it does with trimsheets

[mental geyser]

sure, its not out yet lol we just messin around

[pale field]

okay lemme fetch the texture for ya

[mental geyser]

more color testing to come, im not against just doing a post color match after all proccesing

[rustic talon]

definitely test the neutral model a bit more before settling on it. to me it's a bit weaker, but that may also be what you want from it

lol

[gritty dew]

We're upscaling diorite

[mental geyser]

bird poop*

[rustic talon]

ya. personally i'd probably just do a color match at the end and use whichever model does a better job

[mental geyser]

time to pull out hm-mvgd-hm

[rustic talon]

what is that lol

[mental geyser]

its what i use on current one lol

mixture of some methods ig

[mental geyser]

[pale field]

I couldn't find it

[distant gate]

I demand the goods!

[mental geyser]

i need a 40 series person to serialize an engine

p p pwease

[mental geyser]

||release tommorow likely||

[pale field]

What do you mean by serializing an engine?

[mental geyser]

Compile a model to use tensorrt

The compilation is only compatible with same architecture/series, so I can compile it for 30 series but it won't work on 40

[primal olive]

What's the process? If you make it easy I'll do it and send the results lol

[mental geyser]

Not hard I've got a python venv and a file in a zip and run one command

[primal olive]

In that case I can do it just fine. How big is the download? Won't be home for a while but I could prob get it today

[mental geyser]

9.46GB

[rustic talon]

i was told that engines are only compatible per GPU, not per arch

[mental geyser]

these have been working per arch

[rustic talon]

hm

[mental geyser]

at least im p sure @gritty dew you got pbr working right?

yeh cuz Arros works and he has a 3090

i got a 3080

vovanchos works off a 4060

that was built on skurrtys 4090

[rustic talon]

that's interesting

i tried a model from a 3090 on my 3070 ti and it failed (for ESRGAN)

was then informed by a dev who works with trt like daily that it was per GPU

not sure what happened there lol. maybe things have changed since then? that was about 2 years ago

[mental geyser]

the future is now

https://tenor.com/view/bilbo-baggins-running-adventure-gif-10198122811708180458

we have several a40s at work im tryna get my boss to let me use em lol

---

well if anyone with a 40 series can download this

https://huggingface.co/NightRaven109/PBRFusion/blob/main/autoserializedepth.zip

go into ComfyUI-Depth-Anything-Tensorrt folder and run the bat file called "serialize", then upload the the depth_anything_v2_vitl.engine file it spits out itd be gr8ly appreciated

[rustic talon]

about how large will that file be?

[mental geyser]

644 MB

[rustic talon]

if no one else does it i should be able to in about 3-4 hours

[mental geyser]

💚

Ig I could rent a GPU in runpod for a couple min too lol

Dunno why I didn't think about that

[neon tulip]

https://tenor.com/view/help-is-on-the-way-dear-mrs-doubtfire-robin-williams-gif-12917727

Downloading and getting ready to run now

[mental geyser]

my hero

[neon tulip]

https://drive.google.com/file/d/1D8M7lxmap_g2-8VE9Zd9VXdtUhrNbUok/view?usp=sharing Boom! 40's series engine done

[mental geyser]

[rustic talon]

was just getting ready to do it 😛

[mental geyser]

PBRFusion 2

🙂

Big Update

PBRFusion 2.0 !!!!

• Replaced upscaling with SeeSR and Kims Span Model, Giving much better quality

• Added RTX Remix integration Workflow

• Upscaling will auto go into PBR generation

• Updated to Newest Comfy Version

• Got rid of Atlas Mode Altogether

Notes:

You will need to reinstall, simply too many things changed to make a patch

Zip file went down in size by about 4 GB not as much as I hoped but better then more

This likes to use about 8 GB of VRAM to process, You can try decreasing latent_tiled_size in seesr node to save on vram. I plan on trying to reduce this by taking out the RAM (Recognize Anything model) from SeeSR freeing about 5GB of VRAM. This shouldnt fail with under 8gb of vram but since it will go into shared memory it will be much slower. There is an option on the SeeSR node to not keep model loaded, which will free itself from memory for pbr generation processing but will need to reload everytime once again slowing things down

I also plan to make a way to grab all textures from a capture in remix to batch process, right now it is just the 1 texture per click thing

Notes dont seem long but this is over another 150 hrs of work, hope yall like it 🙂

https://huggingface.co/NightRaven109/PBRFusion/tree/main

[primal olive]

want me to pin

[mental geyser]

sure thanks

[primal olive]

[rustic talon]

aha

[primal olive]

[pale field]

We're on version 2 already?

[mental geyser]

no its 1.6.8.759ab763

[pale field]

Bro accidentally released the sequel

[mental geyser]

now what do i do im bored

||i kid||

[neon tulip]

YIPPEEEEEE! Nice work NightRaven

[mental geyser]

a setting on batch loading workflow is wrong -_-, zip updated

https://imgsli.com/Mjk2MjQy 😮

[pale field]

Can your app detect or create alpha maps for transparency

Like the windows on the door for example

[mental geyser]

not currently most it does is reapplys current alpha (if it exists) to upscale

i suppopse thatd be possible with a trained segment model

[rustic talon]

imo it'd be better to do that manually

[rustic talon]

very impressive

[mental geyser]

BTW for those of you with unlimited VRAM you can increase latent_tiled_size for faster upscaling

[rustic talon]

mwuahah 4096 TB VRAM

[rustic talon]

want me to pin this

[mental geyser]

sure lol

[neon tulip]

When we install 2.0, do we just click run_nvidia_gpu.bat and it should run like normal?

[mental geyser]

yuh

[mental geyser]

if you downloaded it b4 this message you need new workflow lol

[neon tulip]

Ahh bet bet, redownloading

[mental geyser]

oh i can just give the file lol

[neon tulip]

Oh sweet, thanks!

Ground control to major tom

[mental geyser]

that doesnt look like the new workflow

press ok and load new one?

[pale field]

Unlimited 💀

[mental geyser]

24gb = unlimited 🙂

[pale field]

Guess that includes me now

[neon tulip]

Oh I'm so silly, forgot to clear out the old workflow - alls working!

[mental geyser]

oh i forgor! upscale will auto go into pbr gen now, or u can turn off pbr gen and just do upscale

[neon tulip]

Ohhhh, so if I uncheck Material it'll just upscale, but if I have Material checked it'll upscale+generate PBR?

[mental geyser]

yup 🙂

[neon tulip]

NO WAY, click and go! You're the best NightRaven

So for the vram on the latent_tiled_size - would I switch 128 to 24gb for a 4090? 🤔

[mental geyser]

im not sure what the scale is as i didnt have one to test 😝 maybe try 512 and see what vram is at

its tile size not vram size not one to one like last workflow, bigger tile size = more vram but less tiles to process so that = speed

[neon tulip]

Ahh gotcha. 512 seems good, coasting at around 18.7. Would I be able to do something like 513, or would the tile need a certain denominator? 🙂

[mental geyser]

if you click the arrows it will 'step' to the correct interval i think its 32

it wont let you mess it up lol

[neon tulip]

Sweet! On my next batch I'll see what the next interval would be like and report back 🙂 I'm doing an upscale now that we can compare to Fusion 1.0, I'll post the results when they're all done

BEAUTIFULLL, it assigns names on the file for easy ingestion

[mental geyser]

about how fast is it doing 512x512s?

[neon tulip]

It looks like it's averaging 45 seconds for upscale+pbr

[distant gate]

Whenever I load PBRFusionV2.0BatchLoader.json I am recieving error:

Loading aborted due to error reloading workflow data

TypeError: Cannot read properties of undefined (reading 'find')
TypeError: Cannot read properties of undefined (reading 'find')
    at nodeType.onGraphConfigured (http://127.0.0.1:8188/assets/index-CrROdkG4.js:1880:34)
    at app2.graph.onConfigure (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:80196:34)
    at _LGraph2.configure (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:60574:14)
    at LGraph.configure (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:80185:26)
    at LGraph.configure (http://127.0.0.1:8188/extensions/ComfyUI-Custom-Scripts/js/reroutePrimitive.js:14:29)
    at LGraph.configure (http://127.0.0.1:8188/extensions/ComfyUI-Custom-Scripts/js/snapToGrid.js:160:21)
    at ComfyApp.loadGraphData (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:80641:18)
    at async app2.loadGraphData (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:74290:18)
    at async app.loadGraphData (http://127.0.0.1:8188/extensions/ComfyUI-Manager/components-manager.js:771:9)
    at async reader.onload (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:81002:11)
This may be due to the following script:
/extensions/ComfyUI-Custom-Scripts/js/reroutePrimitive.js

All extensions are already up-to-date with the latest versions.

[mental geyser]

Try deleting the I forget what they're called like, comfy.prv files?

Only things in folder should be
ComfyUi
Data
configs
Batchloader.json
Remix.json
Start_nvidia_gpu.bat

And then it will repopulate

[mental geyser]

@neon tulip can you humor me and try latent size at 96 and see if it's faster lol

[neon tulip]

Houston...did I do something wrong https://imgsli.com/Mjk2Mjc4/8/9

[mental geyser]

Hmmm maybe setting latent size so High was actually detrimental

Like I said I hadn't tried it

Try 96 on the preview pic above and see

What are the original resolutions?

It definitely is not the correct result

[neon tulip]

Huh, ran it at 96 and still got the preview quality. The original size is 256x256

Here's the original

[mental geyser]

Ok try , a couple nodes to the right of seesr node ( the one containing latent size) delete color match node rerun

[neon tulip]

This one?

[mental geyser]

Yep

Then change align method to adain above latent size

[mental geyser]

:/

[neon tulip]

Hmm, would you get the same result on your end?

[mental geyser]

I need to bed, so gonna have to put a pause on this till tomorrow sorry!

[neon tulip]

No worries! Have a good rest, I'll look back into it again

[gritty dew]

How doable is rest api with 12gb of vram?

[mental geyser]

Ummmmm, you gotta decrease render resolution in remix lol, and then maybe

[mental geyser]

i do get the same result, seems like a worse case for this haha, using this ill see if i can develop something better later on

as for the other ones, id say if you put them up against the original as well, it does a much better job of staying true to the original, where the old one is good at making things up

you can increase guidance scale to give the model more freedom (like to 3 or 4) but it also slows things down and somtimes is detrimental, so ig it depends what youre going for

i will experiment with it tn and see if i can get a better result or combine some methods 🙂

always room for improvement

all in all anything 240p and below is going to be difficult to bring back to life

[distant gate]

No beuno, all I saw was prenv log files, removed but no change in behavior

[mental geyser]

Try clearing browser cache and reload it

Or maybe I just gotta remake the file and give it

[distant gate]

Not a browser thing, something to do with the custom modules. Ive tried Edge and Chomre both have the issue

Ive gone through 3 or 4 programs now using ComfyUI so its possible I just have so much installed from other builds, version mismatches or something going on funky

Reinstalled all custom modules, still recieving the same error on loading either json

When attempting to process an input, when being given a valid input directory, receiving the attached error. Which I imagine is tied to the workflow not loading properly somewhere

[mental geyser]

oh lawd dont do that

youll break more lol

[distant gate]

We are 6 attempts late, I am one to brute force 😉

[mental geyser]

delete the folder, reextract from zip

and try this

[distant gate]

I will note PBRFusionPortable2.0\ComfyUI\PBRFUSION2.0BatchLoader.json loads without issue. I am also able to batch upscale without issue

The json you just provided still reports the same error on Load:

Loading aborted due to error reloading workflow data

TypeError: Cannot read properties of undefined (reading 'find')
TypeError: Cannot read properties of undefined (reading 'find')
    at nodeType.onGraphConfigured (http://127.0.0.1:8188/assets/index-CrROdkG4.js:1880:34)
    at app2.graph.onConfigure (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:80196:34)
    at _LGraph2.configure (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:60574:14)
    at LGraph.configure (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:80185:26)
    at LGraph.configure (http://127.0.0.1:8188/extensions/ComfyUI-Custom-Scripts/js/reroutePrimitive.js:14:29)
    at LGraph.configure (http://127.0.0.1:8188/extensions/ComfyUI-Custom-Scripts/js/snapToGrid.js:160:21)
    at ComfyApp.loadGraphData (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:80641:18)
    at async app2.loadGraphData (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:74290:18)
    at async app.loadGraphData (http://127.0.0.1:8188/extensions/ComfyUI-Manager/components-manager.js:771:9)
    at async reader.onload (http://127.0.0.1:8188/assets/index-Dfv2aLsq.js:81002:11)
This may be due to the following script:
/extensions/ComfyUI-Custom-Scripts/js/reroutePrimitive.js

[mental geyser]

is that the message that looks like this box?

if you just hit close, it doesnt work?

[distant gate]

Yep

I see the box yes, but let me retry with the new json a upscale

[mental geyser]

the batch loader file in /comfyui is a couple days older then the features in new

i think i had that b ox b4 too and just hit close

but i just extracted the zip and dont have it

[distant gate]

This bypass error again with the new json, Ill try re-extracting it

[mental geyser]

.. o you havent done that yet

[distant gate]

No of course tried the new json first, now Ill try rextracting

[mental geyser]

right the extraction

[distant gate]

Of course I need to redownload so will take time 😦

Chonky zip

[mental geyser]

updating the nodes, they like to mess things up i had to do extra stuff on top for compatibilty

you need to tell the models to be smaller

[distant gate]

https://tenor.com/view/shrinking-shrinking-time-shrink-tiny-getting-smaller-gif-13512377

[mental geyser]

#stillsmallerthencod

[distant gate]

6 minutes to download right now, 26 last night. Bandwith slower in the dead middle of the night eh??

[mental geyser]

Huggingface can be a little jank sometimes

[distant gate]

Alright rextract, problems resolved PBRFusionV2.0BatchLoader.json loads without error, upscales withour error. However I do have concern

This might not be a valid concern, but I did a 8x upscale on a 256 texture to get it to 2048. https://imgsli.com/Mjk2NDc5

Does this look like what you would expect for such an upscale?

[mental geyser]

You set seesr to 8x?

[distant gate]

Yes, which segways into another question. Is that the current correct location to tweak scale?

[mental geyser]

Then yes it expected going above 4x on seesr is detrimental to quality

We get to the higher res from kims span

If you run workflow as default it will output to 2k

[distant gate]

What is the upscale value set to by default

[mental geyser]

4

If you want higher res you can change image resize node to right of seesr node to 4096 4096

[distant gate]

Gotya alright looks fine now, thanks!

[mental geyser]

👍

[neon tulip]

I think maybe it was just a bad batch to compare, I don't think they were all 512x512, the map had custom textures I think. I'll run another one that have more 512x512 today

[neon tulip]

The image integrity is GREAT on PBRFusion 2 https://imgsli.com/Mjk2NDk3

[jade lake]

rawr

[neon tulip]

It stays very true to the original textures for sure

[distant gate]

This looks great!

[distant gate]

Although I can't help feel that in certain cases it wouldn't just be better to outright replace the texture. Such as in the case with these brick walls

As there are times where an upscale will contribute minimally to making an asset 'modern'

[primal olive]

some free pbr textures would do great for a couple of those

[neon tulip]

It looks like Fusion 1.0 was much more lenient to "create" what it imagined the texture would look like, while Fusion 2.0 moreso cleans up the pixelated look from the originals https://imgsli.com/Mjk2NDkw

[distant gate]

Only problem with 1.0's creativity was that it sacrified angular for a sloppy/curved creative view on textures. Where in cases it would likely not make sense

[neon tulip]

The pbr is STELLER, it all seems very accurate. The roughness maps on these are great

I know nothing about roughness maps, but it's making the default shiny stuff look not shiny, and shiny in areas where it should be shiny

[distant gate]

https://tenor.com/view/baller-ballin-shoot-basketball-for-gif-4728986

[rustic talon]

@mental geyser i think a simple but effective improvement would actually be to sharpen the textures a bit with unsharp mask. should lead to more defined PBR generation as well

[neon tulip]

The heightmaps are really impressive

[rustic talon]

ya, very impressive

[jade lake]

Ughhh stop smoothing the edges grrrrr

[mental geyser]

That height map wouldn't be hard to just make in Photoshop and would look alot better

[mental geyser]

I'll look into it

[burnt oyster]

so wait how do I install the PBRFusion

[mental geyser]

#1263287747905327236 message

youll download zip and unzip it hit run_nvidia_gpu

[wary forum]

@mental geyser How to make SeeSR faster?

[mental geyser]

you can try decreasing to 32

[wary forum]

And why i have 4k in input?

1k is maximum

[mental geyser]

what is the original size?

[wary forum]

for input

[distant gate]

This doesnt seem right https://www.youtube.com/watch?v=TuwmhN9v7_Q

[mental geyser]

ohhhhhhhhhhhh

it says "input is 4k" because thats what the 4x does it upscales via some other model and then runs it through stable diffusion

[wary forum]

💀

[mental geyser]

if you did size of 256 itd say input is 1k

[wary forum]

-vid

And if i don't want to say goodbye to my vid?

update is good (maybe), but how to make it work good on my PC?

[mental geyser]

try thisssssss Vovancho

[wary forum]

Already tryin

[mental geyser]

lol same for me ive never tried 32

ok 64 seems to be as low as you can go

like i said in update tho, the main problem is going to be vram for you i think i can get it to use about 5gb less but that will be a future update

[wary forum]

It's very slow(

i tried the upscale 2, but something wrong with texture...

But its very fast

@mental geyser how to change params to do only 2x?

[mental geyser]

you can do upscale as 2 in seesr node and then bypass upscale with model node

[wary forum]

only seesr?

thx)

[mental geyser]

wait!

i forgor

this one lol

[neon tulip]

🥵

[wary forum]

• CheckpointLoaderSimple 342:
  • Value not in list: ckpt_name: 'sd1.5texture.safetensors' not in ['dreamshaper_8.safetensors', 'stable-diffusion-2-base\text_encoder\model.safetensors', 'stable-diffusion-2-base\unet\diffusion_pytorch_model.safetensors', 'stable-diffusion-2-base\vae\diffusion_pytorch_model.safetensors'] 🫤

[mental geyser]

oope thats cuz i was experimenting

one sec

[dreamy pagoda]

How does the PBR Generation fare with textures that have sections with different painted colors? Say a metal sheet that us painted one color in the upper half and another ar the bottom?

[burnt oyster]

Thank you

[gritty dew]

@mental geyser any help?

ah nvm

it still launches

[gritty dew]

Well

[burnt oyster]

Im not ngl, i got no clue what im looking at in the PBRFusion, anybody know any tutorial for ts

[burnt oyster]

I fucked up

[gritty dew]

Just load the workflow file

I can't believe how great the upscale is here especially with the text every other upscaler couldn't get it right

[primal olive]

Put the original back or else

[primal olive]

Thx

[gritty dew]

Thought it wasn't the original

[primal olive]

Damn that's good

[gritty dew]

I got another one with text

that'd be hard to upscale

Original

Upscale

I'm very impressed, by far the best upscale I've seen handle that

I wonder if it can do the clock

Yeah no

Something else that'd typically become blurred

Best result I've seen but not 100%

[burnt oyster]

wait

which file is the flow work in order to fix the blank space

[gritty dew]

load the batch one

[burnt oyster]

and the file path for that would be...

[mental geyser]

Same folder as run_nvidia_gpu.bat, the root of the zip you extracted a file called PBRFusionV2.0BatchLoader.json

And then you gotta read the note boxes in the middle top of the workflow

[burnt oyster]

okay i got this, but that don't look like how it was when I first opened it

[mental geyser]

Cuz it defaults to a blank workspace

[burnt oyster]

okay

[mental geyser]

[burnt oyster]

one other quick question

is there a tutorial for ts

cause it isn't reconizing this

[gritty dew]

@mental geyser idk what I've done but it isn't launching anymore

[burnt oyster]

I'm guessing that there is an input file somewhere and an output somewhere, is just don't know what it is

[mental geyser]

Not yet

Paste in Path to input folder in "GIR Dir Picker" ( the node to the left of the green middle box ), and hit "Queue Prompt" on right side of screen

Outputs will be in (next to run_nvidia_gpu) ComfyUI/output

[burnt oyster]

thankyou

[mental geyser]

You done messed up a aron

[gritty dew]

Well I guess I'll just extract the file again

[mental geyser]

Easiest thing to do would be reextract the zip

Yuh

That'd be faster then reinstalling torch to python env lol

[gritty dew]

Also would you know what was causing the depthanything to throw this error #1263287747905327236 message

[mental geyser]

What GPU?

[gritty dew]

4070 Super

[mental geyser]

Node under red box did you select 40series file

[gritty dew]

yeah

[mental geyser]

🤔

Ok ig you gotta compile the engine yourself

One sec

Cool huggingface is down so we can't rn LOL

[primal olive]

i mean im downloading rn so

[mental geyser]

At least this models card is

[primal olive]

o

also 36 GAMER BYTES\

jesus

[mental geyser]

Chunky models

[gritty dew]

Hang on I think it is working now

[primal olive]

you get what you pay for they say

[gritty dew]

yeah it is fine now

[mental geyser]

I was gonna say nothing, but some of them are mine so they definitely cost something

[primal olive]

I'm looking to try it out to have a set of placeholder PBR textures until I replace them with my manually done/unique textures

[mental geyser]

Nice!

[primal olive]

@mental geyser just to verify, when running this it will create a new layer with the AI textures right

[mental geyser]

The rtxremix workflow creates one and manages it for you yes

It is a one at a time thing

Batchloader you're on your own

[primal olive]

say what now

[mental geyser]

Batch one you can batch 100s of textures, but it just outputs the 4 pngs to output folder then you have to ingest and assign

[primal olive]

oh that's not a huge deal

[mental geyser]

They are all appended with diffuse normal roughness height

[primal olive]

then I'd use that one community script to mass assign?

[mental geyser]

Yep 😊

[primal olive]

[mental geyser]

In future I hope to take in capture and do it automagicly

[gritty dew]

I'll see if i can modify the workflow to create a layer and ingest the textures. Although now I think about it idk how it will assign the textures

[mental geyser]

It will have to be some kind of custom integration with remixs texture craft

[pale field]

i could see a point where you could run this app and it'd do all the PBR upscaling stull, and then write a mod.usda as well

or does it do that in it's current state already, idk

[mental geyser]

That is what we're talking about yep 🙂

Will be much easier to do with the new nodes and API

But will still take a bit of development

[burnt oyster]

Crazy how yall are talking about updating ts to re-texture everything in a single capture and I'm still tryna find the "Green middle Box"

[burnt oyster]

okay thanks

I did smth wrong

[mental geyser]

The path you put in includes a search term? It needs to be a regular folder path

[primal olive]

says I'm missing nodes

[mental geyser]

?

[primal olive]

I got an error when I brought in the UI file. I get this which might be related

[mental geyser]

Where are you loading batch loader json from?

[primal olive]

the ComfyUI folder. Now I'm seeing that there are two in the base directory

[mental geyser]

Yeh those are old I gotta take em out srry

Use root ones next to run_nvidia_gpu

[primal olive]

what's the dif between these two?

[mental geyser]

So what we talked about, Batch can que up all in folder, remix is the click texture in remix hit que prompt it shows in remix 1 at a time

[burnt oyster]

👍

[primal olive]

@mental geyser is there a way to cancel the process since I forgot to increase the value for faster upscaling, 24 gigs of vram so

[mental geyser]

😦 I have not coded in a way to interrupt seesr you gotta close cmd window

And open it back up

[primal olive]

will it auto overwrite what's been done already

[mental geyser]

If files are same name yes

[primal olive]

what are reasonable values to enter here?

[gritty dew]

Someone who can actually use rest api and PBR fusion?

[primal olive]

for someone with 24GB of VRAM and 32GB system RAM

[mental geyser]

Change nothing other then what's circled, you can try 512 although I haven't confirmed if that's actually faster then 128

[primal olive]

if I maxed it at 1024 would my PC explode

[mental geyser]

Lmao probably not ™️ what you don't want to do is start using shared memory

[gritty dew]

So I found out what causes that depth anything error

[mental geyser]

O

[gritty dew]

if you run the prompt with the wrong Tensorrt selected you will get the error but if you change it back to the correct one after it will still throw the error

[mental geyser]

Ah I see

Do you have to restart?

Or just hit go again

[gritty dew]

hit go again

But restarting fixes the issue

[mental geyser]

Good to know thanks

[burnt oyster]

wanted to take a screenshot of how much resorces im using trying to upscale but my resorces are so maxed out I can't

[mental geyser]

Yup

[primal olive]

speaking of that lol

[gritty dew]

Yeah it is like that, It gets worse with the other workflow

[mental geyser]

Well you paid for the whole gpu

We're gonna use

The whole gpu

[burnt oyster]

nah im don't done paying my plan off yet

[mental geyser]

Lol

What GPU u got 8?

[primal olive]

considering this is Kim's upscaling model I'm surprised it's taking as long as it is. Dunno how this fancy AI stuff works tbh

[burnt oyster]

4070 SUPER

[gritty dew]

Can I safely change it to 64 because i have less Vram?

[mental geyser]

Yes

[mental geyser]

It only uses kims if the main one doesn't get to 2k by 4xing

If you do a 512x512 original image it won't use Kims at all

[gritty dew]

is the only downside speed?

[primal olive]

that would explain why a 64x64 input took about 40 centuries

[burnt oyster]

Yo I was getting 4 fps just moving my mouse, that’s how much ts is taking me

[mental geyser]

Cheev id do 96 pal

[primal olive]

btw for reference with the value set to 512 @mental geyser so far my VRAM allocation per operation goes anywhere from 6.2 to 9.0GB. In other words it could prob handle higher maybe

[mental geyser]

Oh God what result you expecting from a 64x64

[primal olive]

it's the batch I didn't choose it

[mental geyser]

Even a 512 image ?

[gritty dew]

Getting stuck on the height map with remix open can I change this?

[mental geyser]

Holy why is it 1024

Is that default?

[gritty dew]

yeah

[burnt oyster]

I’ve been doing this for about atleast 10 minutes and I’m only at 24%

[mental geyser]

O

You can do like 256 cheev

[mental geyser]

What was original res

[burnt oyster]

Huh

[mental geyser]

Resolution of the image you inputted

[gritty dew]

512x512 I downloaded the file from earlier

[burnt oyster]

Whatever default is

[mental geyser]

Default is the res of the image you put in lol

[mental geyser]

Remix being a hog

I say as this app is too

[burnt oyster]

I might be here for up to 10 hours

[gritty dew]

Well it wants 384

[primal olive]

maybe this would tell you something? lol

that was by far the longest one

[mental geyser]

Oh it's splitting into 64 tiles

That ain't right

[burnt oyster]

32 gb of ram might not be enough

[primal olive]

pretty sure it's a VRAM thing

[mental geyser]

No wonder vram util is low

[gritty dew]

Yeah 12gb of vram isn't a lot for this stuff

[burnt oyster]

My VRAM is at 11.5 gb in use

[primal olive]

another 512x

[mental geyser]

8 it really shouldn't be taking that long are you still on seesr step?

Ok auto you ready to change some code

?

[primal olive]

sure

[mental geyser]

Get ur vscode rdy

[primal olive]

it's gonna take a long ass time like this anyways

oh nvm

:)

I wouldn't be surprised if this output was just like the original CG render used

[mental geyser]

Ok go to ComfyUI/custom_nodes/ComfyUI-SEESR/node.py

[primal olive]

can I use some shitter notepad app

[mental geyser]

Yeh sure lol

[primal olive]

👍

[mental geyser]

Open dat

[primal olive]

it be open

[mental geyser]

Ok scroll down a tiny bit there will be args / parameters

[primal olive]

?

[mental geyser]

Yeh but I was wrong go down a bit more

[burnt oyster]

Oh shit I had my settings on 30 series

Might be why it was stuck

[mental geyser]

Perfect one sec

Change vae decoder tile size and encoder tile size to those right values

So 224 and 1024

Save file and then you gotta restart comfy

[mental geyser]

It shoulda errord out odd

[burnt oyster]

🤷‍♂️

[primal olive]

before and after. That look right? @mental geyser

note that the order is flipped for me, encorder is on top

[mental geyser]

Looks good

Also I don't even pass in process size I need to look into that 👀

[primal olive]

feeling adventurous so I'm gonna bump this too

[mental geyser]

Probs fine as it also does math in background if its too much I think

Hopefully that fixes the 64 tile thing

Or maybe it's because of the large latent tile size

[primal olive]

so far it seems much faster

I'm seeing VRAM usage jump as high as 16GB

[mental geyser]

What tile size is it doing

[mental geyser]

Like you showed here

[primal olive]

havn't gotten a 512x yet

[pale field]

that sounds about right, feed enough vram and anything will work

[mental geyser]

Why is it unloading models when y'all have unlimited

[burnt oyster]

Does it usually take this long to get past “SeeSRUpscale”

[primal olive]

that process repeats with every texture

[primal olive]

[burnt oyster]

I mean like I’m stuck at 24% again

[mental geyser]

What is cmd saying rn

[primal olive]

it stays at 24 until that step is done

[mental geyser]

Yeh the steps seesr does aren't integrated into comfyui loading bar

[burnt oyster]

It’s saying input size is 8192x8192

[mental geyser]

👀👀👀👀

[pale field]

bro idk, 24gb seems low

[gritty dew]

What

[burnt oyster]

That’s like 12k

[mental geyser]

So you input is a 2k texture?

[pale field]

bro that's 16k

[burnt oyster]

Idk how that’s possible

I’m using the suit texture from hl2

[mental geyser]

Which is what res

[burnt oyster]

Probably 1080p at the highest

Actually

[primal olive]

that's not a typical texture resolution

check it in your captures/textures directory

[gritty dew]

I Wonder how this would perform with a quadro 6000

[burnt oyster]

Because I used the one NVIDIA already shittly upscaled, it’s probably already at 4k

[primal olive]

DON'T DO THAT

use your capture textures as your starting point

[burnt oyster]

Yeah wasn’t my best idea

[mental geyser]

Lol

[primal olive]

for whatever reason one input was 4096x4096 and it had an our of memory error lmao

[burnt oyster]

It says it’s 80% complete, might as well see it through, right?

[mental geyser]

Auto if you want try editing run_nvidia_gpu.bat add --highvram MAYBE that will speed it up even more should keep things in vram memory

[mental geyser]

Lol sure, although you'll also have to wait for pbr gen cuz it won't show diffuse till PBR is done unless you disabled it

I don't think PBR gen will even run tbh at that res

[burnt oyster]

We’ll get to find out

[primal olive]

like this? @mental geyser

[mental geyser]

Ye

[burnt oyster]

If my pc don’t explode and it actually works you can thank me for finding out your Ai can generate 16k

[mental geyser]

👀

From a already upscaled 4k lol

[pale field]

i think it only works with low resolutions ones because it's restoring the image then upscaling

doesn't really matter at such ridiculous resolutions

[burnt oyster]

cmd says its doing smth

[primal olive]

only 256x textures but maybe this gives some insight

[burnt oyster]

It’s not maxing out everything I got for nothing

[mental geyser]

Still unloading interesting

Well there is --gpu-only lol

Not sure it'd speed up that much anyways tbh

It's a small model it's unloading

[burnt oyster]

I’m getting every error message known to man

[primal olive]

please don't upscale 8K textures

[mental geyser]

Not errors just lovely info

These guys must have been in a race

[burnt oyster]

Well not error message

Like

Usage warning messages

It’s not stopping though

After this I’ll try re uploading the 16k texture if it works

[mental geyser]

Jebus

[burnt oyster]

I’m risking everything in the name of pbr ai

I’ll also be the inventor of 24k if it works

[primal olive]

you should probably enforce some maximum input res because apparently my base game has some rare 4096x textures and this shit takes so long it holds up the entire queue

[burnt oyster]

@mental geyser I’ll give you some credit in the process of creating 24k

[mental geyser]

4x from 16k would be 32

[burnt oyster]

Shit

[pale field]

Bro whyyyyy

😭

[burnt oyster]

That night actually be a whole new creation bro

[mental geyser]

Nay it'd be 64?

[primal olive]

@olive cedar what's that one fact about 32K you tried to tell slask

[mental geyser]

0 chance your PC let's you do that lol

[pale field]

64k texture generation will break the unlimited vram limit

[primal olive]

I don't think GPUs can process 64K textures full stop

[mental geyser]

It'd break your ssd lmit

[pale field]

Probably gonna need a dozen rtx 6000 Ada

[primal olive]

remember guys, resolution increases are exponential

[mental geyser]

Jesus why does ssd autocorrect to ass

[primal olive]

32K is an obscene amount of pixels that would result in a file size that's brain numbingly massive

[pale field]

I mean for generation. But you can just barely view it in the image viewer. I remember downloading the Hubble space telescope deep field image and it was loading like the internet explorer days

[burnt oyster]

And ima have 64k

[pale field]

You weren't wrong, it'd break my ass limit either way

[mental geyser]

👀

[primal olive]

for reference a 4K texture is nearly 17 million pixels. A 32K texture is over 1 billion pixels

[burnt oyster]

It probably won’t appear on my screen as 64k but if we get ts on the Los Angeles sphere thing

[primal olive]

I don't think the LA sphere is that high res

yeah it's not

[mental geyser]

It's like what 192 quadros

[primal olive]

the LA sphere is 16K

[burnt oyster]

The 16k image is almost done

[pale field]

Runs on 150 quadro rtx 6000s

[burnt oyster]

It’s done I just gotta know where the files are located

[mental geyser]

ComfyUI/output

[primal olive]

Bro isn't even joking

Think about that lads

A 16K display requires 150 of some of the highest end productivity GPUs in the world

Working simultaneously

[mental geyser]

For video playback

[burnt oyster]

Okay I might have created a 16k image but I’m also dumb asf so ima need a file path location or smth

[mental geyser]

Do they have to turn off the outside screen if they are using inside?

[mental geyser]

So the run_nvidia_gpu file

In that same folder

Is a folder called ComfUI

And then In that folder

Is a folder called output

[burnt oyster]

not seeing it?

[burnt oyster]

that didn't load for me

[mental geyser]

Click into ComfyUI folder in that image

The top folder in that image?

[burnt oyster]

okay so its able to see that I already upscaled it

but if I ingest it using Remix I might be able to attempt it again

So cause it’s 16k remix isn’t recognizing it as a texture😔

Wait

Nvrm

I was on ingest models not textures

[olive cedar]

#asset-creation message

[burnt oyster]

It’s ingested the texture

[pale field]

I can only imagine the horrors of one single 64k texture

[burnt oyster]

I’ve officially started the Journey to 64k

[pale field]

Find a +16gb Videocard first

[burnt oyster]

Ya gotta believe my friend

[pale field]

It's gonna take 17gb vram to load a 64k texture

https://tenor.com/view/this-is-madness-300-pazzia-questaèpazzia-gif-18373678

[burnt oyster]

So it’s not impossible

Maybe to use it

But it’s rendering

Hella slow

But it’s doing it

It might stall. Out at about 32k

But even that don’t exist in a texture yet

So I’d still be the father

Ya gotta believe

If it don’t work, hope someone with a 4090 can finish my journey

[pale field]

I don't even think It'll make that texture size on that

[burnt oyster]

No one thought I’d make a 16k texture

But it did

I looked at it

It really looks bad because of how much Ai has been used before 16k

But none the less it’s 16k

Ahh

Looking at cmd it’s only generating at 16k again

[pale field]

Hard limit maybe?

[burnt oyster]

Maybe

Still

Are there any 16k textures available

Nah there are 16k textures

[pale field]

What for 😭

What is bro up to 💀

[burnt oyster]

It’s like just asphalt and shit

Ig im the creator of a outfit 3D texture at 16k

When it’s done

I’ll send you guys the file

Some with a 4090 and 128gb of ram can try

The steps are after reviving it, you must ingest it in remix, it’s the only way it won’t recognize it as already upscaled

Anybody with a 4090 and like 10tb of storage, a 7950X3D or 14900k, 128gb or more of RAM, you can attempt this

The only one you might not need is 10tb of storage

[pale field]

Maybe someone in VFX industry has that kinda specs

[burnt oyster]

Uh oh

[pale field]

Congratulations, the pc commited suicide

Good jorb 👍

[primal olive]

We warned ya (rather I did lol)

[burnt oyster]

It turned back on so we’re good