[Mr.Foxx] Commissions - Requests - Pinups - Short Stories

[yano2mch]

Yeah, I have Virtual Memory enabled. But that doesn't help much, in fact it kinda slows it even down bcz Drives are slower than RAMs.

Agreed. VM is fine for what it was loading programs or large blocks of data you may or may not be using. Especially under 32Mb, or with Windows 95 you had 8Mb of ram in many cases, and even Windows 3.xx the max memory you could access in 16bit was 16Mb from some 286 limitations while 1Mb was the 'real mode' memory access from the convoluted paging system.

You could set your Virtual Memory to something small like 8Gb, then you have some swapping but it won't be trying to swap a huge amount, rather only is used/touched the least. Or turn it off altogether if you can get away with it.

I will update my gig to 128GB sometime later, but for now I have to squeeze my workload in 64GB. I am also thinking to use my old RTX 3060 Ti 8gb which is just sitting in a box untouched. If I can connect it using a raiser cable, then I get additional 7gb VRAM to load some models into that during execution. Don't know the process right now, but I have read it someplace that ComfyUI supports MultiGPU.

Minimal experience with ComfyUI. But as long as the layers can be separated out or you are doing multiple models in a process that they can fit in each GPU, i don't see why not.

I think the troubles we are having will be resolved in the next 5 years as they will just start making the amount of Video Ram and Ram on the motherboard much larger. No longer are we in the time where 1Gb memory RAM or Video Ram will suffice, at least if you want the 'latest and greatest'. Games used to run on one core, then they were pushing for 2 cores. Now i don't know how many cores they want. 4Gb Video Ram used to be enough, but now they want you to have far higher. (though if they lowered the resolution they were targeting, texture and model sizes i'm sure you could still get everything needed within 4Gb, but for some reason companies are pushing hyper realistic and 4K/8K. No clue).

A note, i saw supposedly ComfyUI supports GGUF models, quanitized ones. So you might push for 2bit or 4bit models to squeeze more in the same space. I do similarly for LLM's mostly for speed since 70B models are so damn big.

But shared memory space models with a really large RAM i think is the way to go for these systems in the near future. That or they offer GPU's with 64Gb of Video Ram.

 

[Kretua]

Sorry to butt in, but ComfyUI cannot natively work with two GPUs (no CUDA can). So if, for example, you have two 12 GB VRAMs in your machine, that is not the same as a 24 GB GPU. It can only distribute work between the two GPUs. So if two workflows are running, it distributes the tasks between the two cards. But within a single workflow, it will only use one GPU to calculate an image.
Otherwise, Comfy also uses GGUFs.
As for the direction of development, I see that GPU manufacturers, especially Nvidia, do not intend to produce larger VRAM cards, precisely so that you cannot do meaningful work in a home environment, let alone run LLM. For that, you need to buy a dedicated system for thousands of dollars. This is clearly illustrated by the fact that, for example, the affordable 16 GB VRAM cards all have only a 128-bit data bus, which is ridiculous, since speed is precisely what VRAM is all about. Even today's high-speed DDR RAMs slow down when the machine swaps. So if you use AI at a hobby level, you either have to be patient and make compromises, or pay for a GPU farm online. (Which, if I'm not exaggerating, is much cheaper than buying a 24 GB card, not to mention a professional 96 GB one.)
Or, at a professional level, you invest a larger amount and hope that it won't become obsolete within a year.
Best regards.

 

[Mr.Foxx]

Agreed. VM is fine for what it was loading programs or large blocks of data you may or may not be using. Especially under 32Mb, or with Windows 95 you had 8Mb of ram in many cases, and even Windows 3.xx the max memory you could access in 16bit was 16Mb from some 286 limitations while 1Mb was the 'real mode' memory access from the convoluted paging system.

You could set your Virtual Memory to something small like 8Gb, then you have some swapping but it won't be trying to swap a huge amount, rather only is used/touched the least. Or turn it off altogether if you can get away with it.



Minimal experience with ComfyUI. But as long as the layers can be separated out or you are doing multiple models in a process that they can fit in each GPU, i don't see why not.

I think the troubles we are having will be resolved in the next 5 years as they will just start making the amount of Video Ram and Ram on the motherboard much larger. No longer are we in the time where 1Gb memory RAM or Video Ram will suffice, at least if you want the 'latest and greatest'. Games used to run on one core, then they were pushing for 2 cores. Now i don't know how many cores they want. 4Gb Video Ram used to be enough, but now they want you to have far higher. (though if they lowered the resolution they were targeting, texture and model sizes i'm sure you could still get everything needed within 4Gb, but for some reason companies are pushing hyper realistic and 4K/8K. No clue).

A note, i saw supposedly ComfyUI supports GGUF models, quanitized ones. So you might push for 2bit or 4bit models to squeeze more in the same space. I do similarly for LLM's mostly for speed since 70B models are so damn big.

But shared memory space models with a really large RAM i think is the way to go for these systems in the near future. That or they offer GPU's with 64Gb of Video Ram.

Now you are showing off, I don't understand that gibberish

Anyway, there's this AMD Ryzen AI 9 HX 370 and other new upcoming CPUs have larger VRAM baked in. So the future is good.

 

[Mr.Foxx]

Sorry to butt in, but ComfyUI cannot natively work with two GPUs (no CUDA can). So if, for example, you have two 12 GB VRAMs in your machine, that is not the same as a 24 GB GPU. It can only distribute work between the two GPUs. So if two workflows are running, it distributes the tasks between the two cards. But within a single workflow, it will only use one GPU to calculate an image.
Otherwise, Comfy also uses GGUFs.
As for the direction of development, I see that GPU manufacturers, especially Nvidia, do not intend to produce larger VRAM cards, precisely so that you cannot do meaningful work in a home environment, let alone run LLM. For that, you need to buy a dedicated system for thousands of dollars. This is clearly illustrated by the fact that, for example, the affordable 16 GB VRAM cards all have only a 128-bit data bus, which is ridiculous, since speed is precisely what VRAM is all about. Even today's high-speed DDR RAMs slow down when the machine swaps. So if you use AI at a hobby level, you either have to be patient and make compromises, or pay for a GPU farm online. (Which, if I'm not exaggerating, is much cheaper than buying a 24 GB card, not to mention a professional 96 GB one.)
Or, at a professional level, you invest a larger amount and hope that it won't become obsolete within a year.
Best regards.

Yeah, maybe. I don't know exactly, but I read it somewhere that we can allocate Vae and clips on one card and the model on the other and share the load between two GPUs, there are some ComfyUI nodes that do this it said.
And if that's legit, then it's nice to have the feature.

 

[Kretua]

Yeah, maybe. I don't know exactly, but I read it somewhere that we can allocate Vae and clips on one card and the model on the other and share the load between two GPUs, there are some ComfyUI nodes that do this it said.
And if that's legit, then it's nice to have the feature.

Yes, it's perfect for that. To share the load. But a 24GB model won't fit into two 12GB cards because it only handles VRAM separately. Yes, there are multi-GPU nodes. But you've made me curious, I'll look into it.

 

[yano2mch]

Now you are showing off, I don't understand that gibberish

Sorry Was just meaning doubling your effective ram was worth it... when you had very little. And the Ram hungry apps didn't come til later (Photoshop eating 64Mb... of the 80 something meg limit for Win95).

But too much swapping is what makes the system chug. I've seen to too many times, where my system seems to suddenly halt because it's doing so much swapping it isn't doing much processing. Even in linux using ZRam (compressed module) it chugged quite a bit as usable ram vs stored compressed blocks become smaller and smaller til it just.... stops. Or you can wait 2 hours and hope it gets past the minor hurdle so you can kill the program and free memory.

Anyway, there's this AMD Ryzen AI 9 HX 370 and other new upcoming CPUs have larger VRAM baked in. So the future is good.

Mhmm I can hope. I remember integrated graphics back in the 90's where i could select how much video memory vs physical memory. That does seem to be the future. Sides motherboard Ram is cheaper so having more of it is easier.

Yes, it's perfect for that. To share the load. But a 24GB model won't fit into two 12GB cards because it only handles VRAM separately. Yes, there are multi-GPU nodes. But you've made me curious, I'll look into it.

Sounds promising. I may have a lot of legacy and technical knowledge, but this new stuff... i've only scratched the surface. When the hardware is affordable and i don't feel like Windows blows as bad, i'd get more to see what i can do. I certainly hope the future is bright.

 

[TulXXX]

GIF - "Luk Pat Req Aug 2024 by Mr.Foxx"

 

[Kayonshee]

GIF - "Luk Pat Req Aug 2024 by Mr.Foxx"
View attachment 1922494

amazing job!!! here is a other link with a story if you want to bless that tread with your work!!

Thread '[Garak3D] A Desire in the Family'

Dec 13, 2023

This is the new Garak3D comic I was creating, I'll go up to where it left off and I hope someone has information and can continue it

...

• ElCastigador
• incest milf mother
• Replies: 976
• Forum: 3D Comics and Artworks

• 
• 

 

[Deleted member 52101]

By MR.FOXX View attachment 1784899View attachment 1784900

I wish we could get too see more of those two hot girls together

 

[Deleted member 52101]

By MR.FOXX View attachment 1666137

some of the best work from Mr. Foxx!

 

[Mr.Foxx]

Some of my old animations.
GIF conversion would be bad, so download the attached file MP4 files.

 

[Ryze]

By MR.FOXX

 

[Ryze]

By MR.FOXX

 

[TulXXX]

GIF - "Kaloryes PatronReq(Mar 2024)" by Mr.Foxx

 

[Ryze]

By MR.FOXX

 

[Ryze]

By MR.FOXX

 

[yano2mch]

GIF conversion would be bad, so download the attached file MP4 files.

I've seen on other sites where they say it's a gif but it's actually a webm file with a gif extension. Though doing a quick conversion and test i don't see the site being happy with it as webm, or renamed as gif, or webp. That's too bad.

Kinda sad the textured pixel pattern they encode for gifs is really bad usually; I'd prefer to go smacker instead as it does a much better job at selecting a good 256 color palette. But it's an obscure codec (unless you're into games and you see Smacker/Bink logos everywhere). The RAD game tools seems to have a bug with conversions, but workarounds do seem to work.

Though mp4 with HEVC or AV1 encoding works really well too. I'd be happy to encode to AV1 for you.

Here's trying to do a gif with smacker's results.

 

[Mr.Foxx]

I've seen on other sites where they say it's a gif but it's actually a webm file with a gif extension. Though doing a quick conversion and test i don't see the site being happy with it as webm, or renamed as gif, or webp. That's too bad.

Kinda sad the textured pixel pattern they encode for gifs is really bad usually; I'd prefer to go smacker instead as it does a much better job at selecting a good 256 color palette. But it's an obscure codec (unless you're into games and you see Smacker/Bink logos everywhere). The RAD game tools seems to have a bug with conversions, but workarounds do seem to work.

Though mp4 with HEVC or AV1 encoding works really well too. I'd be happy to encode to AV1 for you.

Here's trying to do a gif with smacker's results.

View attachment 1957580

I don’t usually upload things here, but I’m considering sharing both a GIF and a clean MP4 version of my results, since that’s my usual output format. This way, viewers can either download the MP4 or just check out the GIF for a quick preview.

 

[yano2mch]

I don’t usually upload things here, but I’m considering sharing both a GIF and a clean MP4 version of my results, since that’s my usual output format. This way, viewers can either download the MP4 or just check out the GIF for a quick preview.

Mhmm. I'd say a clean mp4 would be preferred, since then converting i don't have to fight with the dithering. And if i AV1 encode it, it would be like a meg or less; While gif previews can be nice gif really wasn't intended for trucolor videos of any length. Different age, different requirements.

The downside though since the gif files would be so large it's almost laughable using them as a preview (14Mb gif vs say 4Mb mp4). But with the limits of what they let for attachments, it's hard to work around. I'd say lower resolution gifs would greatly reduce the size as a preview (probably 720x480). Beyond that whatever you think is best.

 

[Professor Sinister]

I find Animated PNGs make a much better output than Animated GIF since they do support TrueColor. But the file size can be prohibitively huge so at that point, AV1 is the way to go