Self-Hosted AI Image Generation – InvokeAI

Try using AMD's Vega GPUs. HBCC allows you to use Ram as VRAM. Using Optane PMEM or just 4 optane NVME SSDs in a pool as swap space can get you enough VRAM to run LLAMA 405b, the full 810GBs of it.

I'd recommend a Vega 2 GPU like the MI60 if you are just using PMEM or ram, because PCIE 4.0 X16 is the bottle neck. If you are using Vega 64/56 or the Radeon 7 which is Vega 2 but for some reason doesn't run in PCIE 4.0, then optane NVME is just fine.
It's stupid, and totally insane to run such a big model, but token quality matters more than Tokens per second, so I made the jankiest setup ever just to say I could.
If I could run it on a raspberry pi, I would.

I am assuming you performed the image creation on your monster of a machine. What would be really interesting is having a watt meter running and take some readings of how many watts a particular image creation and/or editing consumed. I've read you can create images, but it may take hours on a small rig. You have a fast rig. So watt consumption could be a way of assessing and possibly estimating time when comparing your fast rig to something less.

Invoke is a great tool but pretty much all roads are leading you to use ComfyUI; hitting the token limit vastly affects the output of the generation.

Invoke has major limits with no support for increased batch sizes.

With ComfyUI you can take inpainting to the max with plugins such as Krita AI Diffusion – which is substantially more customizable than adobe's.

With new models like flux coming out that require 32gb of vram I fear that most people will not be able to host these models anymore locally.

Another cool thing you could check out would be Stable Audio Open which is trained on all the sounds from freesounds — it allows you to prompt the model to generate SFX

I understand the appeal of generating backgrounds and music. I've tried these systems, but haven't used anything in my videos thus far. I'm tempted to actually use it as well.

Video creators are already not immune. Some of my transcripts from my larger YouTube channel are in EleutherAI's 825 GiB "the pile" dataset that I didn't agree to. It's being used by companies like Apple, Nvidia, and Salesforce according to news articles I saw in July.

At this point I don't think there is any going back, so I'd rather see it advance quickly to reach real AGI. That way social and economic change must be made, but it is frustrating to see creative works targeted first because it's the easiest type of AI to create. In my opinion the current systems are more like an advanced form of a web search rather than an actually intelligence.

I was messing around with something similar to this a couple a of months back. I liked some of the results I got, but I still needed to touch them up in Affinity. One thing I really noticed is when you add multiple descriptors, especially with colour. So to keep things fantasy themed, a prompt like: "A very pale male Elf, with long silver curly hair, golden catlike eyes, and wearing a tattered blue robe, is standing in a vacant vandalised ally, under the midnight sky." it just wouldn't work, but if I give that to an artist, they'd be trying to picture what I'm trying to explain, focusing on those little details (and adding their own flair).

Adding something like this has been something I've wanted to add to my home lab. 2 question:

I have a Dell T440 running Proxmox with a lot of moderate speed cores, more than enough ram, but my Nvidia p1000 is passed through to a plex VM. Can that be shared easily enough between VMs or is a GPU even essential to doing this?

Whats the training process like? My sailing club has thousands of photos that I'd want to train it on to possibly use for marketing content. Can you train it to know difference between 2 different, but similar classes of boats? Formula 18 vs Hobie 16 if you feel like looking up what those are.

By nanoseconds, the average time for "I'm Jeff" is steadily increasing. If Jeff were immortal and continues Craft Computing into the far future, the intro will eventually take longer than the lifespan of a star. The very last sound ever heard before the heat death of the universe could become the nearly-endless, final F of our reality. Only then will Jeff finally be able to rest.

I need Jeff to ask for pictures of Spiderman with that AD voice

No one has a problem with AI until it starts to impact on their own situation. When image generators appeared a lot of musicans were really happy to use them for cover art for their music and had no problem with the fact that Artist's work had been used without permission to train those generators. But later on when Music generators appeared the Musicians were really angry that their music had been used to train those music machies. The same is true of writers who were happy to use AI Art for their book covers but not so happy to discover that their books have been used to train AI's to write books.

Content creators on youtube are now startng to get angry that their content has been used to train Video creation AI's and will probably get more angry when they find themselves competing with AI generated youtube channels where all of the content-including the people- is totaly AI generated.

AI could have been a positive force in our society by empowering the ordinary person- but you don't empower people by taking their work and using it without permission or payment in order to create machines that replace them, which is what is happening now. In the end the only people who are going to be 'empowered' by AI will be those who own and control it- almost eveyone else will rendered poweless and poor by this technology.

Orwells 1984 will look positively benign compared to the dytopian reality that an AI powered surveillance society will represent- not only will AI likely take your job, it will watch you 24/7 to make sure you adapt nicely and quietly to your new status as an unemployed non entity.

would be scary fun to mess with voice to text to translation to voice, so you could reach non-english audience easier 😀

Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword Buzzword

There, I fixed it for you.

invoke ai "is compatible with"……

Wait… You cant use own safetensor model files? im definately using my SD3. using it with automatic1111's GUI or with ComfyUI atm but ill check invokeai ui out too.
pretty…. "basic" tutorial for stable-diffusion but this will definately get you going.

I'm curious. Have you been playing around with lora's to see if it can give you some more clean results?

This is really great.

JUST STAY YOU. I know all the bs is a fad now, but what doesnt change is just be YOU.

use a layered png file for your backgrounds

Ubuntu 22.04 is actually still supported in the Ubuntu LTS lineup until 2026 under standard support, so any of my fellow linux bros would be wrong to come at you 🙂

That said if it works with 3.11 it probably would work with 3.12 but they probably have not tested it 3.12 is still relatively speaking new.

Michael Jackson on a horse?

convincing? i think not

26:28 Jeff has the "Disney Volume at home" too (Disney has a large screen room thing they use as a "better greenscreen" in a similar way, they call it "Volume")

15:43 As noticed also by other youtubers that dabbled in AI images (Shadiversity among others, on his secondary channel SHAD AI), the AI training on swords and swordfighting is very bad, so to get anything resembling a decent sword and a correct swordfighting pose you have to feed the AI sketches to use as a template.
The same for guns or any other specialist tool

Python is garbage, but you can run multiple versions concurrently on any version of Ubuntu using Virtual Environments which can be created and managed with Anaconda / Conda / Miniconda

Don't ever be tempted to hack a different version of Python into Ubuntu: some internal Ubuntu stuff requires Python, and because Python is garbage, it requires exactly the version that comes with Ubuntu.

Also dechutes gateway, ouch…, right in the beer belly.

Sponsor spots are normally what i skip over but i did appreciate the spaceballs reference "ludicrous intelligence". Well done!!