Help installing a local AI

[rinaldo00]

So I need advice on how to install a local AI for text to image and then in the future text to video to create short clips. I have never used any AI tools before and then I saw an ad for the Ace Browser with built in AI. I started trying out Nova which is part of the browser to do some text to image and I really enjoyed it. It was great at giving suggestions for a nice story.


But I am tired of the daily quotas and the restrictions on not allowing anything risque so I would like to run something like it locally. Do you have any recommendations?

Here are my specs:

Windows 11
AMD Ryzen 7 7800X3D 8-Core Processor - RAM: 31 GB
NVIDIA GeForce RTX 4090 - VRAM: 24 GB


Thanks

 

[M76]

Is it 32 or 16GB RAM that you have since you listed it twice?

If it's 16GB then don't even think of running text to image, let alone text to video models locally. Even 32GB wouldn't be enough.
Minimum 64GB is recommended if you want any decent models to run. Maybe some cut down models can run with 16GB, like WAN2.1 1.3B, but even that's uncertain.

 

[rinaldo00]

Is it 32 or 16GB RAM that you have since you listed it twice?

If it's 16GB then don't even think of running text to image, let alone text to video models locally. Even 32GB wouldn't be enough.
Minimum 64GB is recommended if you want any decent models to run. Maybe some cut down models can run with 16GB, like WAN2.1 1.3B, but even that's uncertain.

My mistake, it is 32GB. I also corrected my post.

 

[rinaldo00]

I read that these were good for text to image (SD 1.5, SDXL) / Flux . Here are their requirements:

The Three Models at a Glance​

	SD 1.5	SDXL	Flux Dev
Release	August 2022	July 2023	August 2024
Parameters	~860M	~3.5B (6.6B with refiner)	12B
Native resolution	512x512	1024x1024	1024x1024+
Minimum VRAM	4 GB	8 GB	12 GB (quantized)
Architecture	UNet	UNet (larger)	DiT (transformer)
Text rendering	Poor	Poor	Good (95%+ single-word accuracy)
License	CreativeML Open RAIL++	CreativeML Open RAIL++	Non-commercial (Dev) / Apache 2.0 (Schnell)
Status	Deprecated, huge ecosystem	Active, maturing	Active, growing fast

SD 1.5 is the Honda Civic of image gen. Cheap to run, parts everywhere, gets the job done. SDXL is the midrange sedan: better in every measurable way, still affordable. Flux is the sports car: noticeably better output, but you need the hardware to match.

edit: fixed the table

 

[M76]

I read that these were good for text to image (SD 1.5, SDXL) / Flux . Here are their requirements:

The Three Models at a Glance
SD 1.5 SDXL Flux Dev
Release August 2022 July 2023 August 2024
Parameters ~860M ~3.5B (6.6B with refiner) 12B
Native resolution 512x512 1024x1024 1024x1024+
Minimum VRAM 4 GB 8 GB 12 GB (quantized)
Architecture UNet UNet (larger) DiT (transformer)
Text rendering Poor Poor Good (95%+ single-word accuracy)
License CreativeML Open RAIL++ CreativeML Open RAIL++ Non-commercial (Dev) / Apache 2.0 (Schnell)
Status Deprecated, huge ecosystem Active, maturing Active, growing fast
SD 1.5 is the Honda Civic of image gen. Cheap to run, parts everywhere, gets the job done. SDXL is the midrange sedan: better in every measurable way, still affordable. Flux is the sports car: noticeably better output, but you need the hardware to match.

SD is the OG model that started it all, it's quite outdated by now. Flux dev is pretty good, it also has a smaller version Flux klein which is also decent (still far better than SD or SDXL) another good one is Qwen image, not sure about it's minimum requirements.

But choosing a model is one thing, first you'll need an environment to run it in. My preferred choice is ComfyUI. ComfyUI comes bundled with a bunch of workflow templates, that you can use OOB. All you need is to download the required models. It will tell you exactly what to download and where to put it, even provides direct Download URLs for each component. But if you want risque things you'll need modified models or loras, the official models are all "safe", which doesn't mean they sometimes don't produce NSFW output, even undesired. But I'd get comfortable with the basics before getting into all that.

You can try the official installer: https://comfy.org/download
But I'd recommend going with this one instead: https://github.com/Tavris1/ComfyUI-Easy-Install
As the latter comes with a ton of addons already preinstalled that you will want anyway if you are serious about this. Or you can try with the official one, then install the "Easy-Install" version separately later as that is completely stand alone.

 

[rinaldo00]

SD is the OG model that started it all, it's quite outdated by now. Flux dev is pretty good, it also has a smaller version Flux klein which is also decent (still far better than SD or SDXL) another good one is Qwen image, not sure about it's minimum requirements.

But choosing a model is one thing, first you'll need an environment to run it in. My preferred choice is ComfyUI. ComfyUI comes bundled with a bunch of workflow templates, that you can use OOB. All you need is to download the required models. It will tell you exactly what to download and where to put it, even provides direct Download URLs for each component. But if you want risque things you'll need modified models or loras, the official models are all "safe", which doesn't mean they sometimes don't produce NSFW output, even undesired. But I'd get comfortable with the basics before getting into all that.

You can try the official installer: https://comfy.org/download
But I'd recommend going with this one instead: https://github.com/Tavris1/ComfyUI-Easy-Install
As the latter comes with a ton of addons already preinstalled that you will want anyway if you are serious about this. Or you can try with the official one, then install the "Easy-Install" version separately later as that is completely stand alone.

Thanks a lot! I am off to try and install the easy install ComfyUI !