AI Generated Images

So, I'm not going to write up a "this is how to install and start ComfyUI and view it in a Web browser" tutorial -- like, lots of people have done that a number of times over online, and better than I could, and it'd just be duplicating effort, probably badly. You'll want to look at them to get ComfyUI up and running, if you have not yet.

My understanding is that if you install ComfyUI and load a workflow -- which I've provided -- and you're missing any nodes or models used in that workflow, ComfyUI will tell you. That's why I attach the workflow JSON file, so that people can use that to reproduce the environment (and should be able to reproduce this image exactly).

The workflow is compressed with xz and then converted to text for posting here, since the Threadiverse doesn't do file attachments on posts, with Base64. On a Linux system, you can reconstitute the workflow file by pasting the text in the spoiler section into a file (say, "workflow.json.xz.base64") and running:

$ base64 -d <workflow.json.xz.base64 | xz -d > workflow.json

There's an extension that I use that makes it convenient to download nodes in ComfyUI: ComfyUI Manager

https://github.com/ltdrdata/ComfyUI-Manager

It works kind of like Automatic1111's built-in extension downloader, if you've used that.

I think -- from memory, as I'm on my phone at the moment -- that I'm only using one node that doesn't come with the base package, SD Ultimate Upscaler, which is a port from Automatic1111 to ComfyUI of an upscaler that breaks the image up into tiles and upscales it using an upscaling model; it lets me render the image at 1260x720 and then just upscale it to the resolution in which it was posted. If you install ComfyUI Manager, you can use that to download and install SD Ultimate Upscaler and whatever else you might want in terms of nodes. Also provides an option to one-click update them when they get new releases.

I don't think that it can automatically download models, though it'll prompt you if you don't have them. I believe -- again going from memory -- that there are two models used here, the NewReality model derived from Flux, and the SwinIR 4x -- I probably should have used the 2x model as I'm scaling by a factor of 2, changed that in my most recent submitted images -- upscaler model.

You can google for either of them using their name, but NewReality is on civitai.com, which you'll need an account on:

https://civitai.com/models/161068/stoiqo-newreality-flux-sd35-sdxl-sd15

There are several versions of the NewReality model, based on various base models. You want to click on the right version, the one based on Flux 1D.Alpha; the default when you go to the page has the SD3.5Alpha version checked. Then click "Download".

That gets dropped into the models/checkpoint directory in the ComfyUI installation.

..I don't remember exactly where I got my copy of the SwinIR upscaler model, think it was from HuggingFace:

https://huggingface.co/mirroring/civitai_mirror/blob/ee44205453f9a8f125b0acf496e2ed13ad074498/models/SwinIR/SwinIR_4x.pth

This is the homepage of the SwinIR upscaler, should also work, though looks like they use different filenames:

https://github.com/JingyunLiang/SwinIR/releases

The SwinIR_4x upscaler gets dropped into the models/upscaler directory in the ComfyUI installation.

You need to restart ComfyUI after adding new models for it to detect them.

If you haven't used a Flux-based model before, it can be slow, as a warning, compared to SD models that I've used.

If you've never done any local image generation before at all, then you're going to need a GPU with a fair bit of VRAM -- I don't know what the minimum requirements are to run this model, might say on the NewReality model page Civitai site. I use a 24GB card, but I know that it's got some headroom, because I've generated at higher resolution than I am here. Hypothetically, I think one can run ComfyUI on a CPU alone, but it'd be unusably slow.