I can answer 1 and sorta answer 4.

1) The refiner does not replace a VAE. Latents made by either the base model or the refiner must still be decoded with a VAE. The refiner is fed the latent of the base model. If you don't want to use the refiner you just decode the latent of the base model with a VAE instead of passing it to the refiner. In ComfyUI you can easily do both at the same time. From my experience, the refiner improves some things and makes others worse. You can just see the refiner as a different checkpoint, rather than part of SDXL.

4) Personally it was using around 11GB of VRAM for 1024x1024 images, but I'm assuming the answer is with the same tricks that allow people to run the 1.5 models on 2GB currently (a moment of silence for my friend generating with a 750TI).

@poiGenAI Thanks, that clears some things up. I got kind of confused by the VAE not being mentioned in the flowchart.

1.) I use the refiner model in img2img as a checkpoint with the SD Upscale script and .25-.35 denoising and seems to work fine.

2.) Can't Answer 3.) can't answer.

3.) I used about 10GB VRAM with generation then about 14GB-16GB VRAM when upscaling/refining. I only have a Quadro P4000 8GB and it didn't crash. Took forever to render (about 25 minutes for 2048x2048). But it got there in the end without crashing.

From what I read on kohya's github, it's nearing 16 for LORA and still in the 40s for fine-tuned models.

issues of rights...

Also non official ones were fake.

Stability have an embargo on release until the 18th July. And even then, what's officially released might not be these models. We'll put out updates!

Assuming it will be possible to fine-tune it on consumer cards...

LoRA training is possible though.

smh give rivals catalyst to push the baseline consumer VRAM say 16 to 20 gigs in the future... maybe

13gb IMO is overkill when most models are 1gb to 2gb and models geared toward celebrity are only like 150mb.

I thought that the devs changed their mind about nsfw.

Don't forget that 0.9 is supposed to be mostly for testing. It's 1.0 that I'm waiting on.

@olternaut latest I've heard (7/17ish) is that it isn't trained for nsfw stuff. but honestly it doesn't matter. just look at all the sd15 based models that have new data trained into them. training new nsfw data into the sdxl model won't be difficult and we'll be seeing a ton of nsfw model versions soon after release.

@rjox I don't think so, otherwise nsfw content would already be present on version 2 of SD

This is what happens when Open Source community starts virtue signaling so much to avoid the fake backlash started by large companies.
While the open source community self-censors, removes copyrighted materials from the training, avoids NSFW, censors any "offensive" speech, etc. making open source AI (from LLM to AI Art) move backwards, large companies like Open(Close)AI, Google, Adobe, Meta, etc. are ingesting copyrighted materials by the billions, allowing all kind of kinkiest NSFW (as long as it's real, because it becomes "empowerment"), and make their solutions the defacto standard while profiting from both the monopoly and exclusives rights to monetize their solutions at ripoff prices.

It's sad, because SD started as an awesome idea, we could have had some awesome solutions with SD2 and more recent. Now we're just asking "how bad" are the new models compared to the previous ones, smh.

@rjox but does that making the improvement going nowhere? if you trained using old, lesser quality from 1.5 synthetic data, then expect the negative impact even dubbed as latest version, its generating ability would be stuck in the past

ONLY SDXL VAE CAN BE USED