Yes. Forge supports loading some number of models. At least newer version of forge. There's a top bar for these additional models next to the "Checkpoint" bar at the top.

In there you must add:
CLIP
T5
VAE

To use Flux models that do not have these built in.
Most models do not have it built in, as it adds a lot of extra space to each model, and it's considered better to keep them separate as you can then re-use the same models.

Compared to SDXL and previous model generations, when these parts were small enough, so the standard was to bake them into the same model file.

NudeTouch The VAE is incorrect. The VAE you linked is the Stable Diffusion 1.5 VAE. The others should be fine. I think this is the file you want: https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/ae.safetensors

It is but what you make of it.

Narz I hear you. It's very hard to say "what's different". It's generative AI. I don't know for real what's changed or not. Which is why I started sharing the sources I am merging with.

My process is this: I take the most recent version, and I look at models released by others and try to see features that I like. Such as an improvement to "Flux Face", or more realistic skin, or stronger male + female NSFW, improved nipples was one thing I tweaked in a recent version. It can be all sorts of things. I look at other models, test them, and try blending them into my models at various degrees.

Oftentimes it's just for overall beauty/style and vibe.

With Flux, since it is still under-devleoped in NSFW, and still often produces body horror, I'm always trying to improve the NSFW capabilities in general.

I spend a couple of hours doing this, and then do a bunch of test generations to see if I'm happy with the model, and if I am, I call it the next number iteratively.

When the change is large and significant, as in, a big step up, I will increment with a major number, i.e. 2.3 to 3.0.

In the case of Persephone, it's based on Fluxmania to some percent, which really helped stabilize males in the model. Most models (mine as well), are overfit on female inputs / training data. This one added a lot of stability back to the model in form of males.

MysticXXX is a key pillar to unlocking NSFW. It's the most stable and well-trained model in this regard. So I often blend in a little bit of it for balance. Like maybe 10%. So I use "SFW" checkpoints which has un-ruined bodies to calm the model down, and then I blend in Mystic as a spice to... spice things up.

This explains the process, but it might not be the answer you are looking for, and if so, I'm sorry but I don't have a better answer for you. I do vibe-merging.

With Flux, testing is so slow and painful, I don't run several thousand generations to review compare specific capabilities as a test suite, which would be great, and I would do and automate if this was done at a professional level. i.e. always using the same prompt and seeds, to see how different aspects developed.

Sure, I usually do Q8 and Q4, would that not do?

ronwee84 Not sure, but Q4 should probably work. But the higher quality you go, the better the results are on average. Much more body horror on lower quants.

I have a Q6 version done. So whenever CivitAI actually lets me finish an upload without breaking along the way, it'll be online.

Q6 uploaded. That took about 16 hours for just the uploading...