IIRC CFG = classifier free guidance. CFG is what is used in Comfyui, I think it was different in A1111/SDN though. Comfy is worth swapping to though, unless you're a settings junkie. Cuz comfy has nothing for settings, but WAY more configurable on the front-end.

@Lazman I have never used anything other than Comfyui.

In my workflow, I use the Fluxguidance node where we set the guidance. I don't have a cfg node unlike when I was playing with sdxl

@diaserikcg114 I still use SDXL cuz I'm not rich, so can't reasonably afford a 4090. 4060ti is what I set my cap at. I paid 2k total for my entire PC (and use my 65" TV as the monitor). I Can run flux, but given the extra resources required for it, I think there's better potential with SDXL. Especially if I can manage to perfect my lora training and train some top notch loras with 1536 res images.

Besides, there's a lot more resources for SDXL, since it's been around longer, and a lot of creators are still hardware limited. Also easier to get SDXL working than flux, since ya don't need extra text encoders, or to sign(click) any legal contract on the hugface site in order to use it. Honestly, even SDXL is still catching up in some areas. There's certain resources I noticed are out for SD1.5 but not for SDXL. like the handfixer controlnet(at least, I haven't found one for SDXL, yet).

@Lazman I started AI with a 1050ti, I understand ;-) I had time to save up to change my PC and buy a super 4080.

Since I use this PC for work, I cut back on all other hobbies. I think many will understand ;-)

I admit that I hesitated for a long time to switch to Flux: sdxl is more flexible to use while Flux only has the quality of its image for the moment.

I thought that would change quite quickly but I am increasingly considering going back to SDXL and only using Flux in img2img to finalize my work.

@diaserikcg114 Yea.. I'm loosely considering trying flux for outpainting, cuz SDXL is not doing it good atm. But there's a good chance I just don't have my settings right. To be fair, I'm also trying a lot of things at once, so it's a bit chaotic. I'm attempting to refine my controlnet setup for directly turning 2D flat images into full 3D/photoreal. Also testing a workflow with 36 prompt windows connected to 6 switch boxes (RGThree node), that way I can easily swap between pre-defined prompt setups.

Like, one for visual settings(scores, realism, visual effects, etc.), one for scene (city, sci-fi, caverns, etc.), one for character/s, one for actions, one for clothing/jewelry, etc. But, sadly, when I tried to run it, I got some error about the controlnet not being iterable. So, I'll have to look into that. it might work without controlnet, but I haven't tested yet, and I got lost in testing new controlnet models after I got the error.

Apart from that, also testing IP adapters(to a lesser extent). Still gotta learn how those work, and see if/how I can tie them together with controlnets(and which ones) in order to maximize control over my workflow/s.

And when I'm not doing any of that, I'm attempting to streamline my settings in Onetrainer. I just attempted to use prodigy instead of Adamw for the first time when attempting to train a lora based on Dundas square in Toronto. Failed miserably. 142 epochs, and the 20-50-142 epoch versions won't generate anything to do with the input data, so, prodigy may be crap, or it's possibly to do with the fact that I didn't caption well enough, or the fact that I boosted the lora rank over 16, up to 64 that run. (and 32 on a prior run with it).

@HailoKnight hype!

@HailoKnight Might be worth verifying, as the site itself also reports no file attached. :heart:

@HailoKnight yay!