I could say the same thing with EPS and still made good images with enough effort. What WebUI and settings do you use for generating images ? Perhaps you can ask OP and others to help you if you are stuck with something instead blaming V-pred not liking you.

@libright I had the same problem even though I tried other workflow settings. Period. This is a good checkpoint, but V-pred is not for us, dude.

@libright And if it's good, why not instead share your own settings or workflow, where you can easily add the loras without any extra effort, rather than starting from scratch. Training was in the same situation and it never worked because they were only focusing more on the network online.

@libright I dont use WebUI anymore and it color likes always yellow had Vpred on or off. Also Vpred AUTO had ON. So, if this yet got yellow, I dont using this. EPS not got same problem. So, what problem, if we also got EPS version yet? V7 had also EPS plus Vpred.

@RolRyu i added an article with the workflow inside of it to download it. It doesn't have a lora loader but you can add it easily to make it workable for your needs.

https://civitai.com/articles/19111/sdxl-txt2img-workflow-in-comfyui-v0357

@libright Sorry bro, but that was worse. Good side: it gave me some tips. So, thanks for that. We'll stick with this Randomizer workflow style for now, because everything is visible at the same time.

Now there's just another problem: When it scales, its color becomes too bright. EPS not do the same. :)

@RolRyu That's strange how the colors gotten too bright for you but at least good to know how its working for you right now.

@libright Yea. That why we yet staying EPS, because not got same. Later maybe can sharing, if want checking. Warning: It's big and possibly cringe, but it we seeing all options, you know. Also Ksampler after it scaling and Face fix and more.

Guys, to use V-pred models you need to enable it in A1111, in comfyUI you can just use a node called ''ModelSamplingDiscrete'' and pass the model through it - Choosing V-prediction. It's super simple and not a mistery. Instead of commenting a quick google should get ya'll fixed up. Aside from that, there's no difference. Usually lower cfg works better for v-pred models, but you can still leave most, including this mode, luminar, which is awesome by the way, at euler_a normal cfg:6. The biggest advantage of v-prediction is to get more variation and prompt adherence sometimes at the cost of strong styles or some lora supports.

@Sythria Tried everything, but I'm still not satisfied with it. Still asked for EPS. Is it that hard?

Luckily there are others here. So, I guess this will be the last one for now.

@Sythria eh. v-preds still have so many issues with adetailers and img2img and even hi-res.