@Zavy thank you !

@Zavy Awesome, thanks!

I also appreciate the pursuit of quality over speed. I do wonder though if Turbo might be able to produce higher quality when run at the same steps as a regular SDXL model (and not just at 1-4 steps). The SAI paper suggests it can achieve higher quality than the base at 50 steps. I tested some of the first turbo merges and I think it's possible it can be better. But if the licensing is more restrictive then that settles the decision anyhow.

I don't need to copy paste. I don't need your workflow. But like every other good model out there, you should have at least a few flagship examples to show. Together with recommended keywords, especially recommended negative prompt to get the best out of the model. Anyway, it is just a suggestion, no need to be touchy.

@Zavy If you look at my join date, you can see I'm a new user and I'm in the process of learning generative AI. Of course it is allowed to copy prompts. People copy my prompts also. That's the reason for those prompt sharing websites exist, testing and learning. I have thousands of images in my pc, I share only a few here, they are experiments. So if you do not care about people using your model and love to be greedy, why are you sharing it here? Why do you use this website? You can just play with your friends in your shitty discord channel. No need for this childish behaviour.

@Zavy You are the immature one here. I'm sure I can find tons of copied prompts in your images also, but you cowardly hide all your prompts. At least I'm being honest. And you are the one throwing false accusations without proof. You talk about courtesy but you just show how arrogant you are.

@Zavy I have decided not to use your models. No point in arguing any more. I'm sure other sensible users will also boycott your models because of your crappy personality.

@nexus99  Why? only lazy and unimaginative people need to see the prompt and gen data. The days of sharing the whole prompt and workflow are over and have been for two months. Just be grateful that @Zavy is still offering models to download for free. Those days may end too as gen AI keeps moving forward. I've already seen a lot of creators move their models to patreon and other private places. Why bite the hand that feeds you?

@nexus99  wow! Did you bump your head or something? You don't make any sense at all.

@nexus99 The reason Zavy no longer provides the prompt is due to this: https://civitai.com/models/119229?modelVersionId=169740&dialog=commentThread&commentId=288878

And Zavy is not alone. Many people got accused of "false/fake prompt" when novices could not replicate the image (because they don't have the correct ComfyUI workflow and only copy & pasted the prompt).

So some people just decide that it is north worth their trouble to provide anything at all.

Thunk out your own prompts, dude.

Yes, at least we can see a sample of negative prompts (as generic one) for this model to start with. Some model requires this negative prompts but nowhere it mentions this model need negative prompt.

@nexus99 You are right. This is the only model in CivitAI that don't have sample negative prompt. I always use this as the starting prompt for my ConfigPresets extension. 

@codegix  It seems the users with no shared images, no ratings, no models are the ones who make bad assumptions. The creators do not owe you anything. You are lucky to get anything for free. Can you read? Then do yourself a favor and read the description where it is explained why no prompts are included. TLDR because of beggars like you that can't figure anything out on their own.

@Zavy Thank you. I didn't know that SDXL negative prompt is not useful. I think that is enough to understand I don't really need the other positive prompt.

@EricRollei21 He has been doing a great job providing the best model, and I genuinely appreciate his work. I don't intend to offend anyone here, but I feel like you have to be chill for insulting me being a 'beggar' without questioning why I asked in the first place.

I was working with the State Plugin, and it requires a 'GENERIC negative prompt' recommended by the creator. Typically, almost all creators in CIVITAI provide these prompts, either in the description or within sample images. I found all of them except his and I know somebody would ask for the same thing and nexus did. It's not hard to write negative prompt but even if I wrote a good -ve prompt it might not work perfectly for certain models. I could randomly choose a negative prompt, but as a creator, he likely understands better what his model is capable of. While it may not be crucial in life and can be experimented with some random negative prompt, having the creator's recommendation is the best. It would have been beneficial if he had mentioned, as he did before, that negative prompts are not very useful in SDXL. This is new information to me. Thus, his response is more informative compared to yours for not being chill. Relax bro.

Ps: I just realized he did write a good description about using the -ve prompt (less is better). He did provide a sample prompt in the description which normally I took it from the sample image. I think this is enough and solved the problem. So, this is just the problem of me for not to read all the description till the end.

Thank you @Zavy.

@codegix Thank you for your calm response🙏.

Things can easily turn into a rude shouting match on the internet 😁.

@codegix Funny how you managed to read the comments but not read Zavy's description, and you had so much time to make a reply here. Next time use it better and read the model creator's words first. Without them freely sharing AI would be so much more difficult or impossible.

@NowhereManGo You are safe from me at least since you've posted numerous and wonderful images. I only attack those that never shared even 1 image but criticize the hand that feeds them. I'm still amazed that people are sharing models for free, and that Civitai allows anyone, even unregistered to download for free. But I can't imagine it will stay that way. I really want to protect the creators since without them we wouldn't have much to work with.

@EricRollei21 I share your sentiment regarding Civitai's model creators.

It is wonderful that so many creators (you among them) share their models for free with us, and for that they deserve our gratitude.

Have this problem with some DPM variants, depending on scheduler. DPM (whatever) + Kerras didn't have the problem. Euler A didn't have the problem.

@Dribbl Hey, it was a broken VAE issue.
Join Discord to discuss further.

Cheers ! 🥂

The explanation is simple. Base SD1.5 is quite unusable except by using word salads. So every SD1.5 model has to be overtrained to get half decent result. There is a high cost to this overtraining, though. These SD1.5 models tend not to follow prompts well, even compared to base SD1.5. SDXL prompt following is way better than any SD1.5 model.

With SDXL, the base model itself is already highly fine-tuned. So most “photo real” models that have mostly pictures of people in their training set will change base by a much smaller amount, hence the smaller variation you see between models, specially for “photo style” prompts.

If you want more variety, you can try one of the non “photo realistic” models, such as Paradox 2, StarlightXL, SDXL Niji SE, and Unstable Diffussion Yamer. These models have been optimized for “non photo realistic”. You can still generate photo style images with them, but you will see that the deviation from base will be larger.

And of course, you can get variety by simply varying your prompt, by describing hair color, eye color, etc. and get variety.

One BIG advantage of the fact that most SDXL models deviate less from base is that LoRAs tends to work very well across all models.

Having done a lot of experiments making/merging Turbo models, the flexibility of the model is a ton lower, they respond less to negative prompts, and do styles much less effectively. They'll produce realistic people in contemporary/modern settings just as crisp if not better as SDXL, however.