nice take, plz enjoy this downvote.

5 to 10 steps is still plenty turbo, that's like a 75-90 percent reduction from what is typical.

1 step in a Turbo model is = to 5 steps in a regular model. So your "1-4" is only equivalent to 5-20 steps. If you cant deal with running 6+ steps to get 30+ step equivalents, then perhaps you are playing with the wrong toys and need to seek another hobby.

same here

@SG_161222 That PR has already been merged into dev, so it might be easier for most people to just run the dev branch rather than manually merge diffs. Or just wait for the next release.

@SG_161222 can you explain how to download this? it links to a raw txt file

@SG_161222 thank you!

From my own tests, it's just a name thing.

@Kyra31 It looks like the file sizes are different between the non-turbo version on huggingface and this one. 6.5 GB here and 6.94 on Huggingface

Well, no, not really. I rather like it :-)

Turbo is for fast inference ... some need quality, some need speed. Don't dictate others needs.

If you're getting "very poor" quality, this is almost certainly due to operator error. I've used this model to produce images of such good quality that people thought they were real photos (even with my sad AMD setup). This is not to say they were perfect, but if they can fool a lot people, the quality is definitely not "very poor." Check your settings: Make sure you're using the right sampler, proper SDXL resolutions/aspect ratios, low CFG and low step count, and turn on hiresfix. Good luck.

@Kyra31 I understand this. I need to download the 'non-Turbo' version, but no matter which download link I use, I end up with the Turbo version.

@epzer No, that's not it! When I say 'the image quality produced is very poor,' I am comparing it to the original version. The image quality generated by the Turbo version is indeed much worse than the original. Unfortunately, I am now unable to download the original version! No matter which download link I choose, I end up with the Turbo version.

@hazc138405 there was already another reply to you explaining that you need to look at the end of the file names to know which model you're getting. They are all labeled "turbo" at the beginning because that's the main title on the model card. I know it's confusing, but this has already been explained. I am getting great images and there are a ton of high quality images generated with the turbo model shown below it. What is more likely, that you're making some sort of mistake or experiencing some technical issue unique to you, or that we are all hallucinating? Chill and find somewhere to ask for help instead of complaining at the creator.

That's you my dude, i have tested it with many different prompts and the quality is good.

I thought the same but look at the black porche example , the settings, if you adhere to them, the results are very impressive and FAST!

Inpainting can generate images the same way the normal model can be used to inpaint. But it's not as good. It should be used when your are going to inpaint.

@diogod Thanks for the reply!

I also mentioning that to authors .... and they are pissed in spite of saying not true to people ....no comment

You need to be more specific in your prompting, default will be an average

Same seed?

same problem here. random seed. is there no prompt to generate random pictures? or at least get more variation? i tried (brown or black hair) but it always brings back brown hair... :(

Same here, but it happens to me with all SDXL models: always shows a very close portrait. The only thing I can do is ask for a "full body portrait" and put "torso" in the prompt.

I think you downloaded the turbo version, try with the other one

Download the turbo model, use CFG 2 and steps 6 or something around those values. Check out the latest prompts in the reviews or my account, model needs some refinement, but already pretty good.

Using Dpm++ SDE works best. That said, its randomly blurry. I can only assume its a turbo thing, some things are very sharp, but most is soft, unusable.

@SG_161222 oh right! thanks. I had no idea

It works for me, but I created the 8-bit model (via Manage models)

If all you are going to tell people about your problem is "doesn't work" you are very unlikely to get any help. Do you get an error message? What is it? Are you expecting something to happen that doesn't? What? Is something happening that you aren't expecting? What?

@jbboncemore2 Nothing happens when it finishes importing it

@Eloyse thx, Imma try that later 

@ImWion step by step how i did it:

1) download the model to icloud, not in the app since civitai doesn't allow direct dl no more
2) import the model via 'Select from File ...'
3) confirm import
4) test the model, it worked in my case but I prefer the 8-bit version for better memory management
5) convert the model to 8-bit
6) test and compare, the results are comparable, just a tiny bit different
7) delete the original (bigger) model

Yeah I'm also confused. I already downloaded the turbo version, now I wanted the "normal" version with baked vae but the filename still says turbo?

@SG_161222 Perhaps the "V3.0 (BakedVAE)" checkpoint shouldn't contain the word "Turbo" in the filename. Just a thought.

Same here, I'v put (open mouth:2.0) three times in the negative prompts, and (mouth closed:2.0) in the positive, and every image generates with lips parted. How do others get closed mouth images?

Got the same at first, what helped me was "Teeth XL" Lora here on Civitai- I use it with 0.4 weight, helps somewhat :-)

I'm hoping that v4 (if it comes) addresses face and hands as first priority - but we'll just have to wait and see, because baby has first priority!!!!! (rightly so)

Link:-

https://civitai.com/models/128371?modelVersionId=140535

I didn't try it with this model but you can try "parted lips, teeth visible" in negative. At least it works with other models

Are you using the right sampler? For Turbo, you need to use DPM++ SDE Karras.

Interestingly enough, after messing around for quite some time now i found out that, when i use hires upscaling, it does reliably add pubic hair to both images, before and after hires. If i don't use it, it's hit and miss, and miss for the most part. The wierd ways of AI i guess.

It is much easier to add pubic hair with inpaint, especially in a full body pic

I'd go with these approaches:

- pick a photo you already have and like, use tagger/captioner like BLIP to see what keywords you might be missing

- if you find a photo that was generated by someone else, just check their generation info

- try to find a relevant LoRA or, if there isn't any, try to train one yourself

It's not a dumb question. But you will get different answers from different people. Some people think turbo is just as good as non-turbo others think its not. Personally, I much prefer non-turbo - it suits the way I work better. But no doubt, someone who loves turbo will come on here and say, turbo's better if you do it right. I would suggest you try both and see which works for you. Sorry if this is a crap answer!!

@Raken cosign

Turbo is faster, but somehow loses some quality, but is better for people with worse GPUs. So people who pursue quality will not choose Turbo, but both versions have their reasonable existence value. For example, I pursue a few seconds to get 2~4 results, Turbo is just needed for me

No, the recommendation is always to use SDXL Base as the base for a LoRA training. For inference later on you would use this checkpoint as a base then.
So your LoRA stays universal and will work with other checkpoints (e.g. V5 in the future) as well

@SG_161222 I tried using this as a LoRA but with very limited results. Maybe it requires different samplers or CFG settings, but I'm not that of a pro to test this efficiently.