It is trickier for me in this model, but the prompt "Rear view" paired with " Looking out of frame" gave me the look I was after. I might post one or two images of that run later.

same query

Are you using the "Facial restoration on" setting in Adetailer? I make sure to turn that OFF for all turbo models/checkpoints. Even at an extremely low strength, face restoration makes Turbo output really smooth, as you said.

@Socalista Thanks! The face restoration in Adetailer was turned off. There must be something else that messes up Adetailer

@Socalista I can confirm this issue with Adetailer is only affecting Turbo models. The facial restoration in Adetailer was off, however....

@jpgranizo I've been able to use Adetailer with my turbo merges/models just fine. I have all the settings set to default except the denoising strength (which I change as needed) and the separate width/height set to 1024x1024. I keep my inpainting mask blur at 4-6

@jpgranizo Perhaps your inpainting mask blur is too high? That'd smooth it out.

@Socalista I can confirm that mask blur is low (at 4) and I changed the separate width/height to 1024. This didn't change anything regarding the Adetailer issue.

Here are my Adetailer settings: https://i.imgur.com/sQsrmRD.png

Here are a few examples I uploaded. Left is without Adetailer and right with Adetailer. What is your opinion?

RealvisXL 3 Turbo:
https://i.imgur.com/PisIdpA.jpg
https://i.imgur.com/ET1JXpQ.jpg

REalvisXL 3 Non-Turbo
https://i.imgur.com/tyZZdgI.jpg
https://i.imgur.com/j8uco63.jpg

@jpgranizo I don't see much of a difference in blur/smoothness between turbo and nonturbo, they're BOTH becoming blurred/smooth. What denoising strength?

@Socalista I agree, both becoming smooth. Adetailer inpainting denoising is 0.4. These are all my adetailer settings: https://i.imgur.com/sQsrmRD.png

Any idea what is going wrong here?

@Socalista II experimented with varying step counts in Adetailer inpainting alongside the Turbo model. Here's a brief rundown:

- Primary generation: I used 8 steps in the primary generation and 8 steps for 1.5x upscale with Hihres.fix.

- Adetailer: Tested with inpainting step settings at 8, 12, 15, 20, and 30 steps. Here's the outcome:
https://i.imgur.com/jpkPiZe.png

Observations and Questions:

1. I think the 20 and 30 step results more closely match the original image in terms of skin texture

2. Should I even use 20-30 steps in Adetailer when paired with a Turbo model, as merged Turbo models work with 5-10 steps?

3. Is the improvement of skin texture at higher steps genuine, or is it just additional noise that gives the impression?

Download it from hugging face

Okay. Thank you.

+1

why not share this chimera model with community? it could be interesting :)

@HMS_Phone Seconded. Model/Pics or it didn't happen.

@HMS_Phone That is what people do to get attention. They are lonely. There is no model.

maybe i will, after a few more merges that dilute some of the ones i merged in that have "do not share merges of this model" in their descriptions. like the HPI1.4 model, for example. when i left this same comment on that model, the uploader reminded me that i'm not supposed to share merges that include HPI1.4, according to its own description, (and then said something about potential legal action by the people who assembled some of the resources he used to make it? i don't know, man)

this really isn't rocket-science, i'm just stealing the best models i can find and smashing them together.

if you really wanna replicate my "basically exact" results, go ahead and merge the only model i've uploaded with the basic Turbo model, then merge it back onto itself (so, in total: merge my model and turbo; merge that result with my model again) in order to maintain fidelity and overwrite some of Turbo's jankiness, and you've basically got "the oh-so-mystical merge" that i was referring to. if you want it to be better, find a better model and repeat that same process, with a preference for ones that have already been merged with turbo. that's all i do.

A+B=1; C+D=2; 1+2=temp_model; temp_model+my_model=model_A; model_A+my_model=model_B; test model_A and model_B to see which is better. rinse, repeat.

8-10 steps, cfg 1-2. test different samplers to see which work. if you're lucky, even the dppm_3m or 3m karras options will work

Chill. 😂I'm literally giving you a workable recipe and 90% of the logic I use to select my merge candidates

I'm not saying it's perfect. But I really frickin' like the results.

Thanks for the tips friend!!!!!!!!! Now i a doing miracles!

Which model did you test?

You talk about RealVisXL but then you mention realvisxlV30Turbo_v30TurboBakedvae on your website, which behaves quite different to RealVisXL V3.0.

The turbo versions are supposed to be used with a low cfg and step count.

It's really bad with lora's for me for some reasons compared to dreamXL or SXXXL for instance.
I also get blurry everything except the face when not using lora for some reason?

This worked for me as well. Faster, and looks better.

use A1111.

Inpaint now.