I found that if you use adetailer with a 1.5 model such as JernauMix or RealCartoon-Realistic paired with a realistic embedding/LoRA of a person (real person or not real) it does a far better job of capturing the likeness of different faces to get rid of the generic 1.5 face. Just be sure to change the adetailer ckpt model to one of the ones I listed previously and make a separate prompt in adetailer with the person embedding/LoRA you want to use, Then add all of the improvement words such as "best quality," detail words, skin tone, makeup words, etc. that you used in your main prompt and any other words that you want to add to change things up. Keep negatives and all the other settings the same. I recommend doing this in A1111 through inpainting. I personally get better results that way.

@ZyloO If you’re talking to me I haven’t had the chance to test it yet but I will tonight and let you know 👍 Thanks for sharing your model!

try my pony-model ;) (and oc you could call the nationality, european, asian, ...)

you can try my "better face model v3"

and if you want real REAL good small faces as simple read my article

https://civitai.com/articles/5002/face-swap-with-adetailer-and-reactor

@ZyloO ... you get not a new face with my lora if you call a specific nationality ?

and faceswap, yes you must choose every image a new face.

i dont work with ComfyUI, but there iam shure you can create some kind of random nationality every image ...

but maybe SD3 is better on that ;)

Please also add MALE face variety in future updates. Thanks!

@ZyloO Amazing to hear thank you, I'm always hesitant to download big models like this but I'll grab this one in that case! :)

it can't do any m2m scenes

That's a problem with a lot of pony models sadly

me too...

try "realistic-eyes" (hyphen is important!) in positive. should fix your issue easily

@themancalledmojo999 Oh thanks. I will try it :)

gayyyyy

Yeah, we definitely need more facial features than that one and only poor man's Chloe Grace Moretz face that plagues every pony model.

any version of DreamshaperXL you like in particular? or just newest?

@Lyser “v2.1 Turbo” should do you right.

@sdvww1110wpww223 what's your settings for the refiner and for this model steps/sampler wise?

@BobTheBurner 

Based on the reactions to my posts, this seems to work well: Steps: 45, Sampler: DPM++ 3M SDE Exponential, CFG: 4 - Refiner switch (dreamshaperXL_v21TurboDPMSDE): 0.6-0.7. ADetailer with `mediapipe_face_mesh_eyes_only` can be helpful depending on how the model feels for a particular generation.

whats tencentarc?

@ZyloO I'm using the same prompts in the samples provided here and getting the oversaturated images. I have to remove BREAK or reduce the tokens to below 75

@ZyloO Ohhh I thought it was a Lora at first glance lol. D'oh!

Oh, @ZyloO , thx man. I didn't know that Pony is a different thing. Through that this is just SD XL. Good to know.

@ZyloO "Oh, and one more important note - these tags are less useful in negatives as you can only go as low as 4 and ideally we should train on quality all the way down to 1, but it makes the training significantly more expensive. So you can still use score_4/5 in negative but they will not push you away from really bad images."

Since the Pony model was never trained on score 1, 2, and 3, using them in the prompt shouldn't make much of a difference. I'd recommend running an experiment on this to confirm.

You mean these?
https://civitai.com/models/330790
Which version are you using?

@2buckchuck yes thats the model, I use the 23.5Plus, the new one works also but older is a bit better me thinks :)

@megaderp thanks for confirming! I'm definitely getting superior images with mklan235plusxxx as a refiner. I need to figure out how to mitigate the slowdown of switching models - I think adetailer is the bottleneck.

For anyone looking for more skin realism I recommend 'low' hires upscaling (to 1.1) using 1x-ITF-SkinDiffDetail-Lite-v1, then following your usual upscaling method with that output.