@Cyberdelia Nah, i can prompt XL based celeb loras to perfect likeness on Pony. Make a try please.

@Cyberdelia Good enough.

@Cyberdelia i am waiting for the same tbh, currently using this model as a refiner with pony real as base

@kvsh88 Do you use Forge? When you switch to refiner?

@michaelarguelles275 Yes I do

@Cyberdelia Some images have been hidden based on moderation preferences set by the creator, view all images using this resource.

Straight over Gallery pics.

@Cyberdelia You set up no XXX?

It's a general problem of SD/SDXL engine. Try to use ADetailer/inpainting for faces.

@Cyberdelia thanks for advice.

@mphobbit @Cyberdelia I installed the ADetailer extension and downloaded the face_yolov9c.pt model. Can I use the default settings?

For upscaling, I’m using the 4x_NickelbackFS_72000_G.pth model with a 1.5× scale, strength set to 0.3, and 15 steps.

@Cyberdelia I didn't try upscaler)

@brunobarbosa_96517 face_yolov8n is good for most checkpoints, but try another variants to find the best one for your tasks.

@mphobbit @Cyberdelia thanks for the answers 😁

@Cyberdelia Thats not optimal, only if you have a powerhouse. 25 steps, 768p enough.

Anyway, its more faster, just use detailers, fixers, or prompt good(extremely detailed xy, 8K, etc)

@brunobarbosa_96517 as an experienced A1111 and ADetailer user, just want to chime in a couple points to help you here, just in case.

- You mentioned you are using the "c" model for face_yolov9 which is, for nearly any purpose, the correct one. While Cyber mentioned the "n" version. So if you saw that and switched to "n", you can go back to "c", you had that right.

- Yes, default ADetailer settings are generally pretty good. Sometimes moving mask blur up to 8 or 16 can help with visible boxes, or if the entire face gets badly over-textured and warped or blurred (which might happen in very close-up shots), then bring that back down to 4. Sometimes face_yolov9 simply won't cooperate, particularly in extreme close-ups - you can try to use person_yolov8m-seg in it's place.

- Actually I use person_yolov8m-seg for nearly every single subject in every single photo. It's a great ADetailer model. If you start using multiple ADetailers on an image, make sure you've got them ordered sensibly. For example, if you do face_ first and then person_ afterwards, person_ will undo what face did.

- Get the "SDXL VAE fp16 fix" from here: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main . It's really the only VAE you'll end up needing for SDXL. Don't get confused on his model page there - you want the one specifically called "sdxl.vae.safetensors" - and you might wanna rename it to help you distinguish the file once downloaded. Throw it in your VAE folder and reboot A1111 and you'll be able to select it. Then you can just leave it selected forever.

- For ADetailer on SDXL, keeping it around 0.4 denoising is generally good. Only very rarely will I get up to 0.5 for a specific case. You also can't go too low or the mask blur will become a problem, since the denoising won't overcome it. You can mostly leave it parked on 0.4 denoising and not worry about it.

- "Inpaint only masked padding, pixels" is the most important setting in the whole thing, after denoising. Basically think of this setting as the painter's "context window" - what it can see as it's doing it's work. You can really move this slider around to achieve different things. For the most part, you'll want to go higher than the default setting, so that the painter "sees" more of the image as it's painting, which results in better overall blending (at the cost of some detail, which can be counteracted with slightly higher denoising - for example, if I max out padding all the way to 256, I might raise denoising to 0.42 or 0.45).

- Another setting to consider is mask erosion/dilation. You'll generally either leave it as-is or go bigger (dilation). Basically it just makes the area actually being painted larger. This can be helpful for blending and for capturing additional things in the ADetailer run - a prime example is that if you increase the mask dilation significantly during a face_ run, you can often get any dangly earrings or necklaces or all the hair in the run, which again, will help with blending.

- All of my tips are based on a 1-megapixel image (1024x1024 or other standard SDXL resolution) that is then upscaled 1.5x via hires.fix. If your image is smaller at the time of ADetailer run, you'll need less padding and less dilation. If larger, you should probably just max out padding for sure, and be aware your dilation setting will have less overall reach.

- Try the Siax 200k upscaler, it's pretty dang good.

- Last point, don't use "Use separate width/height" as suggested by Cyber if you're upscaling with hires.fix first. That'll hurt things in that case.

@shapeshifter83 is face_yolo9c the best for faces?

@imaginaryrealms yea. unless you have VRAM limitations or have some other specific in-depth reason, always just use the largest and most up-to-date of the models.

@shapeshifter83 Thanks, it was very enlightening. I hadn’t read that—I only got the notification a little while ago. I tried to understand as much as I could with Google Translate. I just downloaded the VAE; I didn’t know I needed it.

@Cyberdelia I simply have the SDXL VAE fp16 fix selected in the A1111 webui at all times. While selected, it applies to all phases of initial generation or any inpainting, including ADetailer phases. Remember, ADetailer models are detection models, and don't have any painting input themselves - it's all still coming from the selected checkpoint and prompt.

There's almost never any need to remove it as the selected VAE - only in the odd case of a checkpoint that has a non-standard VAE baked-in, which is rare, and I've only ever seen it with very niche anime-focused models, whereas I generally only play with realistic-focused stuff. The only exception for myself is that I've found the Sharp Spectrum VAE (https://civitai.com/models/455138/sharpspectrumvaexl) to be useful in certain situations where I need strongly increased contrast. But 99% of the time I'm using SDXL VAE fp16 fix.

@Cyberdelia hey man, I didn't mean anything negative with this comment but just a reflection of my lack of skill for prompting. I am trying to get the model to generate multiple character and multiple clothing items (open tops with bra underneath) and then trying to put those items in a less natural position (bra lifted or bra below breasts) and I am struggling with those. So any help would be greatly appreciated!

@Cyberdelia thanks!

@Cyberdelia I know, how embeddings work, i just cant save them the same way as loras(weight plus prompt)

@Cyberdelia I alway add plus prompts, trigger words to my characters. With embeddings there is no option to save trigger words.

Have to agree with you. 5.6 is giving me better results. Just post a a comparison