I see the same complaint (chin cleft) in the flux world as well, so not just a pony issue.

@Cyberdelia looking at the example images on this page, i stand corrected about *every chin having a cleft. Cyberdelia is tops.

@barry9000 Same. I counter with slim round face in the prompts

@Cyberdelia yeah I have both and tested both with identical settings. 6.1 works but 6.2 gets weird textures and colors.

It might not be the model, but unfortunately I don't know enough abut any of this to guess at what else it could be.

@automirage04995 Tried using a VAE?

@arnorwing Yes I used sdxl_vae_fixed

think of it as a flavor of XL that is trained to react to anime and manga type stuff (research "danbooru" tags, concepts etc). so load it the same way you would load an XL model. if you use comfy, you can try downloading images on civitai that have been made with pony and if the workflow is in the image metadata, then a workflow will appear in comfy if you drag and drop the image in there.

Pony models are SDXL models based on a model called Pony Diffusion, originally designed for anime/manga/fury/my little pony images. Turns out, it's a very obedient model, responds well to prompts, and turns out consistently good images, so now there are dozens of variants and merges. You use it just like a regular SDXL model, with the following notes:

- Add "score_9, score_8_up, score_7_up" to the positive prompt. On some versions, you might want to experiment with adding "score_6_up, score_5_up, score_4_up" as well. The original training images still had these ranking tags attached to them, so the model responds to those keywords.

- Use a straight-forward positive prompt. Pony models tend to perform better without a string of terms designed to make them more photorealistic. You can usually get away with terms like "cinematic," "film still," "epic fantasy," etc., without much problem, but adding things like "raw photo, 50mm, 8k, 16k, hdr," etc., tend to override the model and give worse results, unless your particular version of the model was specifically trained with those terms.

- Keep the negative prompt as simple as possible. This takes a delicate balance. No negative prompt, for some reason, tends to create washed out colors, but if you add too many negatives, the image suffers.

- Use a relatively high cfg, around 7-8.

- In A1111, use a clip skip of 2 (or, in ConfyUI, use a "set clip last layer" of -2).

- The sampler/scheduler/steps don't make much difference compared to the regular SDXL.

Other than that, lookup the original Pony Diffusion model on Citivai and read the entire description for more information on how to use it. It responds well to danbooru tags, it recognizes a ton of anime/cartoon characters by name, and it responds well to descriptions of poses/situations. There are also "rating" tags that can be used to specify the level of SWF/NSFW content, and "source" tags that specify the art style. You can read about those on the Pony Diffusion page.

You absolutely NEED the positive and negative prompts alongside to your own prompt

@VWGolf Thank you for the advice! I've already copied the positive and negative prompts along with all the generation details like seed, steps, etc., but I still can't achieve the same result as the image. I'm using the same model, cyberrealisticPony_v62.safetensors, on Web UI Forge. Could there be any other parameters or settings that I'm missing that aren't included in the image details?

@Crispitina Forge generates different images from A1111 or Comfy for the same seed. Forge (or A1111 for that matter) also does not import LORA information or additional settings if they are not part of the pos/negative prompt (i.e. in Civitai plug-in or Foocus metadata). Finally you might not have all the LORA or embeddings. Review the PNG Info for any resources and settings that Forge or you have missed.

You have a great model that can generate more images than there are planets in the known Universe. Just create. Don't spend your day trying to reproduce an image someone else did. That's nonsense. Just experiment ;)

Thanks for your answers guys.