Merged with several anime and then several real pony mixes
If your gens look great up until the very last step and then look pixelated you need to use an XL Vae. There are no vaes baked into this. Use
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors
Description
FAQ
Comments (20)
a lot of pony mixes don't have very deep color when photorealism is aimed for, but this has real potential. I can suggest that you can mix SDXL1.0 Loras into the model with Supermerger :)
Good checkpoint
Could you include some example prompts, please?
I thought I did, but for the images I used:
marvelous digital photo by a professional photographer score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up close-up of a hot woman wearing cyberpunk clothes
just change woman to man instead for a man
@Darkarma Thanks! And no special settings?
@Jellai Nope
I am not able to generate any usable images as all of them have massive dots and artifacts (no LORAs being used and the simplest prompts).
That's more a problem with Pony in general than this specifically. Try avoiding using weights above 1 on words or phrases (loras are fine with higher weights). At one point there was an add-on that fixed it but I can't remember what's it was called.
I was having the same issue, the only way I was able to fix it was to enable the Refiner option and choose another checkpoint like AutismMix Confetti at 95%. Something is wrong with this model as it breaks the images in the last 5%.
@jobobby04 Never had that issue myself where its the last 5% unless you using a non-xl vae.
@Darkarma I am using default settings in Automatic1111 WebUI. So no VAE is loaded unless it has the same filename as the checkpoint
@jobobby04 https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors I use this for the all pony checkpoints see if this fixes it but its not baked in @GhostlyGaze
@Darkarma Thanks, getting the VAE loaded alongside the model fixed the issues.
The same for me, tried to used it but get only images with dots and artifacts, while it looks brown
There's no vaes baked in, you need to manually select one. Or download one, I recommend this one. https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors
@Darkarma Thanks !
Oh nah. This checkpoint ain't getting enough buzz. It's one of the best Pony Models I've used. Imma have to start posting some examples of what it can do
For me all i get are solid black images. Tried multiple options from a fellow AI generator and nothing helped. put the vae in the vae folder, put it in the checkpoints with a clever rename to get a sd vae dropdown and checked the arguments. no luck
Are you using --no-half, --no-half-vae or --disable-nan-check? These are the usual modifiers which fix this issue in Stable Diffusion Automatic. See here for more info: https://www.reddit.com/r/StableDiffusion/comments/17nsx3e/image_generation_looks_fine_but_final_output_is_a/
Gotta talk to one of the civitai staff about making this one available for on-site. Looks great :O