Best preset settings for Pony model?

[mashb1t]

Support for VAE download has been implemented in #3216.
Proposed preset: https://github.com/lllyasviel/Fooocus/blob/bbf90a97b7ebd1808836fb39d9e490966070f4c2/presets/pony_v6.json
Inpainting doesn't work with any inpaint engine other than "None", same as for the playground model.

Happy to hear your insights about CFG (recommended: 5-7, proposal: 7) and positive + negative prompts!

Examples (fantasy themed & unmodified, metadata in images):

CFG (Guidance Scale)	Tiefling Shaman	Mouse Wizard	Bear Druid
7
6
5

[ILikeToasters]

That actually is pretty close to what I have narrowed things down to. Although I am still experimenting with sampler but will say that one works well for the realistic ones I have tested for sure. Only difference I have done is creating a style instead of the words in the prompt.

[mashb1t]

@ILikeToasters I mean, one can always switch again to other inpaint modes if necessary or can adjust the preset to your liking / create a new one. It's about users, who don't know anything about Fooocus, just switched the preset and then get weird results.
100% the new preset key default_inpaint_method will find its way to the playground preset, but pony is debatable. It does not work well but better than just random noise compared to playground.

@BoJangles002 thanks for the reference. The inpaint model isn't trained for pony, so without it you might get better results by default. Also setting overrides in presets practically disables refiners, which would be better then to not even show the refiner select.

[mashb1t]

@BoJangles002 (temporarily) pushed to PR for evaluation, see 49d898f

[mashb1t]

@ILikeToasters @BoJangles002 after running various tests for combinations of presets, models and overrides the result is to only override the inpaint engine instead of setting the whole inpaint mode (playground + pony) + use override refiner steps (pony).

[mashb1t]

done in 33c2292, please test again.

[ILikeToasters]

At the moment that seems to be the "best" results. Although I have not given up on testing different things yet since the best is still not great but especially when doing minor things it does decent I feel.

[BoJangles002]

I would run all those images again at cfg 5/6 and see if any of the details improve. specifically things like the tiefling horns that blend into the background, the trees/birds, and the bear's sash that morphs into his fur.

The only other thing I notice is that you have censored in the negative. Uncensored in the positive is supposedly significantly better but I don't understand the inner workings to know how/why. Something something unet something something dataset.

[mashb1t]

@BoJangles002 thanks for the feedback, will check again with different CFG.

Censored has already been removed from the negative prompt: https://github.com/lllyasviel/Fooocus/pull/3216/files#diff-f106c4c5729f1dc284aafa1254813175da0bca93c15870d94c8fc6d67931fdc6R37

[mashb1t]

@BoJangles002 you can find CFG 7/6/5 now in the table above.

[jaimeastorga2000]

The base Pony model is relatively low quality and unpredictable without a good style LoRA. Recommend either pick a default LoRA or switch to one of the AutismMix models for a default anime look.