creator: Ongelmanratkoja

This is a Pony-based checkpoint merge made with "Train Difference" towards different checkpoints and Lora merges, instead of regular flat merging.

Loras haven't been applied directly to the checkpoint, rather the model has been "Trained" towards that other checkpoint instead.

Goal has been to improve the "base quality" and change the base "native" style of Pony, so it would make better images with less negative prompting.

However, what I noticed is that some Loras which have been trained with "base pony", have also inherited it's drawing style, so when applied to this model it will skew the drawing style.

It's Pony base model, so all related tags and prompts work with this one, and because it's a merge, some stuff might've broken on the way.

I've had fun with it, so thought to share it if others find it interesting. This is my first model here, so comments are appreciated.

Thanks go to all the people doing Stable Diffusion models and Lora related stuff, as they really do all the hard work.

So to start a prompt, pony basic quality trigger words work:

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,

And after enter the prompt/items you want in the picture.

If you use them all, then the model will also draw more stuff into the picture, as pony was trained with all of those. More information about the score tags in article:

"What is score_9 and how to use it in Pony Diffusion"

https://civitai.com/articles/4248

But you can drop off the lower ones and mix up a bit, like in the examples. Like what I've mostly used:

score_9, score_8_up, score_8, score_9

score_9, score_8_up, score_7_up

Etc. Depending what kind of quality or style you want. Or in case if it seems like the picture is missing something, which might come up with the lower scores included.

For long prompts, sometimes it may be good to increase the emphasis on some of the words like this:

(detailed:1.2)

Which will in this example increase attention to the word/items within the brackets by factor of 1.2. So for longer prompts they will be focused on more often.

Because it's a Pony base, you can also utilize the data selection tags of:

source_pony
source_furry
source_cartoon
source_anime

Ratings of:

rating_safe
rating_questionable
rating_explicit

And:

censored
uncensored

Characters, styles and artists work also, but as the "base" style has changed, from my experience it will also draw the artists & styles in a different way:

Some worse, as have lost parts of training data
Some with better quality, but losing artist's original style, as it's skewing towards model style instead of artist's original picture format/style
And some will work the same as before (probably were included in the other models also)

With loras, if you are seeing some weird anomalies with high Lora strength, lower it down to like 0.5 and see if it still happens and/or is the concept of the Lora still applicable and then go up/down depending how it works.

Also if you are using CFG Scale=7, you can try to lower to CFG Scale=5 as that might fix it (or at least improve it).

Some Loras still work fine with 0.8 and even 1.0 strength.

But I think as the model has "learned away" from base pony, there may be problems with Loras that otherwise work fine with other pony-based models.

For the negatives, you shouldn't (hopefully?) need as much as base pony would. Other models have already improved on this aspect, so "Train Difference" incorporated them into this merge as well.

Negatives will still work as usual, but as they also might change the picture composition, you have try things out.

Personally I always start with blank negative prompt to use the model's trained style. Then after some prompting filter out stuff or to fiddle with composition/quality/style. Like adding words which shouldn't even be in the picture to alter the composition.

Some helpful negative triggers below. NOTE: Some of these will force a change in the style.

How to avoid Real Face

(realistic, lip, nose, tooth, rouge, lipstick, eyeshadow:1.0)

How to avoid too muscular body shapes:

(abs, muscular, rib:1.0)

How to avoid Bokeh

(depth of field, bokeh, blurry:1.0)

How to remove mosaic & censorship

(censored, mosaic censoring, bar censor, convenient censoring, pointless censoring:1.0)

How to remove blush

(blush, embarrassed, nose blush, light blush, full-face blush, shame, ashamed, shy:1.0),

How to remove some NSFW effects

(trembling, motion lines, motion blur, emphasis lines:1.0),

How to remove double navel (happens when using euler a & hiresfix)

(double bellybutton)

How to remove watermarks, etc

(watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name:1.0)

Some Lora's have been trained against simple white background, so to remove those:

(simple background, white background:1.0)

Dimensions Aspect Ratio

1024 x 1024 1:1 Square

1152 x 896 9:7

896 x 1152 7:9

1216 x 832 19:13

832 x 1216 13:19

1344 x 768 7:4 Horizontal

768 x 1344 4:7 Vertical

1536 x 640 12:5 Horizontal

640 x 1536 5:12 Vertical

Some others work too which are in between, or higher to 1440, but character shapes will bend or get long limbs.

Also, may be good to note that some Loras have been trained sit specific aspect ratios, so they work better in those.

Sometimes messing around with the resolution a bit will change the contents and how they are drawn, esp there isn't enough room in the specific resolution for all items in the prompt (long / complex prompts).

Same will also happen by adding negatives which shouldn't even be in the picture, they still affect the composition in one way or another, some more and some less.

Example pictures have been made using SD Forge https://github.com/lllyasviel/stable-diffusion-webui-forge