The spoonfeeders guide to PDXL

Back to Main Page

Model Choices

PDXL

SeaArtFurryXL

Honorable mentions
Bananastrike:
is an updated V2 of compassmix, same ups and downs apply. Supposedly improves on backgrounds but I find it just slaps a really boring default look to pictures. I'm somewhat biased against the SeaArt family of models as is so take my subjective opinion here with a grain of salt, the team working on these does good stuff, just not really my thing. The fact that they are all exclusively on lightning is really ass by the way.

The unfathomable boundless depths of realistic Pony merges:
There's sooooo many out there. To understand why, you have to realize there's basically no good models out there for realistic porn, so like flies they swarmed to pony and all tried to merge realism models into it with varying success. Some of them are kinda decent? Your mileage may vary, I'm not a huge fan of these. All of them tend to have the same issues, massive human bias, massively lobotomized composition. But I also find realistic stuff pretty uncanny so I'm biased.

Refiners
You can use a realism model as a refiner to get the composition of base Pony, and some of the default style/shading of your chosen refiner model. As an example you could do PonyXL -> Refiner 0.7 into PonyRealism (this makes backgrounds kinda nice).
I've seen people do their base gens on Pony, then refine with SeaArt, that might be something you could try as well. I don't have positive experiences with it.

Your own merge
Recently I've been simply using a merge of PonyXL with PonyRealism because it helps keep things consistent with inpainting.
0.7(ponyDiffusionV6XL_v6StartWithThisOne) + 0.3(ponyRealism_v21MainVAE).safetensors (70% Pony/30% PonyReal)
0.5(ponyDiffusionV6XL_v6StartWithThisOne) + 0.5(ponyRealism_v21MainVAE).safetensors (50% Pony/50% PonyReal)

Merging is easy, in the WebUI you just choose model A, model B, choose the ratio, hit merge. I recommend not renaming them so they become easier to recognize.

My Verdict

You have Autism(actual,not the model) and a lot of time or do this as a hobby?
PonyXL, just vanilla. Since most loras are directly trained on base Pony, it also means loras affect it the strongest. Since it's such a blank slate, it means you can turn Pony into something fairly personalized.
Most Pony merges simply act like base Pony, merged with a strong style lora. That's how I treat them at least. They aren't bad, but I wouldn't call them strict 'improvements' as a 'main model' at all. Some of them are gigacope imo.

You like Anime or smooth looks
Can just go with AutismMix. The Anime bias is a blessing or a curse, it's one of the earliest merges and one of the most competent ones. If you want Pony but aren't ready yet to spend hours lora mixing, Autism is a good choice. Good beginner model.

You come from SD1.5 / You don't have a lot of time experimenting / Your favorite artist has no lora
CompassMix or base SeaArt. I do think CompassMix improves noticably on SeaArt, but both essentially still suffer from the same problems. I don't use these much, some swear on them. Pony was a big jump in compositional strength in general. SeaArt struggles with a lot of things that you take for granted on Pony, duo scenes for one.

THE TIERLIST

Tiers ordered from left to right. Tiers assume competence in the model. This is extremely subjective and entirely based on how I use AI to make funny pictures, which is in part HEAVY inpainting and manual editing. If a model is not mentioned it was not worth mentioning. If this was a list of beginner models it would look very different. Realism merges are excluded.

S: PonyXL,
A: AutismMix, EasyFluff/QueasyFluff(SD1.5),
B: SeaArtXL, IndigoFurryXL, BeastMixXL, YiffyMixXL, CompassMix,
C: BananaStrike,

General tips for PonyXL

Important SDXL Setting!
Go to settings > Stable Diffusion > Stable Diffusion > Emphasis mode: No Norm.
This fixes an issue in SDXL where under specific prompts you will get broken generations at seemingly random, fixing itself with a minor adjustment to the prompt. The side effect of this is that emphasis works more akin to ComfyUI's, which means the emphasis is generally much stronger. That is to say, if you are used to doing things like (wide hips:1.5), you will have to reduce it down to (wide hips:1.2). 1.3 is the highest you should ever go in this mode. It's worth switching to No Norm right away so your prompting stays consistent.

Score tags
The more score tags you include the 'better' your image gets, but the dumber the composition variance. What I mean by that is it turns more and more into 1girl slop considering you are excluding a lot of 'bad' images to give it more variance. I find cutting positives off at score_6_up and negatives at score_5, or none at all, is good enough.

Source Tags
The source tags should be focused depending on both material and style. You can prompt source_furry AND source_anime if you want furshit in anime styles for example. You generally do not need rating tags unless you want to make something SFW (and even then you just prompt clothed).

Danbooru tags? E621 tags? Natural language?
Danbooru and E621 tags BOTH work. In fact, some tags do not always have equivalents and it can be good to check if a concept exists on the opposing booru. Natural language is often cope, but it does occasionally reinforce more abstract concepts.

MLP Bias
Pony obviously has a strong bias towards ponies. Including source_pony in negatives helps a lot. Other bleedovers include puffy_anus, equine_penis, and so on. I add these when they come up.

Locations/Backgrounds
Pony is very sensitive towards styling the overall image after the location or backgrounds. I like detailed background as a tag, then followed by combinations for locations such as (jungle, waterfall), (forest, stump), (cave, coral), (bar, cyberpunk), (medieval, fantasy). These tend to give wildly different results in both composition and color balance, I consider these to be practically artist styles.

Style Tags
Styling can also be heavily influenced by tags like realistic, watercolor (artwork), oil painting (artwork), etc. Keep in mind these are soft style tags, if you are prompting photography (artwork), you WILL inevitably end up with photos in your image.

Lora Reliance
Pony gets a lot more focused with Loras, especially multiple. It basically ends up like EasyFluff style artist mixes. However a shit ton of loras online are badly trained. A ton are overtrained, some are clearly badly tagged. I would avoid generic style loras unless it does something very specific for your image. Importantly you DO NOT have to use furry based artist loras to generate furshit. Pick artists you like on a composition, coloring, detailing basis.

Schizo negs
From my experience the usage of schizo negs, that is using excessive negative prompting of very vague technical faults such as "moiré pattern or downsampling" are NOT required and do not give consistent positive enhancements. You see these A LOT on civit because someone started them and everyone copied. As a general rule, if it does not have a booru tag, do not use it.

Character lora
Don't forget character loras will inevitably carry a ton of style from it's training material. From my experience, especially for characters Pony already partially knows, you can either reduce the lora weight a lot or schedule them to be disabled through an extension after half the step count.

Resolution
Remember SDXL is trained on 1024x1024. Deviating up and down is fine !as long as it's relative!. 1152x896 is a res I like, but feel free to change it to more traditional aspect ratios. Again, you wanna pick aspect ratios !around! 1024x1024, something like 1512x1024 as your base res is NOT valid.

Hashes
Artist tags were obfuscated in PonyXL through what people refer to as hashes. It's not exactly a hash but whatever. What this means is you can call a 3-4 letter prompt and get a weak version of someones style. For example (gjem) approximates Wamudraws style. One documentation sheet. Long story short, most of these suck to use, they have a ton of bleedover. Not gonna comment on the author of Pony doing this in the first place.

Some random stuff
1: (censored) in negatives does not uncensor your image or avoid censoring, in fact it often makes pussies look like hot ass.
2: I prefer to edit/inpaint watermarks out rather than try to counter them with negative prompts. I have no basis for potential downsides of using those negatives, I just have a hunch it bleeds over.

Sampler Choice

Good:
DPM++ 2M Karras: Good allrounder. 30 steps. 7 CFG. Basically the default.
DPM++ 3M SDE Karras: Different sort of detailing. Lower CFG to around 5.
DPM++ 2M CFG++ Karras: An attempt to merge the functionality of attention guidance into a sampler. Very cool? 50 steps, 1CFG. You may need to edit your webUI to allow lower CFGs even, it's designed with very low ones.
Restart: Slower. Very nice out of the box detailing. Forgot what step count this needs.

Mixed:
EulerA: I have a soft spot for ancestrals but Euler does kinda suck at details.

Bad:
Anything (you) use.

For Inpainting: The same sampler as the basegen or DDIM (DDIM is supposedly useful for inpainting for high detail shading and such since it's context aware, I have had mixed experiences with this, occasionally it just makes areas terribly smudgy.)

Upscale choice

I mostly use Hires.fix with 4x_NMKD-Siax_200k at 1.5-1.7 scaling and 0.3 denoise. Depending on the style you are working with you can go higher on the scaling or denoise. For more advanced upscaling and detailing tips check Inpainting & Img2Img.
If you do install 4x_NMKD-Siax_200k you may need to create the ESRGAN folder in your "Model" folder yourself, it's slightly confusing.

Useful Loras

PDV6XL artist tags
Link
Attempts to give PDXL artist knowledge similar to Easyfluff etc. The second file under downloads has a list of artists. I personally have mixed experiences with this, but I can't deny I've seen some VERY good gens done with it. I believe my results are mixed because not all artists are represented equally well in the lora, some probably work more than others, I've seen people say they sometimes even work better than dedicated loras.
Do not expect the exact same artstyle with your artist mixes as you would in SD1.5 models, the weighting will be different. If artist tags are what holding you back from messing with PDXL, please give this a try.
A kind anon has opened a mega with comparisons: PDXL Artist Comparison

ExpressiveH
Link
Makes H scenes more... expressive. Works pretty well, but has a noticeable style footprint. Scheduling the lora to turn off after half steps might be worth.

Add detail/More detail
Don't, they sloppify your image.

Generic Style Loras
Styles for Pony
Styles for Pony(different author)
Vixons Styles
Your mileage may vary. Sometimes you have to have them at low weights, sometimes they bake your image hard. Occasionally they look incredible. Often you have to finetune their strength to the rest of your image and or other loras you might be using. I use the 2.5D ones from the second link a lot because they seem fairly well trained, a lot of the ones from the first links, especially the early ones, are not. You can usually tell when a style lora is badly trained when your images become incredibly grainy even on hires.

Other Download Locations
_Ka_De's website with Loras
Giga autist who makes really good loras, but doesn't post most on Civit because reasons. Nice guy, makes some really good resources too if you want to browse some. The lora guide in particular is VERY in depth and well written.
Trashcollects
/sgd/ run repository.
PonyXL notes
Similar to the above but anime focused.

Any artist loras you can get your hands on
Some are dubiously trained, some are overtrained, but the ones that do work are what transform Pony into something usable. Base Pony is like a blank slate, a hodgepodge where all artists were mixed together into primordial soup. Loras can form the smarts of the model into something usable. Not all artists have a lora unfortunately. Also do not limit yourself to just furry/anime lora. Nowadays I look for specific color schemes, how lines are drawn, that sorta thing.

Useful Embeddings

You don't really need embeddings on PDXL. The model is well trained enough as is, most SDXL models don't benefit much from embeddings in the first place, because they don't have glaring issues you need to blanket fix.
The only embedding that I have used for convenience sake at one point were the zPDXL embeddings, there's always a positive and negative embed. But it's really not required and I don't use it anymore unless I'm testing.

General prompt advice

Properly organizing your prompt is just as important as the prompting itself. It makes it easy to find stuff to change, it makes it a whole lot easier during inpainting.
Some universal things to keep in mind:

• Prompts get divided into 75-long token blocks. You can see your current token length at the top right of the prompt box in the WebUI. Once this is filled, a new block gets created. The BREAK keyword forces the current token block to end before its filled.
• Just how models were trained on a specific resolution, they were also trained on a specific token count. 75 is actually the maximum, the workaround that essentially happens is it divides the generation into multiple blocks for each step. The more token blocks, the less overall weight tokens will contribute to an image. You may notice a drastic change in the image once you breach a 75 or 150 token threshold.
• Using BREAK is only advised if you know what you are doing, in most cases using a proper format means it's not needed.
• In some models like the SD1.5 models, it is prudent to have things like quality and artist tags first, then use a BREAK, so all of it is in it's separate block and you don't run into certain issues. In Pony I generally feel no need.
• BREAK CAN be used to divide character definitions, but I consider successes to often be a bit biased and mostly luck. Regional prompting sucks in Pony so Inpainting is your friend.

Here's how I structure my prompts in PDXL. This is by all means very subjective and primarily based on organization, not what might be optimal for the tokenizer:

And here is an example for a simple duo scene. Not using everything, no embedding etc:

Keep in mind that dividing characters with just raw prompting is difficult. This prompt has a natural bias towards a male lucario and a female renamon, in other pairings you need to get lucky. Defining it instead as (female renamon) does not naturally lead to a female renamon, the tokenizer sees all the prompting overall and decides that way. The BREAK, "Can" reinforce this, but it's still luck.
On a personal level I often don't bother with a natural language prompt describing the scene (also known as boomer prompt). I find tags to be much more succinct and predictable.

Prompt Templates

A semi random collection of subjectively decent quality prompt templates. This is supposed to get you started. You can right click save any of these, and import them in the WebUI under PNG-Info. Not all of the prompting on these is "perfect", so if anything looks off, blame the human error of the author or his bad taste. Trying to feature a variety of styles but I'm not always into something enough to explore it.

All of these are done on Vanilla PDXL. You can use these in AutismMix just fine, they will look slightly different.

All of these are raw gens with no inpainting, no hires or upscaling.

Some like Oouna or Meesh require trigger words in the prompt.

Important: Check the lora weights from the imported template pic. If you put every lora at 1.0 it will look bad.

These are meant as starting points and to showcase the 'meta' of lora mixing as opposed to the artist tag mixing of the SD1.5 models. 'Teach a man how to fish' sort of thing. Experiment with your own mixes too.

PonyXL (Vanilla)

AutismConfetti(Similar on regular Autism)

FAQ

My generations look normal until the last step and then they finish garbled.

Remember to install the SDXL vae, you may also have some legacy face restoration setting toggled on in the settings.

My generation randomly looks garbled on the same prompt, but magically fixes itself when I do a minor change.

Known problem, navigate to settings -> search "emphasis", set the mode to "no norm". There was an extension that fixed this but the option is now supported here in the settings, just not enabled by default.

Pony sucks muh EasyFluff is so much better.

I love 1girl! I love on_back!
I love easynegative, badhands-4, boring-e621-v3, bwu, dfc, ubbp, updn, (((deformed))), blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), fused fingers, messy drawing, broken legs censor, censored, censor_bar, multiple breasts, (mutated hands and fingers:1.5), (long body :1.3), (mutation, poorly drawn :1.2), black-white, bad anatomy!

Please, try the PDXL artist tags Lora. it's genuinely good, jokes aside. Easyfluff is also good, both can be good!