This article only talks about the stats for the Pony Diffusion V6 XL model.
TLDR: Honestly, the stats didn’t reveal anything super exciting. But I’d say the part every newbie should definitely check out is about CFG Scale, Sampler, and Steps. You can save buzz if you use more cost-effective configurations!
In the previous article, Image Generation Guide for Beginners, I intentionally avoided discussing some topics, like the CFG Scale, Sampler, and Steps. I believe there isn't a single correct answer regarding their usage.
In this article, I will explore what statistics reveal about the most popular configurations, loras, prompts, and what users tend to prefer.
The full list of what will be covered:
The most popular combinations of CFG Scale, Sampler, and Steps
The most popular Loras
The most popular positive prompt words
The most popular negative prompt words
The most popular rating
The most popular image size
Bonus: the best times to post for maximum popularity :)
As always, I’ll keep the text brief and include more images! If you want to read boring details on how I gathered statistics, read until the end. The main thing to remember is that I'm going to compare the stats of the most popular images (those with over 200 reactions) with all the other images.
The most popular CFG Scale, Sampler, and Steps combinations
Images with 200+ reactions
Click here to see the full image.
This stats focuses on images that have gotten more than 200 reactions, which suggests that the creators probably know their stuff better than the average person when it comes to making images.
Most of the images that got over 200 reactions were made with the "Euler a" sampler, CFG scale: 4, and 29 steps.
Why 29 steps is popular? Because generating one image on Civitai with 30 steps costs one extra buzz.
Why 24 steps is popular? Because generating two images on Civitai with 25 steps costs one extra buzz.
Basically, the most popular configurations were chosen because of the price, not because they are the best.
But here is a comparison of some of them:
Click here to see the full image.
The difference isn’t that big and mostly comes down to style. In my experience, "Euler a" gives off brighter images and a little 3D effect, while "DPM2 a" gives a plain effect.
I personally prefer "DPM2 a" over "Euler a".
I usually use this configuration: "DPM2 a", CFG scale: 3, and 15 steps. And increase the number of steps only if the image has defects.
All images
Click here to see the full image.
This is a rundown of stats on all the images, whether they're popular or not.
The first place: "Euler a", CFG scale: 7, and 30 steps. The reason is obvious - this is the default "balanced" setting by Civitai.
The second place: "DPM++ 2M Karras", CFG scale: 7, and 25 steps. No idea why that is. If anyone has a clue, drop a comment!
And some comparison:
Click here to see the full image.
The most popular Loras
Images with 200+ reactions
Click here to see the full image.
There are tons of different Loras out there, each with its own style, but the most popular images usually come from just a few of them:
ExpressiveH (Hentai LoRa Style) エロアニメ:ExpressiveH
Incase Style [PonyXL]:v3.0 (PonyXL Edition)
Vixon's Pony Styles - gothic neon:gothic neon v1.0
All Disney Princess XL LoRA Model from Ralph Breaks the Internet:v2.0
All Disney Princess XL LoRA Model from Ralph Breaks the Internet:v1.0
Let's look even closer, at the most popular loras combinations.
Click here to see the full image.
If you take a closer look, you'll see that they’re basically the same. The only difference is like 1 or 2 loras, and even the weights are the same in different combinations.
Here is a visual comparison between them:
Click here to see the full image.
All images
Click here to see the full image.
Not much has changed in the top 10 Loras compared with the most popular images. But there are a few newcomers in the top 20, like ThePit Style, Geekpower AI Artstyles, and Penis Size Slider.
Click here to see the full image.
The most popular combo is actually the same as what's hottest in the most popular images stats. But here’s the really cool part: out of over 200000 images, only about 1000 feature this top combo. That’s awesome because it shows the community is really experimenting with different combinations, even if these are just different weights.
The most popular positive prompt words
The first image displays the most popular words in positive prompts among images with over 200 reactions, while the second image shows the most popular words in positive prompts across all images.
Click here to see the full image.
Click here to see the full image.
The word "safe_pos" totally caught my attention. Is it some kind of magic word that makes images go viral? If anyone knows, drop your thoughts in the comments!
All other statistics are not very interesting, except for some fun facts:
The community likes to generate images with blue eyes, but then give more reactions to images with green and purple eyes.
The community likes to generate images with huge and big breasts, but then give more reactions to images with small breasts.
The word "beautiful" is popular among images with over 200 reactions, but not among other community members. Can it be a real secret?
Ok, let's be serious. The real words that potentially can influence the quality and exist only in images with over 200 reactions are: detailed background, dynamic angle, volumetric lighting, godrays.
Let's check them out:
Click here to see the full image.
The most popular negative prompt words
Click here to see the full image.
Click here to see the full image.
Nothing interesting here as well, except for the magic word "safe_neg".
The most popular rating
Browsing level is an internal name in civitai for the rating:
Browsing 1: PG
Browsing 2: PG-13
Browsing 4: R
Browsing 8: X
Browsing 16: XXX
Looks like everyone’s noticed that PG and PG-13 images aren’t super popular in Pony Diffusion V6 XL. But if you take a closer look at the stats, you’ll see there's a better shot at hitting 200+ reactions with PG images. Probably because some users have filtered out everything else.
The most popular image size
There is almost no chance of getting 200+ reactions if you don't create images in 832x1216 format. Even though, other formats are relatively popular among the community.
Bonus: the best times to post for maximum popularity
That title is too clickbait, right? Nothing to see here, I promise!
First, let's look at how many images community members posted at 10-minute intervals.
The timezone UTC: 00:00 corresponds to 19:00 in New York.
Interestingly, there are peaks at XX:00 minute times, indicating that many people schedule their images for posting.
Now, let's split images into buckets:
from 4 to 9 reactions
from 10 to 49 reactions
from 50 to 199 reactions
from 200 to 499 reactions
more than 500 reactions
Click here to see the full image.
Now, I urge you to be very careful with statistics.
Statistics say posting at midnight (00:00) is the best time to rack up over 200 reactions. But at that hour, a ton of images are also getting way fewer reactions.
So, it can really go either way. You might want to try posting when popular images are going up or pick a time that isn’t so crowded.
Like I said, there’s really nothing to see here. I don't have an answer for you about the best time to post images.
Some boring details about the statistics
How I gathered it
I used an official REST API that is available to all users, so anyone can gather the same statistics.
Check out the docs here: https://github.com/civitai/civitai/wiki/REST-API-Reference
You'll need to create an API key in your account settings: https://civitai.com/user/account
Here are the params I used:
params = {
"modelVersionId": modelVersionId, # Model filter
"limit": 200, # Max results per page
"sort": "Newest", # Sort by the newest images
"period": "Day", # Set period to last month
"nsfw": "true", # Filter
"cursor": next_cursor # Cursor for continuation
}
Civitai has limited the number of images that can be retrieved in a single response to 200. So, you'll have to keep sending requests in a loop to get more. This is the purpose of the cursor provided.
They also cap the total at 50k images. After 250 requests, you’ll hit a point where you get an empty cursor, meaning you can’t pull all the images at once.
The community posts around 50k images daily with Pony Diffusion V6 XL, so you’ll be making requests for a few days if you want to gather some decent stats.
The "nsfw" flag can be a bit confusing and doesn't really work the way the documentation says it does. If you leave it out, you'll only see PG images. If you set it to "true," you'll get R, X, and XXX rated images. If you go with "Soft," you'll get a mix of PG and PG-13 stuff. So, I ended up having to make two requests: one with "nsfw": "true" and another with "nsfw": "Soft" just to get all the images.
What I gathered
I made two types of requests:
1. With parameters "sort": "Newest" and "period": "Day" to gather statistics that I mention as "all images" in the article. I gathered this info over four days:
2. With parameters "sort": "Most Reactions" and "period": "Month" to gather statistics of popular images. Because the statistics are limited to 50k images, it mostly contains images with 4+ reactions, but I filtered out all images that have less than 200+ reactions.
So, as you can see, the comparison between all images and images with 200+ reactions is not an apple-to-apple comparison.
Final thoughts
I was way more excited about doing the research than actually writing this long article. By the end, I was pretty wiped out and didn’t really proofread it like I should have. If you spot any mistakes, just shoot me a private message or drop a comment, and I'll fix it!