What this is about

When I first started using Stable Diffusion, I was somehow overwhelmed by the incredible amount of options available to me. From different prompts to this huge range of available checkpoints, all the options and tweaks and I didn’t even know about LORAs, Controlnets and whatnots. However, with so many choices at my disposal, I found myself hit by option paralysis. In order to overcome this, I decided to experiment with brute-forcing different options.

As a tabletop roleplaying game enthusiast, my focus was on generating avatars and tokens for PCs and NPCs in VTT. I began by selecting checkpoints that were specifically tagged for fantasy and RPG genres. Using a basic prompt, I then iterated over every race and class in the D&D 5e Players Handbook, creating four images each - two for female characters and two for male characters. All of these images utilized the same two seeds and a three-word character specification to ensure maximum comparability. The end result was a huge collection of 432 character images for each checkpoint, providing insight into the knowledge that each checkpoint has about the classes and races in the Players Handbook as well as the aesthetics that are baked into it.

To keep the focus on the characteristics of the respective checkpoints and not my personal perferences, I tried to keep the prmpt and negative prompt as simple and unspecific as possible. This way the differences between the checkpoints aren’t overshadowed by the specificas of the prompt. The negative prompt consists mostly of an extreme short version of keywords to reduce disfigured hands and head (to short to be really effective) and some attempts to overcome the fact that a significant part of the images available for training are photos of painted miniatures that are not the intended imaginery here. For the sake of the images of gnomes, I also added “garden gnome” to the negative prompt.

Content Warning

OK, there are a few things you might encounter in the collection:

Caveats

Some of the checkpoints are also capable of creating NSFW images and for some of those, a simple nswf in the negative prompt doesn’t seem to be enough. This seems to happen mainly with checkpoints that don’t contain any detailed rpg- or D&D-related training data. If the traning data only contains mal babarians and most of them are bare chested, well… I screened the images generated for those that I consideren NSFW and reran them with a slightly modified prompt putting more emphasis on the SFW nature of the intended result.

What this is not about

I don’t want to rank, judge, promote or bash any of the checkpoints and their respective creators.

Can I get the prompt?

Sure, here it is:

%gender% dungeons and dragons %race% %class% in a scenic fantasy background, dungeons and dragons character artwork, masterwork, best quality, ultra high res, 8K, UHD,  photorealistic, focus on character, full body portrait, (symmetric eyes)
Negative prompt: monochrome, (cartoon:2), figurine, text, watermark, logo, back to viewer, 3d, render, on a base, extra fingers, deformed face, distorted face, mutated, disfigured, poorly drawn face, bad anatomy, garden gnome, chibi, close up, nsfw
Steps: 50, Sampler: Euler a, CFG scale: 5, Seed: 2889396479, Face restoration: CodeFormer, Size: 512x512

The 2 Seeds are 2889396479 and 2889396480.

About the Checkpoints used here

I picked checkpoints from civitai.com with the tags of #rpg or #fantasy from the to of the all time highest rated list with more than 5k downloads at the time of checking and some with promising or interesting preview images or skipped checkpoints that were extremely focussed on anime or that I actively disliked. I’m not yet done with the initial list of checkpoints and might add additional ones after I’m done with these.

Some of the checkpoints needed a little extra tweaking (or in some cases less tweaking to the prompt) to work properly within the parameters of this project.

Checkpoint Version Notes
A-Zonya Tools V3
A-Zonya Tools V2 (*)
A-Zonya Tools V2 offset (*)
Deliberate 2 5 images had to be redone due to nsfw content
RPG V4
Protogen x3.4 1 Protogen has a couple of trigger words that I didn’t use.
Dungeons N Waifu’s 2.2
526Mix V1.4.5
526Mix V1.4 (*)
Rev Animated v1.2.2 There was one image that had to be redone with more emphasis on the nswf-negative-prompt
Perfect World V3 (Baked) I ommitted this because it seems to have a different kind of ‘fantasy’ in mind
Artisanica v1.0 Artisanica shows a noticeable bias to create female bodies even when asked to create male characters and in the 3 cases where there were nsfw results, putting more emphasis on the sfw/nsfw positive/negative prompts didn’t do much.
Clarity Clarity 2 The checkpoint claims that ‘SFW Photography is also possible’ but it wasn’t easy. Roughly 5% of the images were topless women or topless men with female body features. Getting these covered up wasn’t easy. Increasing the emphasis on swf and nswf in the positive and negative prompts to the points wgere the image broke didn’t work and sometimes replacing the nswf in the negative promt helped more that adding the replacement to it.
Lyriel v1.6
Lyriel v1.5 (*)
M4rv3ls & Dungeons v2.5
M4rv3ls & Dungeons v1.0 (*) Using the correct VAE created one nsfw image while using a different one was completely sfw.
v1-5-pruned-emaonly v1.0 Added as a global baseline
Level4 v5.0 Baked VAE fp16 A couple of images had to be redone because of nsfw content
Dreamshaper 6
Dreamshaper 5 (*)
Dreamscapes and Dragonfire 3

(*) These are initially hidden, use the “Show/Hide” button to change visibility.