New version 3 is trained from the pre-eminent Protogen3.4. I found that training from the photorealistic model gave results closer to what I wanted than the anime model. This model is fantastic for discovering your characters, and it was fine-tuned to learn the D&D races that aren't in stock SD. Works best with 768x512 and other resolutions larger than 512x512
Try a prompt like:
Positive: modelshoot style digital illustration medium shot of an [older woman] elf wearing a long patterned (kilt) and [breastplate], holding a partisan spear and a shield, wearing an (embroidered) traveling quilted cloak, with white hair up in a bun, art by tony sart, wlop and rossdraws and Magali Villeneuve, sharp focus, intricate background of a countryside, hills, farmland, farm building, trees, trending on artstation, trending on cgsociety, rustic, fine art, vivid colors, detailed face and eyes [Brenda Fricker | Katie McGrath | Saoirse Ronan], [flowers and lavender and herbs in belt], [cape], stars, galaxies, nebula, runic symbols, wooden earrings, [samurai], purple gravity energy aura
Negative: glasses, old, animals, watermark, sexy, thin, (blank background), [hands], (anime), photorealistic
Which I used to find an important NPC for my campaign!
Species in new version: aarakocra, aasimar, air_genasi, centaur, dragonborn, drow, dwarf, earth_genasi, elf, firbolg, fire_genasi, gith, gnome, goblin, goliath, halfling, human, illithid, kenku, kobold, lizardfolk, minotaur, orc, tabaxi, thrikreen, tiefling, tortle, warforged, water_genasi Classes in new version: Artificer, Bard, Barbarian, Cleric, Fighter, Druid, Monk, Paladin, Rogue, Ranger, Sorcerer, Warlock, Wizard, Noble, Townsperson
Originally posted to HuggingFace by 0xJustin
Dataset used to train: https://huggingface.co/datasets/0xJustin/Dungeons-and-Diffusion
Good prompts to try things out:
[natural colors], full body tiefling (knight), [watercolor digital 2D painting], (strong shading), hard shadows, blurry, elegant, wearing robes, style of (saga comic) Lois van Baarle and charlie bowater and Sui Ishida, messy, disheveled, thick brushwork, detailed face and eyes, concept art
portrait (painting) of tabaxi, de Rivia closeup, suit, collar, formal attire, D&D, fantasy, intricate, elegant, highly detailed, artstation, concept art, matte, sharp focus, (brush strokes), (oil on canvas), hearthstone, art by Titian and Greg Rutkowski and Rembrandt van Rijn and Alphonse Mucha (inspired by Reddit post)
Description
Trained from Protogen
FAQ
Comments (22)
The new update looks great
Any chance of a LoRA version?
Gosh in the weeks since I trained this new methods keep coming out. Got any primers on LoRA? I work in ML, so I'm pretty well versed in the jargon
Might want to add that since V3 contains Protogen 3.4, there are additional licenses besides the CreativeML Open RAIL-M license. Protogen 3.4 contains the seek.art license and the dreamlike license
Can I make a plasmoid?
Thats one of the only races I haven't added! Couldn't find enough art. There might be enough style in it to get something dnd flavored and prompting in ooze and translucent skin and the like
It won't, for the life of me, produce a red skinned tiefling. It refuses to acknowledge red at all
Two days late, but I achieved the red skin by playing with the negative prompt and specifying that I didn't want white skin.
@isellmids Yeah red skin is tough for whatever reason. Try both 'red_skinned' and 'with red skin' . For men, [hellboy] also helps bring out the red skin.
Try working around this prompt:
red skin (tiefling), [Watercolor expressionist] painting of a (tiefling) warlock, [lana condor] dark swirling magic, with (red skin), [antelope horns], toga robe, marble, diamond necklace, by Pino Daeni, claude monet, Klimt, gaston bussiere, (lois van baarle), [abstract square background], [disheveled], (vibrant colors), digital illustration, masterpiece, gorgeous artwork, fine art, radiant light rays, god rays, demonic, intricate, highly detailed, evocative, rich colors, galaxies, astral, purple nebula
Negative prompt: elf, nudity, nsfw, braided, circular, interwoven, (((anime))), (hyper-realism), ((realistic)), real, (black and white), (pores), (albino), (clear skin), big eyes, full body, ((hyperrealism)), (photograph), (photo), ((high contrast)), ((outlines)), (hard edge), black lines, ((sharpened)), [[cross-eyed]], ((happy)), signature, watermark
No matter how I tried to reproduce the sample images, none of the resulting examples, even with a full copy of the seed, gives a similar result.
I'm getting some issues with images coming out completely washed. I'll use some of your prompts and change a word or two. Eventually, those slight changes will apparently break the image and all of a sudden it's coming out completely brown or gray.
hi. I had same trobule. I read this post
https://huggingface.co/0xJustin/Dungeons-and-Diffusion/discussions/13
and installed VAE, images generated have colors.
I think you have not install VAE too.
I know that they are a new race to d&d, but what about Harengon (Rabbitfiolk)? Could they be added?
A little late, but there is enough baked in that all sorts of creatures can be prompted in- try to prompt for a tabaxi, but specify rabbit head, or a species of rabbit, and I've found the model can figure it out
I am not sure what I am doing wrong but I cannot get a single similar output to what is shown.
Same seeds and parameters but images are widely different. Prompts are also very hard to get right. Can more examples be shown?
Same. Didn't take a single generation setting.
I'm trying to make a ranger and his pet but it has proven impossible thus far.
i would strongly recommend just generating the ranger, and then inpainting the pet afterwards
@BigBustah Thank you for the advice, I'll give it a try.
Useless. I tried to generate dragonborn, and it's worse dragondors I've ever seen. I tried to generate tiefling, and its generate bad anatomy human. I tried to generate human, and it failed the face. This is the worst checkpoint I've ever seen
I'm sorry, there may be mistakes when writing the review, I used an online translator.
A nice and lightweight model. If you spend some time understanding how it works, it produces amazing results. She is very responsive when interacting with LORA.
Disadvantages of the neural network:
1) Whether she smears her face heavily or makes jokes with the anatomy.
2) I may not have the skills, but every attempt at upscale cripples me mentally. Maybe the problem is in my program, but it's extremely difficult for me.
3) Sometimes, like any model, it ignores prompt, but at some point it stops doing it, that is, it needs either a lot of prompt or a little prompt.
For that, we get a juicy picture, just like from the illustrations. We get interesting charismatic characters. I've been looking for something like this for a long time.
There aint no Yuanti 😭😭
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.