Do NOT ask what happened to versions 004 through 009
This version aims to solve a number of problems with previous versions:
Too little variety in generation because of a small, uniform dataset
Outputs occasionally look "low-resolution" because of the game's sprites being made at slightly less than 1024x1024
Too little interoperability with other LoRAs / base models
This is done by adding a new dataset trained on frames selected from the game's two opening cinematics, and improving the quality of the existing dataset.
Known limitations with this version:
No longer very good at generating the game's iconic "Afflicted" scenes
Will often generate jagged edges, and is very biased towards creating deciduous trees
Short keyword dictionary:
DDCinematic - accesses the new dataset trained on the game's opening movies
DDAncestor - tries to create a likeness of the Ancestor - not tested very much
DDAreaCove - tries to create the Cove, seems inconsistent
DDAreaRuins - same but with the Ruins
DDAreaWarrens - same but with the Warrens
DDAreaWeald - same but with the Weald
DDAreaNone - tries to create a likeness of the Hamlet or the Darkest Dungeon as depicted from the game's title screen
DDHero<HERO_NAME_HERE> (e.g. DDHeroGraveRobber) - tries to generate the likeness of a specific hero - seems relatively inconsistent?
The preview images are a little single-minded, so I'm eager to see what people can do with creative prompting, different base models and LoRA-mixing!
To-do list for the dataset in future versions:
The Flagellant, Shieldbreaker and Musketeer once I get their relevant DLCs
General graphics of the Crimson Court, Colour of Madness and Butcher's Circus DLCs
Do NOT ask what happened to versions 004 through 009
This version aims to solve a number of problems with previous versions:
Too little variety in generation because of a small, uniform dataset
Outputs occasionally look "low-resolution" because of the game's sprites being made at slightly less than 1024x1024
Too little interoperability with other LoRAs / base models
This is done by adding a new dataset trained on frames selected from the game's two opening cinematics, and improving the quality of the existing dataset.
Known limitations with this version:
No longer very good at generating the game's iconic "Afflicted" scenes
Will often generate jagged edges, and is very biased towards creating deciduous trees
Short keyword dictionary:
DDCinematic - accesses the new dataset trained on the game's opening movies
DDAncestor - tries to create a likeness of the Ancestor - not tested very much
DDAreaCove - tries to create the Cove, seems inconsistent
DDAreaRuins - same but with the Ruins
DDAreaWarrens - same but with the Warrens
DDAreaWeald - same but with the Weald
DDAreaNone - tries to create a likeness of the Hamlet or the Darkest Dungeon as depicted from the game's title screen
DDHero<HERO_NAME_HERE> (e.g. DDHeroGraveRobber) - tries to generate the likeness of a specific hero - seems relatively inconsistent?
The preview images are a little single-minded, so I'm eager to see what people can do with creative prompting, different base models and LoRA-mixing!
To-do list for the dataset in future versions:
The Flagellant, Shieldbreaker and Musketeer once I get their relevant DLCs
General graphics of the Crimson Court, Colour of Madness and Butcher's Circus DLCs
Darkest Dungeon 2
The Runaway
The Duelist