Medieval Helper LoRA

☕ Support My Work: If you enjoy my local tools and models and want to support me with real cents rather than just platform Buzz, you can do so here: https://ko-fi.com/monsieurbaka. Your support directly speeds up new software updates and model releases!

About the Problem

Popular base models are generally good at creating an overall medieval warrior appearance, but they often struggle with historically plausible armor and weapons.

The connections between armor elements can start looking random, real weapon constructions get simplified or mixed together, and proper weapon grips and hand positioning are often broken.

Instead of historical examples, the model can easily turn everything into yet another piece of Fantasy armor: armored bras, decorative pauldrons, and random scale armor inspired by yet another TV series about The Witcher.

Medieval Helper LoRA was trained to improve the model’s understanding of these concepts and bring generation closer to real medieval armor and weapon constructions.

The training used 119 images collected from historical reenactments and museum exhibits.

Important: this LoRA cannot fix fundamental problems of the base model. For example, Klein 4B and Klein 9B may still have severe anatomical issues. In such cases, it may be useful to use third-party helper LoRAs.

Trigger

No special trigger word is required for any of the versions.

The concepts should be invoked directly using their natural names.

Versions

Krea 2 V2

V2 expands the dataset. The list of concepts has been expanded (see below), and stylized images were added during training for contrast.

Krea 2

At the moment, this is the base model where the LoRA shows the most interesting behavior.

What I particularly like is how the understanding of individual concepts improves after enabling the LoRA: the model becomes noticeably better at distinguishing specific armor elements, weapons, and their construction.

Krea 2 also handles more complex and diverse prompts well, so this is where you can experiment with combinations of different historical concepts, artistic styles, and more complex scenes.

Klein 9B

For Klein 9B, I recommend generating at around 8 steps.

Among the samplers, I particularly like res_multistep. In my experience, it helps reduce some of the model’s characteristic problems, especially situations involving multiple legs or strange variations of two-handed swords.

However, the LoRA itself does not eliminate Klein 9B’s fundamental anatomical problems. Its primary purpose is to improve the understanding and reproduction of the trained armor and weapon concepts.

Klein 4B

It is essentially the same situation as with Klein 9B, but the base model has more fundamental problems.

The LoRA improves the understanding of specific concepts, but it should not be expected to fix fundamental errors in the model itself.

The main advantage of this version is commercial-free use of the base model.

Anima

I did not like the result of training.

Known Concepts (Main list for all versions)

Below is a list of concepts that were sufficiently represented in the dataset and were considered trained. The list is sorted by the number of images in which each concept appeared.

Armor & Helmets

full plate armor
close helm
articulated pauldrons
pointed sabatons
pointed visor
gambeson
pivoting visor
fluted plates
sallet helmet
hourglass gauntlets
buckled front closure
articulated fingers
segmented finger plates
mail skirt
greaves

Weapons & Handling

longsword
wooden haft
spike point (halberd)
halberd
back spike
axe blade
flanged mace
flanged head
buckler
buckler grip
long grip

Shields & Mail:

central boss
mail hauberk
mail aventail
ventilation holes(helmet)

Non-European / Other Cultures:

samurai armor
lamellar cuirass
kusazuri
menpo
mask helmet
cloak
kabuto helmet
lamellar armor
round shield
conical helmet

Addons exclusively for Krea 2 V2:

Horses & Mounted Combat
horse armor
chanfron
crinet
peytral
flanchards
horse caparison
mounted knight (seat, rein grip, lance carry as a composition concept)

Shield Attachment
heater shield
kite shield
forearm strap (enarme)
sword belt hanger

New Weapon Types
falchion
messer
sabre
rapier / swept hilt
smallsword
estoc
stiletto
bardiche
glaive
bill
voulge
war scythe
corseque / ranseur

New Helmets
barbute
kettle hat
morion
frog mouth helm
nasal helm
armet
great helm

Armor Details
gorget
tassets
poleyns
leg armor
leather arming straps
arming sword

Non-European
mongol armor

What This LoRA Is Designed to Do

The main idea behind the project is not to make the model generate “medieval” as a separate style.

The goal is much narrower: to give the model a more consistent visual understanding of specific elements of historical armor, weapons, and ways of using them.

Therefore, concepts such as full plate armor, close helm, mail hauberk, longsword, samurai armor, mask helmet, halberd, and others should work independently and be combined with ordinary prompts.

The LoRA is not a replacement for a good base model and is not a universal fix for anatomy. It is specifically a concept helper designed to help the base model better understand the historical construction of armor and weapons.