Trained on screencaps from both "A New Dawn" anime teasers.
Very small dataset of only 20 images. This LoRa will be upgraded as soon as the studio releases more material.
I did two different training approaches.
v0.1 A = More cleaner output, but compositionally more tamed.
v.0.1 B = Looks better in widescreen format with a somewhat better and more interesting cinematics..
Weight: 0.8-1.2
Steps: 25-40
Sampler: For me Euler worked the best with "Synta-XL Illustrious Mix Testing"
Helpful prompts: light mint theme, pastel colors, cool colored, vivid, high contrast
Helpful negative prompts: I've tried to manually caption as many recognizable objects as possible in the training images. So when you find some object randomly appear and mess up the scene then just exclude them via negative prompting.
For example: electrical fan, building, railing
Describe the feet if your character is squatting or sitting with knees to chest, as the training images had some of those but the feet were always covered. So by adding "shoe" or "socks" or just "feet" it should fix that.
The prompts "detailed foreground, blurry foreground" in combination with "pillar" or "column" and "door" in an indoor scenery add more compositional dynamic.
The natural language sentence "This composition follows the rule of thirds." can enhance anime-cinematography.
These are just my little guidelines on how to use a LoRa that suffers from a way to small dataset in terms of nailing this very beautiful artstyle and navigate to a satisfying result. This model will improve when new material emerges!