Recommendations (v1.0 - Anima)
Trigger: valkyrie \(clash of clans\) (escape the brackets in ComfyUI) plus the outfit tags orange hair, short layered hair, blue eyes, brown crop top with gray trim, brown loincloth, brown wrist guards, brown boots with gray cuffs. Weight: 1.0 (0.8–1.2); use 0.8 for outfit changes.
Settings: 30 steps, CFG 4, er_sde / simple, 896×1152 or 1024×1024. Add signature, watermark, text, artist name to the negative. More tips are in the version notes.
The Valkyrie from Clash of Clans: orange hair, blue eyes, brown crop top and loincloth, wrist guards, boots, a big double-headed axe. The trigger gives her hair and face; the outfit tags give the classic game outfit.
Description
Anima Base v1.0 version, trained with ai-toolkit on 20 images.
Trigger: valkyrie \(clash of clans\). Escape the brackets in ComfyUI, otherwise they are read as weight syntax.
Weight: 1.0 (0.8–1.2). Use 0.8 for outfit changes.
The classic game outfit needs tags: orange hair, short layered hair, blue eyes, brown crop top with gray trim, brown loincloth, brown wrist guards, brown boots with gray cuffs. The trigger alone keeps her orange hair and face, but the outfit is random fan art (a green dress, a blue bodysuit). With the tags the brown outfit comes out in every test. Base Anima with the same tags draws a generic barbarian with a short bob (weight 0 in the comparison).
Tips:
holding a large double-headed axeorholding axefor the axe; there is no axe without it.chibigives the chibi style;3d rendergave a soft 3D look in 1 of 2 tries.Outfit changes: weight 0.8, keep
orange hair, short layered hair, blue eyesand drop the outfit tags (a white hoodie in the previews).Proportions are exaggerated, as in the dataset.
Works at 896×1152 and 1024×1024.
Note: the dataset is mostly explicit. With nsfw, nude, topless, nipples in the negative none of the 26 test images were explicit.
Tested with: 30 steps, CFG 4, er_sde / simple.
Negative: worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, sepia, signature, watermark, text, artist name (add nsfw, nude, topless, nipples for SFW). The dataset has artist signatures; the signature tags keep the corners clean.
DiT-only LoRA (no text encoder weights), rank 16 / alpha 16, 2500 steps, lr 2e-5.







