Style LoRA Study · Realistic · AI Lab

Valkyrie in Motion — Style LoRA Study

Stack

Role: Dataset curation, captioning, LoRA training, epoch-sweep validation
Base model: Krea 2 (open-weights)
LoRA: Self-trained painterly ornate LoRA ↓
Tools: ComfyUI, self-trained style LoRA + Krea 2, directed generation
Year: 2026

Process

Pushing a style LoRA into motion

A validation piece from an epoch sweep of a self-trained painterly LoRA — testing whether the style holds under a dynamic action pose it never saw in training. Training recipe lives in the LAB card.

The sweep it came from

Fixed seed, out-of-distribution probe, checkpoint by checkpoint:

Epoch sweep

From the AI Lab

Model Card Painterly Ornate Fantasy — LoRA Ornate semi-realistic fantasy character style. 150-image dataset, light-touch epoch selected.

Overview

A personal artist-study LoRA for ornate painterly fantasy characters. Trained as a style exercise; used for the Dark Legion set and the Valkyrie in Motion action study.

Card


Base model	Krea 2 (open-weights, turbo)
Dataset	150 curated images, prose captions, no trigger word
Network	LoRA dim 32 / alpha 32 · final loss 0.0735
Validation	Cross-concept sweep (knights, archers, rogues, modern fashion); early light-touch epoch selected as standard

Notes

The interesting finding: the light-touch epoch keeps base-model anatomy and composition while carrying the surface language — heavier epochs win on style but lose flexibility.

Epoch sweep — out-of-distribution probe

Fixed seed, fixed prompt (a subject not in the training set), swept across checkpoints. This is how the operating epoch gets chosen — by evidence, not vibes:

Epoch sweep

Open in LAB →