Senko-san_NAI-V5-Curated_ORIGINAL-ANIME_ULTRA (仙狐さん) - V1 Showcase
I compared my Senko-san LoRAs on FLUX.2 Klein 9B Base and FLUX.3 Klein Base, including a test of Black Forest Labs’ FLUX.3 prompt upsampler.
The upsampler is a separate language model that expands a scene request into a more detailed description of the character, clothing, pose, lighting and surroundings. I reviewed its output and corrected invented character details before generating the images.
The three-column grid shows:
FLUX.2 with the original prompt
FLUX.2 with the expanded prompt
FLUX.3 with that same expanded prompt
Seeds, steps, CFG, image size and LoRA strength were matched within each row, with each model using its native sampling schedule. The first two columns show the effect of prompt expansion; the last two compare the model/LoRA combinations using identical text.
Across these and the additional outfit and action tests, a few differences stood out:
FLUX.3 had lovely lighting and shading, especially in the sunset scenes. The dress folds looked more detailed, and some expressions captured Senko’s little cat-like smile better.
FLUX.2 remained very competitive for her overall look. In some images, I preferred its face and proportions.
FLUX.3 handled tail-brushing more clearly, with a recognizable brush touching the fur. However, its tail wasn’t visible in the hoodie examples, while FLUX.2 showed one.
Both missed character details. Ears could be too tall, eyes weren’t always quite right, and the flower ornament was sometimes missing or changed.
Hands and interactions remained uneven, particularly the boardwalk hand-holding and pretzel exchange.
Expanded prompts helped with some scene details, but didn’t consistently solve character accuracy or anatomy. FLUX.3 produced beautiful results; this comparison showed scene-by-scene tradeoffs, with neither model consistently coming out ahead.