Background-Improver

Background Improver — Krea 2 Raw style LoRA (v2, people-free)

Always-on style LoRA whose ONLY training data was 18 amateur phone photos of real environments (bedrooms, living rooms, kitchens, beaches, gardens, a car, a lake dock) — ZERO people. Goal: backgrounds that get real detail, real textures, and physically sensible room/scene logic instead of the blurry basic-shapes default, while stacking under character LoRAs.

How to use

• No trigger word. Load it and it is on; LoRA strength is the only knob.
• Checkpoints = training steps 1250 to 2750. Earlier = subtler; later = stronger scene redesign (later steps visibly re-lay-out rooms/pools — coherent, but more change).
• Suggested start in a two-stage Krea workflow: stage 1 only at 0.6-0.8 (v1 of this LoRA at 1.0 in BOTH stages overcooked to painterly).
• When stacking with a character LoRA, name the outfit in the prompt.
• Captions were content-only with caption dropout 0 — the realism/detail register is baked in, never prompted.

Status — honest

Training samples showed un-prompted background detail appearing (cluttered rooms, packed supermarket shelves) and no empty-scene bias, BUT bare-model samples also drifted the default (unprompted) woman: body change, occasional tattoos — none of which exist in the dataset. Inference A/B under character LoRAs NOT yet run. Treat as experimental; test checkpoints against LoRA-off with a fixed seed before production.

Recipe

Krea 2 Raw, rank 16/16, lr 1e-4 adamw8bit, flowmatch linear, balanced, 1024, batch 1, grad-ckpt off, caches + unload_text_encoder on, caption_dropout 0.0, stopped at 2750/3000.