We ran 120 AI generations across the hardest real-world edge cases, found every failure mode, and fixed what we could. Here's exactly what works, what doesn't, and why.
Basic architecture lock + room-type rules
Mandatory furniture, darkness rule, aggressive premium
MANDATORY style overrides, anti-reimagination declutter
AI places furniture from scratch in an empty room
AI replaces furniture with a new style while preserving architecture
AI removes clutter, boxes and mess — nothing structural is changed
AI improves lighting and atmosphere without changing furniture or layout
Click any bug to expand details and fix notes.
These failures are fundamental to how Flux Kontext Pro works. No amount of prompt engineering fixes them — they require architectural changes to the pipeline.
When the input has saturated, well-composed furniture (dark velvet sofa, matched color scheme), the model keeps it. Text instructions cannot override dominant visual conditioning.
When the room has exposed brick, track lights, and industrial aesthetic, the model refuses to place velvet or birch furniture. Only a rug is placed. Modern + neutral_sale work fine in the same room.
When a room has zero architectural anchors visible (all walls hidden by shelving, no windows, no light source), the model cannot remove clutter without reference points — it reimagines the entire room.
On deeply underexposed rooms, some seeds latch onto the dark input and don't diverge — the output stays dark. Other seeds brighten correctly.
Test photos from Pexels (free license). AI by Flux Kontext Pro via fal.ai. Report date: 2026-05-13.