Krea 2 nails fridges and couches but can't generate a normal-length bed
QuirksNFeatures · reddit · 2026-09-22
A user reports Krea 2 gets proportions right for refrigerators, dining tables, couches, chairs, and doorways — but almost always generates beds that are comically too short, worst for King/Queen sizes, often shorter than they are wide. Prompt tricks, adding people for scale, and LLM-assisted prompting all failed; the poster asks whether a prompt trick exists or if training a LoRA would help. A good example of a spatial-prior blind spot in image models, likely from beds being perspective-skewed in training data.
More from Multimodal
- Qwen Image 2.1 int8 tests: color tints and hand glitches persist, transparency support shines — Ant_6431 · 2026-09-22
- Creator open-sources Skills and prompt templates distilled from 463 viral AI videos — dotey · 2026-09-22
- Xiaomi ships MiMo-V2.6 on OpenRouter: 1T-parameter open omnimodal model with 1M context — CurieuxExplorer · 2026-09-22
- HYPER3D Agentic Mode lets you edit 3D models with plain English prompts — Scobleizer · 2026-09-22
- Early hands-on: Suno v6 feels higher fidelity with noticeably different stylization — Kyrannio · 2026-09-22
- Seedance 2.0 test: tiny details like fingers and smiles make AI film directing work — gen_ericai · 2026-09-22