Where2Place
Multimodal · 2024-06-15
Where2Place, released with RoboPoint, has 100 real-world cluttered images, each with an instruction describing a free-space placement region; a 30-item hard subset covers unseen relation types. The model outputs points, each item scores the fraction of its points inside the target mask, and the benchmark score is the mean over items.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Luna | 51.6 |
| Gemma 4 12B IT | 49.0 |
| Qwen3.5-4B | 47.2 |
| LFM3-VL-2B | 44.9 |
| Qwen3.5-2B | 30.2 |
| Gemini 3.5 Flash-Lite | 21.0 |
| Gemma 4 E2B IT | 18.2 |