remyxai/SpaceQwen2.5-VL-3B-Instruct
SpaceQwen2.5-VL-3B-Instruct by remyxai is a 3.75 billion parameter multimodal vision-language model, fine-tuned from Qwen2.5-VL-3B-Instruct using LoRA. This model specializes in enhancing spatial reasoning capabilities, particularly for understanding 3D positional relations between objects in images. It is optimized for tasks requiring detailed spatial analysis, such as inferring distances and relative positions.
Loading preview...
SpaceQwen2.5-VL-3B-Instruct: Enhanced Spatial Reasoning VLM
SpaceQwen2.5-VL-3B-Instruct is a 3.75 billion parameter multimodal vision-language model developed by remyxai. It is fine-tuned from the Qwen/Qwen2.5-VL-3B-Instruct architecture using LoRA (Low-Rank Adaptation). The core innovation of this model lies in its enhanced spatial reasoning capabilities, achieved through data synthesis techniques and publicly available models to reproduce the work described in SpatialVLM.
Key Capabilities & Features
- Advanced Spatial Reasoning: Excels at understanding and inferring spatial relationships between objects in a scene, including distances, relative positions (e.g., above, left-of, contains, closest to).
- Multimodal Input: Processes both image (RGB) and text inputs for Visual Question Answering (VQA) tasks focused on spatial queries.
- Fine-tuned on OpenSpaces: Trained on the OpenSpaces dataset, which contains approximately 10,000 synthetic spatial reasoning traces.
- Strong Benchmark Performance: Demonstrates competitive performance on various spatial reasoning benchmarks, including:
- SpatialScore: Shows strength in 3D positional relations categories.
- OmniSpatial: Assessed for dynamic reasoning, complex spatial logic, and perspective-taking.
- SpaCE-10: Achieves notable scores in Object-Object and Object-Scene spatial relations.
Use Cases & Considerations
This model is particularly well-suited for applications requiring precise spatial understanding from visual data. It can be deployed via Ollama, Hugging Face Transformers, or GGUF with llama.cpp.
Limitations: Performance may degrade in cluttered environments or with challenging camera perspectives. It inherits multimodal biases from its base model and is not recommended for safety-critical or legal decision-making.