GeoReader-Qwen2.5-VL-3B-MindCube: A Vision-Language Model for Geospatial Analysis
2 min readupdated
The GeoReader-Qwen2.5-VL-3B-MindCube model is a full checkpoint (set of weights) based on Qwen2.5-VL-3B-Instruct, fine-tuned for spatial perception and geometry tasks. The architecture combines a vision-language model (VLM), a frozen geometric reader, an adapter, and a gate. The solution delivers high accuracy on the MindCube-Tiny dataset and requires a specialized GeoReader loader instead of a standard Transformers AutoModel call. This model is useful for researchers and developers working with spatial image analysis and computer vision.
Key features
- High accuracy: accuracy on MindCube-Tiny is 87.14% (915 out of 1,050 tasks).
- Geometric analysis: detailed evaluation of rotation (60.5%), relative positions (93.83%), and surrounding objects (92.4%).
- Hybrid architecture: combines the base Qwen2.5-VL-3B VLM with a separate geometric reader and adapter.
- Reconstructed geometry: training was performed using DA3 Nested Giant-Large 1.1 settings for multi-view inference.
What you can do
- Spatial relationship analysis: evaluate the relative placement and orientation of objects from images.
- Accuracy testing: benchmark the model on the MindCube-Tiny dataset using the provided scripts.
- Research experiments: study the performance of geometric adapters and gates in VLM architectures.
Example
Running the model requires a special GeoReader loader instead of standard Hugging Face methods, as well as submitting prompts with a response limit of 128 tokens.
Read more
DetailsBase model: Qwen/Qwen2.5-VL-3B-Instruct
Metrics: MindCube-Tiny accuracy 87.14% (915/1,050)
Features: requires a special GeoReader loader instead of AutoModel
Data: trained on the MindCube DA3 dataset package










Comments