vra / Thinking-with-Visual-Primitives-pytorch

What is Thinking-with-Visual-Primitives-pytorch?

This GitHub repo is an unofficial PyTorch reproduction of DeepSeek's Thinking with Visual Primitives, a pipeline that trains vision-language models to embed bounding boxes and points directly into chain-of-thought outputs for precise visual reasoning. It closes the "reference gap" in tasks like object grounding, counting, maze navigation, and path tracing, turning vague descriptions into structured coordinates. Python-based with PyTorch, users get one-command data generation, multi-stage training (pretrain, SFT experts, distillation), and CLI inference on images or JSONL batches.

Why is it gaining traction?

It democratizes a cutting-edge paper with 12GB GPU configs, bilingual READMEs, and scripts for visual comparisons across epochs—far more accessible than opaque official repos. Devs dig the end-to-end reproducibility, from COCO-derived data to eval metrics, plus quantization for low-VRAM runs. As an unofficial PyTorch take on DeepSeek's primitives, it hooks experimenters tired of black-box VLMs.

Who should use this?

Vision-language researchers reproducing multimodal papers, agent builders needing spatial reasoning (e.g., robotics pathfinding or AR overlays), and PyTorch tinkerers fine-tuning Qwen-VL bases for grounding tasks. Ideal for teams evaluating unofficial PyTorch optimizations like training loops on custom primitives.

Verdict

Grab it for a battle-tested reproduction pipeline with strong docs and MIT license, despite 48 stars and 1.0% credibility signaling early maturity—no tests yet, but visualization/eval scripts fill the gap. Fork if DeepSeek's thinking primitives intrigue you.

(198 words)