Show HN: VQASynth – pipelines to synthesize VQA datasets (github.com)

Inspired by the recent work in SpatialVLM, we reproduce similar data synthesis pipelines using openly available models.

We compare our results to alternative annotation pipelines like RAM-Grounded-SAM.

Our repo uses a simple pipeline in docker compose to produce datasets suitable for fine-tuning multimodal models like LLaVA.