FrankenVAR: Enabling generative vision modality for text-only LLMs (opens in new tab)

The idea looks technically solid to me. Would be great if someone committed a few grand to check if it works on Llama-8B.