Reverse Direction

Key Insight

A standard VLM only runs one way — pixels in, text out — because its output layer can emit text tokens and nothing else. Adding an image-token output head (a second prediction layer that emits discrete image tokens drawn from a VQ-VAE codebook) lets the same model predict picture codes as easily as words, which a decoder then turns back into pixels — converting an understanding-only model into one that can generate images too. This is the cheapest path toward any-to-any: instead of retraining a native multimodal model from scratch, you graft an output head onto a model that already "speaks image" on the input side and teach it to speak image on the output side as well.