Tuna-2: Facebook Research Shows Pretrained Vision Encoders Are Unnecessary — Pixel Embeddings Achieve SOTA on Multimodal Benchmarks

arXiv / HuggingFace Daily Papers·medium signal

Facebook Research's Tuna-2 demonstrates that simple patch embedding layers can replace pretrained vision encoders (VAE, representation encoders) entirely for multimodal modeling. The encoder-free design achieves SOTA on multimodal benchmarks and shows stronger visual understanding at scale on fine-grained perception tasks. This could trigger an architectural simplification wave, eliminating the modular vision encoder bottleneck.