Pixel is All You Need
Researchers investigated whether current visual generation modules can perform symbolic reasoning over text in pixels, finding that they fail to do so without an underlying text module or prompt rewriter. To test if this stems from pixels or training paradigms, the team trained a family of purely pixel-space generative models from scratch without any text pathway. They found that these models exhibit hallmark behaviors of large language models, including predictable performance improvements with scale, enhanced reasoning through frame-by-frame intermediate steps, and cross-modality reasoning on a single canvas. Open article →