wdrink/arm

ARM is a research project from ByteDance Seed and the Institute of Trustworthy Embodied AI that presents a single AI model capable of understanding images, generating images from text descriptions, and editing images based on instructions, all using the same underlying approach. The repository accompanies an academic paper and currently contains the paper's figures, benchmarks, and citation information rather than runnable code, as the authors note that more updates are coming. The core idea is to convert images into a compact sequence of discrete tokens (small numeric symbols that represent visual concepts), which lets the same kind of model that handles text also handle images. Once images are expressed as tokens, the model can treat understanding a scene, generating a new image, and editing an existing one as variations of a single task: predicting the next token in a sequence. This is the same approach language models use to predict the next word. The paper reports competitive results on standard benchmarks for each of the three tasks. For text-to-image generation, ARM is compared against other models on tests called GenEval, DPG, and WISE. For image editing, the authors found that adding a reinforcement learning step (where the model is trained using feedback on the quality of its edits) raised its score on an editing benchmark from 5.75 to 6.68. For image understanding, ARM outperforms several other models that also use discrete visual representations on a suite of visual question-answering tests. One finding highlighted in the README is that the autoregressive model and a separate diffusion-based rendering component play distinct roles: the autoregressive model determines what content to generate and how it is arranged, while the diffusion component handles pixel-level rendering detail. The repository is research code associated with a 2026 arXiv paper and does not appear to include a public demo or installation instructions yet.