部分通路自回归成像

Parti frames text-to-image generation as sequence-to-sequence modeling, analogous to translating a text sequence into another type of sequence. The target is a sequence of discrete visual tokens rather than words in a second language. A ViT-VQGAN tokenizer encodes training images into these tokens, and its decoder can reconstruct images from token sequences. The text-to-image model uses an encoder to represent the prompt and a decoder to generate image tokens autoregressively. Each next-token prediction is conditioned on the text and previously generated image tokens. That dependency requires sequential generation steps; it does not by itself determine the latency of every possible implementation. The researchers compared models from 350 million through 20 billion parameters and reported consistent improvements in their evaluations. Those findings describe their trained models and evaluation conditions rather than a guarantee that increasing any model always solves every prompt. The larger model could handle more demanding compositions, but the paper documents failures and inconsistencies, including counting, attribute binding, spatial relationships, and text rendering. PartiPrompts is a released benchmark containing more than 1,600 English prompts across categories and challenge aspects. It is an evaluation resource, not the name of the entire training dataset or a public image-generation application. Useful comparisons examine what the system produced for a prompt, what failed, and which model configuration was used. Individual appealing examples are informative demonstrations, but neither an example nor an aggregate score proves reliable performance on every new request.