Autoregressive fashions for textual content generally generate repetitive and low-quality output as a result of errors accumulate through the steps of technology. This subject is commonly attributed to publicity bias – the distinction between how a mannequin is educated and the way it’s used throughout inference. Denoising diffusion fashions present an alternate strategy during which a mannequin can revisit and revise its output. Nonetheless, they are often computationally costly, and prior efforts on textual content have led to fashions that produce much less fluent output in comparison with autoregressive fashions, particularly for longer textual content and paragraphs. On this paper, we suggest PLANNER, a mannequin that mixes latent semantic diffusion with autoregressive technology, to generate fluent textual content whereas exercising international management over paragraphs. The mannequin achieves this by combining an autoregressive “decoding” module with a “planning” module that makes use of latent diffusion to generate semantic paragraph embeddings in a coarse-to-fine method. The proposed technique is evaluated on numerous conditional technology duties, and outcomes on semantic technology, textual content completion, and summarization present its effectiveness in producing high-quality long-form textual content in an environment friendly method.