Branching diffusion cuts the work of long text generation

Can continuous diffusion reduce the work of generating long sequences without making the text harder to generate? Token-level diffusion still processes a state for every output token at each step, while compression makes each remaining state carry more information. The research tests whether a system can address both pressures.

Branching Latent Diffusion compresses text into blocks and generates them in groups, cutting generation computation while decoding block spans in parallel.

The authors ask whether continuous diffusion can generate long text with less work while still producing usable sequences. Branching Latent Diffusion compresses text into block latents, generates them in groups, and decodes their token spans in parallel.

For approximately 1,024-token sequences, BLD used 1.1 TFLOPs per sequence versus 96.4 for ELF-L. It also had higher batch throughput than ELF-L and an autoregressive baseline, but longer single-sequence latency than ELF-L. Its score on a measure comparing generated with held-out text was below both baselines.

These results come from unconditional generation of 1,024-token OpenWebText sequences. The study did not include human evaluation or a direct measure of long-range coherence; its comparisons with ELF differ in tokenizer and training budget. It tests one model scale, sequence length and full-system compression ratio.