Paper summary: FIFO-Diffusion: Generating Infinite Videos from Text without Training

Problem

Text-to-video diffusion models are trained to produce short clips of fixed length. Making long videos usually needs extra training or stitching tricks.

Method

FIFO-Diffusion is purely an inference technique:

• Diagonal denoising: keep a queue of consecutive frames at increasing noise levels. Each step denoises all of them a little, dequeues the fully clean frame at the head, and enqueues a new pure-noise frame at the tail. The video can go on indefinitely.
• That creates a mismatch with training, where all frames share one noise level. Latent partitioning narrows this gap, and lookahead denoising lets noisier frames benefit from cleaner frames ahead of them.

Key results

According to the abstract, the method generates conceptually infinite videos without additional training, uses a constant amount of memory regardless of length, and is well suited to parallel inference on multiple GPUs.

Limitations

The extracted text has no dedicated limitations section. The approach depends on the quality of the underlying pretrained video model, and the authors themselves describe diagonal denoising as a double-edged sword because of the training–inference discrepancy.