Skip to content

FIFO-Diffusion

Generate long text-conditioned videos without training

Visit official site

What FIFO-Diffusion is

The technique is described as a novel inference approach built on pretrained diffusion models for text-conditioned video generation. It generates videos of unlimited length without training by iteratively denoising a queue of consecutive frames with gradually increasing noise levels.

What you can do with it

  • Generate text-conditioned video without training
  • Work with several video generation baselines
  • Produce high resolution video output
  • Start from a text description of the scene

How it works

  • A fully denoised frame leaves the head of the queue while a new random noise frame enters the tail
  • Diagonal denoising runs iteratively over consecutive frames
  • Latent partitioning narrows the gap between training and inference
  • Lookahead denoising makes use of forward references

Example uses

  • Generate a fireworks display above Sydney Harbour
  • Produce a 4K clip of penguin colonies on the Antarctic ice
  • Create a high quality 4K video of an astronaut floating in space

Pros & cons

✓ What we like

  • No training is needed to generate video
  • Several video generation baselines are supported, including VideoCrafter2 and Open-Sora-Plan

! What to watch out for

  • Latent partitioning is used to narrow the training inference gap
  • Lookahead denoising relies on forward references

FAQ

Does it need training?

The material states video can be generated without training.

Which baselines are named?

VideoCrafter2 and Open-Sora-Plan are listed as supported video generation baselines.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review